Institutions today face a threat that older public relations manuals never imagined. A convincing video of a chief executive announcing a product recall, a cloned voice of a university president making a racist remark, or a fabricated press release written by a language model can now be produced in minutes and shared with millions before the organisation even knows it exists. This article develops a structured response to that problem. It reviews recent scholarship on #deepfakes, #generative_AI, #misinformation_psychology, and #crisis_communication, and it uses that literature to build a four phase protocol that communication teams can adopt. The four phases are Anticipate, Authenticate, Answer, and Absorb. The article argues that #synthetic_media is not simply a new channel for old rumours. It changes the evidential basis of public trust, because audiences can no longer assume that a recording of an event is proof that the event happened. This shift produces two linked dangers for organisations: false content that looks true, and true content that can now be denied as fake. The article explains why traditional crisis models, which assume that facts can be established quickly and communicated clearly, are poorly suited to this environment. It then sets out concrete pre crisis, response, and recovery measures, including content provenance systems, verification chains, prebunking, correction design, platform escalation, and measurement indicators. The framework is illustrated with reference to universities, government bodies, health agencies, corporations, and civil society organisations. The article closes with a research agenda for scholars of #public_relations and #reputation_management. Keywords: crisis communication, deepfakes, synthetic media, generative artificial intelligence, misinformation, institutional reputation, public relations, content provenance, prebunking, media forensics 1. Introduction For most of the twentieth century, public relations practice rested on a quiet assumption. If an organisation was accused of something it had not done, it could eventually prove the accusation false. Documents could be produced. Witnesses could be interviewed. Recordings could be played. The truth might take time to establish, and it might never fully catch up with the rumour, but the truth existed in a form that could be shown to people. Photographs and video, in particular, carried a special weight. They were treated as a kind of final evidence, a backstop that could settle disputes when words alone could not. That backstop is now weakening. Generative models can produce video, audio, images, and text that most people cannot reliably tell apart from recordings of real events. #voice_cloning needs only a few seconds of sample audio. Face swapping tools are available as consumer applications. #language_models can generate hundreds of variations of a false statement, each written in a different voice, each tuned for a different audience. What was once the work of a well funded film studio is now within reach of a single motivated individual with a laptop and a grievance. The academic literature has begun to map the consequences. Studies show that people are poor at detecting manipulated video and are often overconfident in their ability to do so (Kobis et al., 2021; Groh et al., 2022). Synthetic faces generated by machine learning are not only hard to distinguish from real faces, they are sometimes rated as more trustworthy than real ones (Nightingale and Farid, 2022). Exposure to a political deepfake may not always change attitudes in a simple way, but it does something arguably worse: it increases uncertainty about what is true, and that uncertainty reduces trust in legitimate news (Vaccari and Chadwick, 2020). Fallis (2021) describes this as an #epistemic_threat, because it degrades the evidential value of the recording itself. Rini (2020) makes a similar argument, calling video and audio an epistemic backstop whose collapse would leave public discourse without a shared method for settling factual disagreements. The consequences for institutions are direct. A university whose admissions director appears in a fabricated video accepting a bribe faces immediate reputational damage, regardless of whether the video is later shown to be synthetic. A hospital whose logo appears on an AI generated memo announcing a fake contamination event faces public panic. A government agency whose spokesperson is voice cloned into announcing an evacuation faces a genuine safety emergency. And a company whose executive is caught on a real recording behaving badly may now escape accountability by claiming that the recording is fake. This second effect, often called the #liars_dividend, is a distinctive feature of the synthetic media era, and it means that the technology damages both the falsely accused and the legitimately exposed. Crisis communication scholarship has produced sophisticated models for how organisations should respond to accidents, product failures, misconduct, and rumours. #situational_crisis_communication_theory tells us to match the response strategy to the level of attributed responsibility (Coombs, 2019). Image repair theory catalogues the rhetorical options available to an accused organisation. The social mediated crisis communication tradition explains how information moves through influential creators, followers, and inactive audiences. Yet these models were developed in an environment where the primary uncertainty was about interpretation and blame, not about whether the underlying evidence was real. When the evidence itself is contested, the standard playbook stalls. This article addresses that gap. It asks a practical question with theoretical implications: what should a communication team actually do, before, during, and after a synthetic media attack on its institution? To answer it, the article proceeds in stages. Section 2 defines the key concepts and distinguishes synthetic media crises from adjacent phenomena. Section 3 reviews the theoretical resources available in crisis communication and misinformation psychology, and identifies where they fall short. Section 4 explains the method used to build the framework. Section 5 offers a typology of synthetic media attacks on institutions. Section 6 explains why classic playbooks fail. Section 7 presents the four phase protocol, which is the central contribution. Section 8 discusses law, regulation, and ethics. Section 9 addresses measurement. Section 10 applies the framework across sectors. Sections 11 to 13 discuss implications, limitations, and future research. The intended reader is a student of communication, journalism, or management, and also the practitioner who must write the plan that the student is studying. The language is therefore kept plain, but the argument is built on peer reviewed evidence, and every claim is anchored to a source. 2. Conceptual Background 2.1 Defining synthetic media, deepfakes, and AI generated misinformation Terms in this field are used loosely, which causes confusion in both research and practice. This article uses the following distinctions. Synthetic media is the broad category. It refers to any media artefact that is generated or substantially modified by an algorithmic system rather than captured directly from the world. This includes AI generated images, cloned voices, machine written text, and computer generated video. Synthetic media is not inherently deceptive. A company that uses a synthetic voice to narrate a training video in twelve languages is using synthetic media honestly. The ethical status depends on disclosure and intent, not on the technology. Deepfakes are a subset of synthetic media in which a real person is depicted saying or doing something they did not say or do, in a way designed to appear authentic. Kietzmann et al. (2020) define them in terms of the underlying learning architecture, in which competing networks improve the realism of the output. Karnouskos (2020) emphasises the deceptive purpose. Whittaker et al. (2021) note that the term now covers a wide spectrum, from satire and entertainment to targeted harassment. For this article, the defining features of a deepfake are impersonation and the appearance of authenticity. AI generated misinformation is broader still. It includes machine written false news articles, fabricated documents, synthetic user reviews, and coordinated inauthentic comment campaigns. Ferrara (2024) catalogues these uses and notes that the main change introduced by generative systems is not novelty of content but scale, speed, and personalisation. A single operator can now produce the volume of material that once required a staffed organisation. It is useful to separate misinformation from disinformation. Misinformation is false content shared without the intention to deceive, often by people who believe it. Disinformation is false content created and spread deliberately, with a strategic goal. Most synthetic media crises begin as #disinformation and then propagate as #misinformation, because the people who share the fabricated video sincerely believe it is real. This matters for response design, because a strategy aimed at the original bad actor is different from a strategy aimed at the sincere but mistaken sharer. Finally, an important asymmetry deserves a name. The #detection_asymmetry describes the gap between the effort required to create synthetic content and the effort required to prove it false. Creation is cheap, fast, and requires no expertise. Verification is slow, expensive, requires specialist tools, and often produces a probabilistic rather than definitive answer. Leibowicz et al. (2021) call the resulting situation a detection dilemma, because publicising detection methods teaches adversaries how to defeat them, while keeping them secret prevents public trust in the results. 2.2 Institutional reputation as the asset at risk Reputation is often described casually as what people think of an organisation. A more useful working definition, consistent with the crisis literature, is that reputation is the aggregated set of expectations that stakeholders hold about how an organisation will behave. It is an asset because it lowers the cost of transactions. Students enrol at a university partly because they expect its degree to be respected. Patients accept advice from a health agency partly because they expect it to be honest. Investors accept risk partly because they expect management to be competent. Reputation is therefore a form of stored trust, and #institutional_reputation is drawn down whenever behaviour fails to match expectation. The crisis literature has long argued that reputational damage in a crisis depends less on the event itself than on the #attributed_responsibility, and on the organisation's crisis history (Coombs, 2019). A synthetic media attack complicates this logic in a specific way. In an ordinary crisis, the organisation knows what happened and must decide how to frame it. In a synthetic media crisis, the organisation may be entirely innocent of the alleged conduct, but the public sees what appears to be direct evidence of it. The organisation is placed in the strange position of denying something that stakeholders believe they have witnessed with their own eyes. Denial is normally the weakest crisis response, because publics treat it as evasive. In a deepfake crisis, denial is the only accurate response, and yet it carries all the credibility costs that the literature associates with evasion. This is the central strategic paradox of synthetic media crisis communication, and any useful protocol must address it. 2.3 Why synthetic media constitutes a distinct crisis type Crisis typologies usually sort events by attributed responsibility. Victim crises, such as natural disasters, carry low responsibility. Accidental crises carry moderate responsibility. Preventable crises, such as misconduct, carry high responsibility. A deepfake attack does not fit comfortably in any of these categories. The organisation is a victim of an attack, which suggests low responsibility, but the content depicts preventable misconduct, which suggests high responsibility. Stakeholders must therefore decide which frame to apply before they can decide how much blame to assign, and that decision depends on whether they believe the content is real. This gives synthetic media crises three distinguishing features. First, the crisis is evidentiary before it is reputational. The first battle is over what is true, not over who is to blame. In most crises, those questions are settled in the opposite order. Second, the crisis is reversible in principle but not in practice. If the content is proven synthetic, the accusation collapses. Yet research on the #continued_influence_effect shows that corrected misinformation continues to shape beliefs and judgements even after people accept the correction (Walter and Tukachinsky, 2020; Ecker et al., 2022). Proving the video false does not fully undo the harm. Third, the crisis is recursive. Every successful deepfake makes the next genuine recording easier to dismiss, and every false claim of a deepfake makes real evidence weaker. Twomey et al. (2023) found that exposure to deepfake material was associated with generalised distrust in video evidence, not merely distrust of the specific clip. The damage accumulates across the information environment rather than staying contained within a single incident. These features justify treating #synthetic_media_crisis as a distinct category requiring a distinct protocol, rather than as a variant of rumour management. 3. Theoretical Framework 3.1 Situational Crisis Communication Theory and its limits Situational Crisis Communication Theory, developed over two decades by Coombs and colleagues, is built on attribution logic. Publics look for causes when something bad happens. The more they attribute the cause to the organisation, the more reputational damage follows, and the more accommodative the organisational response must be to protect reputation. The theory offers a ladder of #response_strategies, from denial and attack the accuser at one end, through excuse and justification, to full apology and compensation at the other (Coombs, 2019). The theory remains valuable, and its core prediction still holds: matching response to attribution reduces damage. Yet SCCT assumes that the facts of the case are broadly agreed, and that the contest is over interpretation. In a deepfake crisis, the facts are the contest. The theory has little to say about how an organisation should establish that an event did not occur, or about how it should manage a public that has seen apparently direct evidence. Claeys and Coombs (2020) extend the theory by examining why practitioners often choose suboptimal responses, drawing on behavioural economics. Their analysis is highly relevant here, because a synthetic media attack creates exactly the conditions under which decision makers panic: extreme time pressure, high emotional arousal, ambiguous evidence, and severe consequences for error. Under those conditions, teams tend to either over respond, amplifying a fabrication that few people had seen, or under respond, hoping the story will die. Both failures are predictable and both are avoidable with a pre agreed protocol. 3.2 Image repair, renewal, and the problem of denial #image_repair theory catalogues the defensive rhetoric available to an accused actor: denial, shifting the blame, providing an excuse, minimising the offence, taking corrective action, and offering mortification. #discourse_of_renewal, associated with Sellnow and Seeger (2021), takes a different view, arguing that crises can be opportunities for organisations to reaffirm their values and rebuild forward looking relationships rather than merely defend past conduct. Both traditions are useful for synthetic media crises, but in an unexpected way. The organisation cannot use most repair strategies, because it did nothing wrong. It must use denial. The task, therefore, is to make denial credible, which the classic literature never had to solve because credible denial was assumed to follow automatically from evidence. The renewal tradition offers a partial answer. An organisation that responds to a fabricated attack by reaffirming its values, explaining how it verifies its own communications, and committing to protect its stakeholders from future deception converts a defensive posture into a forward looking one. It changes the question from "did you do this?" to "how do we all protect ourselves from this?" That reframing is the strategic heart of the protocol proposed later in this article. 3.3 Social mediated crisis communication and information flow The #social_mediated_crisis_communication tradition distinguishes between influential social media creators, social media followers, and social media inactives, and traces how crisis information flows between them and through traditional media. This model matters for synthetic media because it explains the pathway by which a fabricated artefact becomes a mainstream story. Typically, a deepfake is seeded in a small, sympathetic, or hostile community, where it circulates among followers. It gains volume. A larger creator or a partisan outlet picks it up. Only then does mainstream media report it, often initially as a claim rather than a verified fact. Each of these stages offers a different intervention point, and the correct response at stage one is almost never the correct response at stage four. This is why blanket rules such as "always respond immediately" are dangerous. Responding loudly to a fabrication that only two hundred people have seen may introduce it to two hundred thousand. The literature on strategic silence, and the practical experience of platform trust and safety teams, both support a threshold based approach in which the decision to respond publicly depends on measured diffusion rather than on internal anxiety. 3.4 Misinformation psychology: what corrections can and cannot do The most important theoretical resource for this problem does not come from public relations at all. It comes from cognitive and political psychology. Four findings are central. The first is the continued influence effect. People who accept a correction often continue to use the discredited information when reasoning. A meta analysis by Walter and Tukachinsky (2020) confirmed that corrections reduce but do not eliminate reliance on misinformation. Ecker et al. (2022) explain the mechanisms, including the difficulty of updating a mental model once a causal story has been installed. For practitioners, the lesson is blunt: a #correction is damage control, not repair. The second is the value of detailed refutation over bare denial. Chan and Albarracin (2023) found in a meta analysis of science relevant misinformation that corrections work better when they supply an alternative causal account rather than simply asserting that the claim is false. Telling people that a video is fake is weaker than telling them who made it, why, how it was made, and what actually happened instead. The third is inoculation, also called #prebunking. Lewandowsky and van der Linden (2021) and Traberg et al. (2022) review a large body of experimental evidence showing that warning people in advance about a manipulation technique, and giving them a weakened example of it, builds resistance. Roozenbeek et al. (2022) demonstrated that short inoculation videos improved the ability of social media users to recognise manipulation techniques at scale. This is arguably the single most transferable finding for institutional communication, because it can be deployed before any attack occurs. The fourth is a warning about warnings. Ternovski et al. (2022) found that informing people about the existence of deepfakes reduced their belief in genuine footage as well as fabricated footage. In other words, awareness raising has a cost. It creates general scepticism, which is exactly the condition that empowers the liar's dividend. Any institutional education campaign must therefore teach verification behaviour, not merely doubt. Telling audiences "you cannot trust video" is harmful. Telling audiences "here is how our organisation lets you check whether a message is really from us" is helpful. Together these findings support a design principle that runs through the entire protocol: build verification pathways rather than distributing doubt. 3.5 An integrated framework: the four phase Synthetic Media Crisis Protocol Drawing these strands together, this article proposes a framework with four phases. The names are chosen to be memorable for teaching purposes. Anticipate. Everything the organisation does before an attack: #risk_mapping, provenance infrastructure, authentication channels, prebunking, staff training, legal preparation, and simulation. Authenticate. Everything the organisation does in the first hours: triage, forensic assessment, internal verification, diffusion measurement, and the decision on whether and how to go public. Answer. The public response: message design, spokesperson selection, channel strategy, platform escalation, stakeholder segmentation, and legal action. Absorb. Everything after: correction reinforcement, trust rebuilding, learning, disclosure, and institutional change. The remainder of the article develops each phase in detail, but the logic of the sequence should be stated clearly at the outset. The quality of an organisation's response in phase three is almost entirely determined by the work it did in phase one. An institution that has no #content_provenance system, no authenticated channel, and no pre agreed verification chain will spend the critical first six hours arguing internally about whether the video is real, while the fabrication spreads. An institution that has done the preparatory work can publish a verified denial, with evidence, within the first hour. The difference is not talent. It is preparation. 4. Method This article is an #integrative_review rather than a systematic review or an empirical study, and it is important to be transparent about what that means and what it does not claim. The purpose of an integrative review is to synthesise findings from separate literatures that have not previously been read together, and to produce a framework that guides practice and generates testable propositions. The three bodies of literature integrated here are: crisis communication and public relations theory; empirical research on deepfakes and synthetic media perception; and cognitive research on misinformation, correction, and inoculation. Sources were selected according to three criteria. First, they had to be peer reviewed journal articles, scholarly books, or reports from established research institutions. Second, priority was given to work published within the last five years, on the grounds that the technological environment before that period is no longer representative. A small number of older foundational works are cited where a theory has no more recent authoritative statement. Third, sources had to make an empirical or conceptual claim that bears directly on institutional response, rather than on purely technical detection methods, since detection engineering lies outside the competence of a communication function. The synthesis proceeded thematically. Claims from each literature were grouped according to the phase of the crisis lifecycle they addressed, and gaps were identified where a phase was well theorised in one literature and neglected in another. The resulting protocol is therefore best understood as a set of #evidence_informed propositions rather than as an empirically validated model. Section 12 sets out how each element could be tested. The framework is normative in an important sense. It assumes that institutions have a legitimate interest in defending themselves against fabricated content, and that they should do so through verification and transparency rather than through counter deception. The article does not endorse the use of synthetic media by institutions for persuasive purposes, and section 8 explains why. 5. A Typology of Synthetic Media Threats to Institutions Practitioners cannot prepare for a threat they cannot name. The following typology organises the attack surface into eight categories, each of which requires a different anticipatory measure and a different response. 1. Executive impersonation. A synthetic video or cloned voice depicts a senior figure making a damaging statement. This is the archetypal case, and it is the most damaging because senior figures are treated as authoritative sources whose words carry institutional weight. Voice cloning is the most operationally dangerous form, because audio is easier to fake convincingly than video, and because it is used in real time fraud. A finance officer who receives a call in the chief executive's voice authorising an urgent transfer faces a decision under time pressure, without visual cues, and with strong social hierarchy incentives to comply. This is not only a reputational threat but a direct financial and security one, which is why #executive_impersonation must be treated jointly by communication, security, and finance functions. 2. Fabricated institutional documents. AI generated memos, letters, press releases, policy documents, or examination papers bearing genuine looking letterheads and signatures. These are cheap to produce and are often more persuasive than video, because audiences examine documents less sceptically than they examine faces. A fabricated internal memo announcing redundancies, tuition increases, or a product defect can move markets and mobilise protest before it is checked. 3. Synthetic testimony. Fabricated customers, patients, students, employees, or witnesses giving an account of harm. Nightingale and Farid (2022) showed that synthetic faces are perceived as trustworthy, and Sundar et al. (2021) found that video modality can increase the spread of false content in messaging applications. A campaign built from a dozen synthetic testimonials, each with a face, a name, and a story, is difficult to rebut because the organisation cannot prove a negative about people who do not exist. 4. Fabricated events. Images or video depicting an incident that never happened: a fire at a factory, contamination in a food plant, violence on a campus, flooding at a facility. These are particularly dangerous for public safety institutions, because the correct public response to a real event and to a fabricated one are opposite, and confusion between them can cost lives. 5. Counterfeit endorsement and brand hijacking. A synthetic version of an organisation's spokesperson, mascot, or logo promoting a fraudulent investment scheme, medical product, or political position. This category harms the public directly and damages the institution by association. It is a growing problem for health agencies and universities, whose credibility is the very thing that makes their counterfeit endorsement valuable to fraudsters. 6. Coordinated synthetic amplification. Language models used to generate large volumes of authentic seeming comments, reviews, letters to editors, and social posts, creating an illusion of #grassroots_opinion. Zerback et al. (2021) demonstrated that #astroturfing comments have measurable persuasive effects and that inoculation can partly counter them. Ferrara (2024) argues that the marginal cost of such campaigns has collapsed, which changes the strategic calculus for adversaries. 7. The liar's dividend. The use of the existence of deepfakes to deny #genuine_evidence. An executive caught in a real recording claims the recording is synthetic. A government denies authentic footage of an abuse. This is the mirror image of the other categories and it is a threat to the credibility of institutional accountability itself. It is also, uncomfortably, a temptation for institutions under pressure. An organisation that falsely claims a genuine recording is a deepfake commits a serious ethical breach and, if discovered, destroys its own verification credibility permanently. 8. Insider and hybrid attacks. Combinations of synthetic media with genuine leaked material, or with real access. The most effective attacks are rarely wholly false. They embed a fabricated element within a substantially true story, which makes the fabrication far harder to isolate and far more credible. Weikmann and Lecheler (2023) note in their synthesis of visual disinformation research that decontextualised authentic images and lightly manipulated content remain more common and often more effective than fully synthetic material. Two observations follow from this typology. First, the majority of these threats do not require sophisticated technology. #cheap_fakes, decontextualised real images, and simple text fabrications remain the dominant form of visual disinformation, and an institution that prepares only for high end video deepfakes will be defeated by a fabricated PDF. Second, most of these categories cross functional boundaries. #crisis_preparedness for synthetic media cannot sit inside the communication department alone. 6. Why Classic Crisis Playbooks Fail Six specific failure points explain why an organisation with an excellent conventional crisis plan can still respond badly to a deepfake. The speed gap. Conventional plans assume an internal verification period, often measured in hours, during which the facts are established before any statement is issued. Synthetic content reaches peak #diffusion far faster than that. By the time an organisation has convened its crisis committee, the fabrication may have been seen by its entire stakeholder base. The plan must therefore contain pre approved holding actions that can be executed within minutes by named individuals without a committee meeting. The burden of proof inversion. Conventional plans assume the accuser must produce evidence. Synthetic media arrives with apparent evidence already attached, which inverts the burden. The organisation must now prove a negative, which is logically difficult and rhetorically weak. The only escape from this trap is to have established, in advance, a positive method by which authentic institutional communication can be identified, so that the absence of that marker becomes itself a form of evidence. Detection is probabilistic and slow. Communication teams often assume that a technical tool can simply confirm whether a file is fake. In practice, #forensic_detection produces confidence estimates rather than certainties, performance degrades on compressed and re shared content, and detection models generalise poorly to generation methods they were not trained on (Farid, 2022). Groh et al. (2022) found that combining human judgement with machine assessment outperformed either alone, which supports a hybrid verification approach but not a push button one. Leibowicz et al. (2021) further show that detection capability is itself a strategic resource that cannot simply be made public without degrading it. Denial is structurally weak. As set out earlier, the accurate response is the response the public trusts least. The organisation must therefore invest the denial with something the public can check independently. Corrections underperform. Even a perfect, timely, well evidenced correction will leave residual belief. Walter and Tukachinsky (2020) and Ecker et al. (2022) establish this clearly. Practitioners who expect a correction to close the matter will be disappointed, and will often compound the error by repeating the correction in ways that increase familiarity with the false claim. Amplification risk. In no other crisis type does the standard advice to communicate quickly and widely carry the risk of causing the crisis. A rebuttal broadcast to an audience that has not seen the fabrication introduces the fabrication to them. This is why the protocol below places a formal #diffusion_threshold decision between detection and response. The conclusion is not that conventional crisis theory is wrong. It is that conventional theory addresses the second half of the problem, the contest over meaning, while synthetic media crises begin with a contest over reality. A complete protocol must add a stage that classic models do not contain. 7. The Four Phase Synthetic Media Crisis Protocol 7.1 Phase One: Anticipate Anticipation is the phase in which most of the value is created, and it is the phase most often neglected because it produces no visible output until an attack occurs. Threat and asset mapping. The organisation should begin by asking a simple question: what would an attacker fabricate, and about whom? The answer is rarely random. Attackers fabricate content that fits an existing narrative. A university with a live controversy about admissions will be attacked with a fabricated admissions scandal, not with a fabricated laboratory fire. The mapping exercise should therefore identify, first, the individuals whose voice and image carry institutional authority; second, the existing accusations circulating about the organisation, since fabrication almost always confirms an existing suspicion rather than inventing a new one; and third, the decision points at which a fabricated communication could cause immediate operational harm, such as payment authorisation, evacuation, admission, or recall. Digital footprint audit for principals. #voice_cloning requires sample audio, and face synthesis is improved by varied footage. Senior figures with hundreds of hours of public speaking online are more exposed than those without. The purpose of the audit is not to hide leaders from the public, which would defeat the purpose of their role. It is to establish awareness of exposure and to inform the priority of protective measures. Content provenance infrastructure. This is the most important structural investment available. #provenance_systems attach cryptographically verifiable metadata to media at the point of creation, recording the device, the time, and any subsequent edits. Industry standards for #content_credentials have developed rapidly, and the strategic logic for institutions is compelling. Rather than trying to prove that a fabricated video is false, the organisation makes it possible to prove that its authentic videos are authentic. Every official image, video, and audio file the institution publishes should carry provenance data, and this should be advertised to stakeholders as a feature they can check. Provenance does not detect fakes. It authenticates originals, which is the more tractable problem and the one the institution actually controls. A single authenticated channel of record. Every institution should designate one channel, ideally one it owns rather than rents, as the definitive source for official statements, and should tell stakeholders repeatedly that a message which does not appear on that channel is not from the institution. This is a simple, cheap measure with disproportionate value. It converts the verification question from the impossible "is this video real?" to the trivial "is this on the official page?" A #channel_of_record does not stop a deepfake from spreading, but it gives every journalist, employee, regulator, and member of the public a costless method of checking, and it makes the organisation's denial checkable rather than merely assertive. Verification codes for high risk transactions. Where a synthetic voice could cause direct harm, such as authorising a payment or ordering an operational action, the organisation must remove voice as a sufficient authorisation. Call back procedures, shared verification phrases, and #out_of_band_verification should be mandatory for defined transaction classes. This is a security control, but it is a communication team's responsibility to ensure it exists, because the reputational damage of a successful #voice_cloning fraud falls on communication. Prebunking and stakeholder inoculation. Drawing directly on inoculation research (Lewandowsky and van der Linden, 2021; Roozenbeek et al., 2022; Traberg et al., 2022), the organisation should proactively teach its stakeholders about the techniques of manipulation before an attack. The evidence supports short, technique focused messages delivered through channels the audience already uses. Crucially, and in light of the finding by Ternovski et al. (2022) that generalised deepfake awareness reduces trust in genuine content, the inoculation message must be constructive. It should not say "beware, videos can be faked". It should say "if you see a message that appears to come from us, here is exactly how to confirm it, and here is where our real announcements always appear". The measure of success is not increased scepticism, it is increased #verification_behaviour. Media and platform relationships. Journalists are the most consequential intermediate audience in a synthetic media crisis, because their decision to report a claim as verified, as alleged, or not at all determines the diffusion curve. Relationships with newsroom verification desks, #fact_checking organisations, and platform trust and safety contacts should be established before they are needed. An organisation that has to find the right person at a platform during a crisis has already lost several hours. Legal readiness. Pre drafted #takedown_notices, an understanding of the relevant defamation and impersonation provisions in each jurisdiction where the institution operates, standing instructions to counsel, and, where applicable, awareness of transparency obligations under emerging AI regulation. The legal team should be part of the protocol, not consulted after it. Simulation. The final anticipatory step is rehearsal. A #tabletop_exercise built around a fabricated video of the institution's own leader, run under realistic time pressure, will reveal more weaknesses in an afternoon than a year of policy drafting. The exercise should test the decision chain, not the messaging. Who declares a synthetic media incident? Who authorises a public response? Who contacts the platform? What happens if the incident begins at three in the morning on a public holiday? Simulation converts a document into a capability. 7.2 Phase Two: Authenticate When suspected synthetic content is detected, the organisation enters a defined authentication phase with a fixed maximum duration. The purpose of a time limit is to prevent the paralysis that Claeys and Coombs (2020) predict under conditions of ambiguity and pressure. A recommended structure is a first assessment within sixty minutes and a response decision within three hours, with a pre approved holding statement available immediately if diffusion is already rapid. Step one: internal ground truth. The fastest and most reliable verification is almost always internal, not technical. Was the chief executive in Geneva on the date the video claims? Does the memo use the current template? Was any such statement approved? Institutions have records, calendars, approval logs, and witnesses. A communication team that can establish within twenty minutes that the depicted event could not have occurred has achieved more than a forensic laboratory would achieve in twenty hours. The protocol should therefore begin with a checklist of #internal_verification questions, not with a detection tool. Step two: artefact analysis. Technical assessment of the file itself, including #provenance_metadata where present, compression artefacts, inconsistencies in lighting and audio, and analysis by detection tools. The critical point is that this step should be understood as producing supporting evidence with a confidence level, not a verdict. Farid (2022) and Groh et al. (2022) both make clear that detection is fallible and that human expertise remains necessary. The team should record what was tested, by whom, and with what result, because that record becomes evidence later. Step three: provenance tracing. Where did the artefact first appear, who posted it, and what is the account history? #origin_analysis often provides the most persuasive material for the eventual public response, because a fabrication traced to an account created three days earlier tells a story that a technical confidence score cannot. Step four: diffusion measurement. The organisation must measure how far the content has actually travelled, in which communities, and at what rate of acceleration, before deciding whether to respond publicly. This is where the diffusion threshold operates. A pragmatic three tier structure works well. Contained. Low volume, confined to a small community, no journalist enquiries, flat growth. Recommended action: monitor, document, request platform removal, prepare response materials, and do not amplify. Prepare a reactive statement for use only if asked. Emerging. Growing volume, crossing between communities, first journalist enquiries, stakeholder questions arriving through official channels. Recommended action: proactive notification to directly affected stakeholders through the channel of record, a factual statement, and active platform escalation. Response is targeted, not broadcast. Escalated. High volume, mainstream media coverage, market or operational impact, stakeholder alarm. Recommended action: full public response, senior spokesperson, evidence release, and sustained follow up. The purpose of this structure is to protect the organisation from the two symmetrical errors identified earlier: amplifying an obscure fabrication, and ignoring a spreading one. The threshold must be defined in advance, in numbers appropriate to the organisation's scale, because in the middle of an incident every team overestimates the visibility of the content that is causing them distress. Step five: the declaration. A named individual, with a named deputy, has the authority to declare a synthetic media incident and to activate the response protocol. Without a named #decision_authority, the organisation will debate rather than decide. 7.3 Phase Three: Answer The public response has six components. Component one: message architecture. The evidence on correction design is unusually clear and unusually actionable. Chan and Albarracin (2023) show that corrections work better when they replace the false account rather than merely negate it. Ecker et al. (2022) recommend a structure often described as a #truth_sandwich: lead with the truth, warn briefly about the falsehood without repeating it in detail or reproducing the fabricated material, explain the manipulation technique used, and close by restating the truth. The organisation should never republish the deepfake itself, and should never use the fabricated image as the illustration for its own correction, a mistake that institutions make with remarkable frequency. Applied to a synthetic media crisis, a strong response contains five elements: a clear statement of what is true; the assertion that the artefact is fabricated; the evidence for that assertion, stated in terms a non specialist can check; an explanation of how the fabrication was made and why, which supplies the alternative causal account the psychology literature calls for; and an instruction on how stakeholders can verify authentic communications in future. The fifth element converts the incident into an inoculation opportunity. Component two: speed and sequencing. The first public statement should be issued as soon as internal ground truth is established, even if forensic analysis is incomplete. It is entirely acceptable, and indeed credible, to say that the organisation has confirmed the depicted event did not occur and that technical analysis of the file is continuing. What is not acceptable is silence, because #silence in an evidentiary crisis is read as confirmation. Stakeholders should also be sequenced. Employees should learn of the fabrication from the organisation, not from social media, because employees are the organisation's most credible external advocates and their confusion is visible to everyone they know. Jin et al. (2020) found that employee backup materially improves the effectiveness of corrective communication against crisis misinformation, which makes #internal_communication a strategic priority rather than an administrative afterthought. Component three: the spokesperson problem. If the deepfake depicts the chief executive, should the chief executive respond? There is a strong argument for yes, because a live, unscripted, interactive appearance is currently difficult to fake in real time and demonstrates presence in a way that a written statement cannot. A live streamed session in which the leader takes unscripted questions is a powerful #authenticity_signal. There is also an argument for a second voice, because an independent verifier, such as an external forensic expert, a regulator, or a fact checking organisation, carries credibility that the accused institution cannot generate for itself. The strongest configuration uses both: the leader appears personally to deny, and an independent party confirms the technical finding. #third_party_verification is the single most valuable currency in this crisis type, precisely because the organisation's own word is structurally compromised. Component four: platform escalation. Parallel to public communication, the organisation pursues #content_removal through platform reporting channels, using the specific policy categories that platforms recognise, typically synthetic and manipulated media, impersonation, or fraud. Vizoso et al. (2021) document the divergent approaches platforms have taken to manipulated media, and the practical implication is that an organisation must know each platform's policy language in advance, because a report filed under the wrong category is slow or is rejected. Removal is rarely fast enough to prevent the initial spread, but it caps the long tail and it creates a documentary record. Component five: legal action. Legal remedies are slow and are usually not the solution to the immediate crisis. Their value is threefold. They deter repetition. They generate discovery that identifies the originator. And, importantly, the announcement that #legal_action is being taken is itself a credibility signal, because publics reason that an organisation would not sue over a genuine recording. Legal strategy should therefore be integrated with communication strategy, and counsel should understand that the announcement of proceedings has communicative value independent of their outcome. Component six: what not to do. Four errors recur. Do not reproduce the fabricated content. Do not respond to a contained fabrication with a mass broadcast. Do not attack the audience who believed it, since most of them were deceived rather than malicious, and attacking them converts them into adversaries. And do not, under any circumstances, use synthetic media in the response, for example by generating a synthetic apology or a synthetic recreation. An institution whose credibility rests on the distinction between real and fabricated content cannot afford to blur that line for its own convenience. 7.4 Phase Four: Absorb The final phase is the one most often skipped, because relief follows the end of an acute incident and attention moves on. Yet the psychology of #continued_influence means that the reputational damage of a synthetic media attack persists after the acute phase closes. Correction reinforcement. A single correction is insufficient. Vraga and Bode (2021) found that both preemptive and responsive correction have value and that repetition of the accurate account, from multiple sources, improves outcomes. The organisation should therefore restate the accurate account periodically, in different formats, through different voices, over the following weeks, without re raising the false claim in detail. Martel and Rand (2023) review the evidence on warning labels and find them broadly effective, which supports working with platforms to ensure that any remaining copies of the fabrication carry contextual labels. Search and archive management. Fabrications persist in search results, in archives, and in the training data of future models. The organisation should ensure that the authoritative account is well indexed, clearly dated, and easy to find, because in six months a stakeholder searching the topic must encounter the correction rather than the artefact. Transparency reporting. Publishing an account of what happened, how it was detected, and what was done builds the institution's authority as a source of #verification. This is the renewal move described by Sellnow and Seeger (2021), applied to a new problem. The organisation converts a defensive episode into a demonstration of competence. Institutional learning. The #after_action_review should ask what the attack exploited. Almost always, the answer includes a pre existing vulnerability: an unresolved controversy, a distrusted leader, a communication channel that stakeholders did not know how to verify, or an internal control that relied on voice recognition. Synthetic media attacks are opportunistic. They succeed where trust was already thin. The most durable defence against fabricated content is therefore an institution whose genuine conduct is consistent with its stated values, because a fabrication that contradicts a stakeholder's settled experience of an organisation is far less likely to be believed. This is not a moralistic aside. It is a structural finding, and it means that #reputation_resilience is built in ordinary time, not in crisis time. 8. Governance, Regulation, and Ethics 8.1 The regulatory environment The regulatory picture is developing quickly and varies by jurisdiction, so practitioners must check current requirements rather than rely on any single account. Broadly, three regulatory strands are relevant to institutional communication. The first is #transparency_obligation. Several jurisdictions have moved toward requiring that content generated or manipulated by artificial intelligence be disclosed as such, particularly where it depicts real persons or events. For institutions, the practical implication is that any use of synthetic media in their own communications, including synthetic presenters, generated imagery, or cloned voices in advertising, is likely to attract a labelling requirement. Organisations should build disclosure into their content workflow now rather than retrofit it later. The second is #platform_responsibility. Regulation in several markets now imposes duties on very large platforms to assess and mitigate systemic risks arising from disinformation, and to provide functioning notice and action mechanisms. This gives institutional complainants more leverage than they had five years ago, and it makes familiarity with formal reporting mechanisms a practical necessity. The third strand is the older body of law: defamation, impersonation, fraud, passing off, trademark, and the various rights that attach to a person's image and voice. Citron (2022) argues that existing legal frameworks have been slow to protect individuals against intimate and reputational harms in digital environments, and that remedies remain uneven. For institutions, the practical position is that legal remedies exist but are jurisdictionally fragmented and slow, which reinforces the argument for communicative and technical defences rather than reliance on law. 8.2 The ethics of institutional synthetic media use An organisation that campaigns against deceptive synthetic media while quietly using synthetic media in its own persuasion is in an untenable position. Three ethical rules follow. Disclose. Any synthetic element in institutional communication should be labelled, clearly and in a way the audience will actually notice. A dubbed synthetic voice in a training video is legitimate if disclosed and deceptive if not. Do not synthesise persons without consent. This includes deceased figures, employees, and, in particular, any depiction of a real person saying something they did not say. De Ruiter (2021) argues that the distinctive wrong of deepfakes lies in the violation of a person's ability to determine how they present themselves to others, a wrong that is committed even when the content is flattering and even when no third party is deceived. Never claim a genuine artefact is fake. The liar's dividend is available to any institution under pressure, and it is a form of corruption that destroys the shared evidential basis on which all institutions ultimately depend. Diakopoulos and Johnson (2021) and Gregory (2022) both emphasise that the systemic damage of casual denial exceeds the individual benefit gained from it. An institution that has invested in provenance and verification and then, on one bad day, denies a real recording, has burned the asset it built. #communication_ethics here is not a constraint on effectiveness. It is the precondition of it. 8.3 Coordination beyond the organisation No single institution can solve this problem alone. #sector_coordination, through professional associations, university networks, health authority coalitions, and industry bodies, allows for shared verification resources, shared threat intelligence, and shared public education. Gregory (2022) argues for exactly this kind of collective infrastructure, noting that the burden of verification currently falls hardest on those with the fewest resources. Smaller institutions, community organisations, and those operating in languages with less detection tool coverage are disproportionately exposed, and a purely organisation by organisation approach will leave them undefended. 9. Measurement and Evaluation A protocol that cannot be measured cannot be improved. Evaluation should occur at three levels. Preparedness indicators, measured in ordinary time. These include: the proportion of official content carrying provenance data; stakeholder awareness of the authenticated channel of record, measured by survey; the proportion of high risk transactions protected by out of band verification; the time from simulated detection to simulated first statement in exercises; and the existence of tested contacts at each major platform. These are #leading_indicators, and they are the only ones an organisation can improve before an attack. Response indicators, measured during an incident. Time to detection, time to internal ground truth, time to first public statement, time to platform removal, reach of the fabrication compared with reach of the correction, and the proportion of media coverage that reports the artefact as verified fake rather than as contested claim. This last measure is particularly important, because a story framed as "organisation denies video" is a defeat, while a story framed as "fabricated video targets organisation" is a win, and the difference is largely determined by whether the organisation supplied #checkable_evidence and independent confirmation quickly enough. Recovery indicators, measured afterwards. Belief in the false claim over time, measured by stakeholder survey; trust in the institution compared with pre incident baseline; verification behaviour, meaning the proportion of stakeholders who report checking the channel of record; and search visibility of the correction relative to the fabrication. Two cautions apply. First, an organisation cannot measure recovery without a baseline, which means #reputation_monitoring must be continuous rather than activated at the moment of crisis. Second, absence of a detected attack is not evidence of security. It may mean the organisation is not monitoring effectively. Detection capability should itself be tested, for example through authorised exercises in which benign synthetic content is released and the monitoring function is assessed on whether it notices. 10. Application Across Sectors The protocol is general, but its emphasis shifts by sector. Universities and schools. Educational institutions are unusually exposed because they have many quasi official voices, weak central control of communication, high volumes of public footage of academic staff, and an audience of students who are heavy users of the platforms where fabrications spread. The most common attacks are fabricated administrative announcements, impersonation of academic staff, fabricated examination material, and synthetic harassment of individual students and staff. The priorities for a university are a clear channel of record for official announcements, a rapid response route for individual staff and students who are targeted personally, and student facing #media_literacy education, which universities are uniquely placed to deliver and which Guess et al. (2020) show can measurably improve discernment between accurate and false content. Government agencies and public safety bodies. The risk here is operational as well as reputational. A fabricated evacuation order, a false public health instruction, or a fabricated statement about an election process can cause direct physical harm. The priorities are #authenticated_emergency_channels, standing relationships with platforms for expedited removal, prebunking before predictable high risk periods such as elections and disasters, and coordination with news media on verification protocols. Diakopoulos and Johnson (2021) discuss the specific ethical and operational demands of the electoral context, and Dobber et al. (2021) show that microtargeted political deepfakes can shift attitudes toward the depicted figure under some conditions. Health institutions. Health bodies face #counterfeit_endorsement, fabricated safety alerts, and synthetic testimony about treatment harm. Trust is the operative currency and it is easily damaged. Van der Meer and Jin (2020) examined misinformation treatment in public health crises and found that both the source and the strategy of correction matter, with rebuttal from a credible source outperforming other approaches. The priority for health institutions is a pre established, publicised verification pathway, so that a patient encountering an alarming claim has a reflexive habit of checking a known source. Corporations. The dominant risks are executive impersonation for #financial_fraud, fabricated product safety claims, market manipulation through fabricated announcements, and synthetic review campaigns. Mustak et al. (2023) provide the most useful business focused synthesis and stress the integration of communication, security, and legal responses. Corporate priorities include out of band verification for financial authorisation, coordination with investor relations and regulators, and pre agreed escalation with exchanges where fabricated announcements could move a share price. Civil society and non governmental organisations. These bodies are frequently the target of state aligned disinformation and are the least resourced to respond. Their most valuable defences are collective: shared verification services, coalition responses, and relationships with independent forensic and fact checking organisations. Gregory (2022) documents the resource gap and argues that verification infrastructure should be treated as a public good rather than left to individual organisational capacity. 11. Discussion 11.1 Theoretical implications The framework proposed here suggests three amendments to crisis communication theory. First, existing typologies of crisis type should be extended to include an evidentiary crisis category, in which the primary contest is over whether the triggering event occurred rather than over who is responsible for it. Current typologies sort crises by attributed responsibility and implicitly assume shared facts. Synthetic media breaks that assumption. Second, the theory of response strategy selection needs to accommodate the credibility paradox of denial. SCCT correctly predicts that denial damages reputation when it is used inappropriately, but it does not model the case where denial is both accurate and necessary. The evidence reviewed here suggests that denial can be made credible through three mechanisms that the classical model does not address: independent third party confirmation, checkable provenance, and pre established verification channels. These are structural rather than rhetorical solutions, and their inclusion suggests that crisis communication theory must extend into infrastructure design. Third, the field should treat preparation as a communicative act, not merely as a management process. A provenance system, an authenticated channel, and a prebunking campaign are not administrative preliminaries to communication. They are communication, delivered in advance, and they determine the credibility of everything said afterwards. This is a meaningful theoretical shift, because crisis communication scholarship has traditionally located its object of study in the response phase. 11.2 Practical implications For practitioners, the central message is that the work must be done before the attack. The organisation that begins thinking about deepfakes when a deepfake appears will lose. The specific investments with the highest return are, in order: an authenticated channel of record that stakeholders actually know about; provenance data on all official media; out of band verification for high risk transactions; a named decision authority with a defined time limit; a pre agreed diffusion threshold; and one realistic simulation per year. None of these requires exotic technology. All of them require organisational decision. 11.3 Limitations This article has clear limitations. It is a conceptual synthesis, not an empirical test, and the protocol has not been validated against outcome data. The evidence base it draws on is itself uneven: the psychology of correction and inoculation is well supported by experimental work, while the specific effects of synthetic media on institutional reputation have been studied far less than the effects on political attitudes. Much of the deepfake literature concerns political and individual harms rather than organisational ones, and generalising from the former to the latter carries risk. The technological environment is also moving faster than the research cycle. Findings about detection accuracy, human perception, and the realism of synthetic content date quickly, and any specific claim about what audiences can or cannot detect should be treated as provisional. The protocol is designed to be robust to this, because it relies on authentication of originals rather than detection of fakes, and authentication does not degrade as generation improves. That design choice is deliberate, and it is the article's main practical bet. Finally, the framework is written from a perspective that assumes institutions are, in the main, legitimate actors defending themselves against illegitimate attack. That assumption does not always hold. The same protocol could be used by a bad faith institution to defend itself against genuine evidence of wrongdoing. This is the liar's dividend problem in its sharpest form, and it cannot be solved from within the institution. It requires independent journalism, independent forensic capacity, and legal accountability. Communication scholars should be candid that the tools they develop for institutional defence are also available for institutional evasion. 12. A Research Agenda Several propositions follow from the framework and can be tested. The first concerns pre established verification channels. The framework predicts that stakeholders who are aware, before an incident, of an organisation's authenticated channel of record will be significantly less likely to believe a fabricated communication and significantly more likely to accept the organisation's denial. This is directly testable through experimental design and has not, to the author's knowledge, been examined. The second concerns the credibility of denial under different evidence conditions. Experimental work should compare bare denial, denial supported by internal evidence, denial supported by technical forensic evidence, and denial supported by independent third party confirmation, measuring belief and reputational outcomes. The framework predicts that third party confirmation dominates. The third concerns the amplification threshold. Field or simulation research should examine the conditions under which public rebuttal increases rather than decreases the reach and belief of a fabrication. Practitioners currently make this judgement intuitively, and the literature offers little guidance. The fourth concerns the liar's dividend in organisational settings. Most existing work on this phenomenon is political. Research should examine how stakeholders respond when an organisation claims that genuine evidence is synthetic, including how quickly and how permanently that claim damages the organisation if it is disproved. The fifth concerns occupational and cultural variation. Nearly all of the experimental literature on deepfake perception has been conducted in a narrow range of countries, languages, and media environments. Detection tools also perform unevenly across languages. #comparative_research is needed, and it is needed most urgently in the environments where verification infrastructure is weakest. The sixth concerns the long term erosion effect. Twomey et al. (2023) suggest that deepfake exposure generalises into broader distrust of video evidence. #longitudinal_research should examine whether repeated institutional deepfake incidents produce a durable decline in the evidential authority of recorded media, and whether verification education mitigates or accelerates that decline. The finding by Ternovski et al. (2022) that awareness campaigns can reduce trust in genuine content makes this the most urgent open question in the field, because it implies that the most obvious intervention may make the problem worse if it is designed badly. 13. Conclusion Synthetic media has changed the terms on which institutions defend their reputations. The change is not that lies have become possible, since lies have always been possible. The change is that the evidence which once ended a lie can now be manufactured, and the evidence which once proved a truth can now be denied. Institutions that continue to plan for the old environment, in which facts could be established and then communicated, will find that they lose the first battle before they have understood that it started. The argument of this article is that the correct response is not to invest primarily in detecting fakes. Detection is a technical arms race that communication departments cannot win and should not be asked to fight. The correct response is to invest in #authentication: to make the institution's genuine communication verifiable, to tell stakeholders repeatedly and clearly how to check it, to protect high risk decisions from voice and video authorisation, to build relationships with the platforms and verifiers who will be needed, and to rehearse the decision chain until it is automatic. An organisation that has done these things enters a #synthetic_media_crisis with something to point to. An organisation that has not entered it with nothing but its word, at precisely the moment when its word is worth least. The four phases set out here, Anticipate, Authenticate, Answer, and Absorb, are a way of organising that work. They are not a formula, and no protocol can substitute for judgement under pressure. But the evidence reviewed in this article supports a clear conclusion, and it is one that students of #public_relations should carry into practice: in the era of synthetic media, the credibility of what an institution says in a crisis is determined almost entirely by what it built before the crisis began. #crisis_preparedness is no longer a chapter at the back of the manual. It is the whole of the defence. Hashtags: #CrisisCommunication #Deepfakes #SyntheticMedia #AIMisinformation #PublicRelations #ReputationManagement #StrategicCommunication #MediaLiteracy #ContentProvenance #Prebunking #Disinformation #GenerativeAI #InstitutionalTrust #CommunicationResearch #MediaForensics References Ahmed, S. (2021). Fooled by the fakes: Cognitive differences in perceived claim accuracy and sharing intention of non political deepfakes. Personality and Individual Differences, 182, 111074. DOI: 10.1016/j.paid.2021.111074 Appel, M., and Prietzel, F. (2022). The detection of political deepfakes. Journal of Computer Mediated Communication, 27(4), zmac008. DOI: 10.1093/jcmc/zmac008 Bennett, W. L., and Livingston, S. (Eds.). (2020). The Disinformation Age: Politics, Technology, and Disruptive Communication in the United States. Cambridge University Press. Chan, M. S., and Albarracin, D. (2023). A meta analysis of correction effects in science relevant misinformation. Nature Human Behaviour, 7(9), 1514-1525. DOI: 10.1038/s41562-023-01623-8 Cheng, Y., and Chen, Z. F. (2020). The influence of presumed fake news influence: Examining public support for corporate corrective response, media literacy interventions, and governmental regulation. Mass Communication and Society, 23(5), 705-729. DOI: 10.1080/15205436.2020.1750656 Citron, D. K. (2022). The Fight for Privacy: Protecting Dignity, Identity, and Love in the Digital Age. W. W. Norton. Claeys, A. S., and Coombs, W. T. (2020). Organizational crisis communication: Suboptimal crisis response selection decisions and behavioral economics. Communication Theory, 30(3), 290-309. DOI: 10.1093/ct/qtz002 Coombs, W. T. (2019). Ongoing Crisis Communication: Planning, Managing, and Responding (5th ed.). SAGE Publications. de Ruiter, A. (2021). The distinct wrong of deepfakes. Philosophy and Technology, 34(4), 1311-1332. DOI: 10.1007/s13347-021-00459-2 Diakopoulos, N., and Johnson, D. (2021). Anticipating and addressing the ethical implications of deepfakes in the context of elections. New Media and Society, 23(7), 2072-2098. DOI: 10.1177/1461444820925811 Dobber, T., Metoui, N., Trilling, D., Helberger, N., and de Vreese, C. (2021). Do (microtargeted) deepfakes have real effects on political attitudes? The International Journal of Press/Politics, 26(1), 69-91. DOI: 10.1177/1940161220944364 Ecker, U. K. H., Lewandowsky, S., Cook, J., Schmid, P., Fazio, L. K., Brashier, N., Kendeou, P., Vraga, E. K., and Amazeen, M. A. (2022). The psychological drivers of misinformation belief and its resistance to correction. Nature Reviews Psychology, 1(1), 13-29. DOI: 10.1038/s44159-021-00006-y Fallis, D. (2021). The epistemic threat of deepfakes. Philosophy and Technology, 34(4), 623-643. DOI: 10.1007/s13347-020-00419-2 Farid, H. (2022). Creating, using, misusing, and detecting deep fakes. Journal of Online Trust and Safety, 1(4). DOI: 10.54501/jots.v1i4.56 Ferrara, E. (2024). GenAI against humanity: Nefarious applications of generative artificial intelligence and large language models. Journal of Computational Social Science, 7, 549-569. DOI: 10.1007/s42001-024-00250-1 Gosse, C., and Burkell, J. (2020). Politics and porn: How news media characterizes problems presented by deepfakes. Critical Studies in Media Communication, 37(5), 497-511. DOI: 10.1080/15295036.2020.1832697 Gregory, S. (2022). Fortify the truth: How to defend human rights in an age of deepfakes and generative AI. Journal of Human Rights Practice, 14(3), 702-714. DOI: 10.1093/jhuman/huac059 Groh, M., Epstein, Z., Firestone, C., and Picard, R. (2022). Deepfake detection by human crowds, machines, and machine informed crowds. Proceedings of the National Academy of Sciences, 119(1), e2110013119. DOI: 10.1073/pnas.2110013119 Guess, A. M., Lerner, M., Lyons, B., Montgomery, J. M., Nyhan, B., Reifler, J., and Sircar, N. (2020). A digital media literacy intervention increases discernment between mainstream and false news in the United States and India. Proceedings of the National Academy of Sciences, 117(27), 15536-15545. DOI: 10.1073/pnas.1920498117 Hameleers, M., van der Meer, T. G. L. A., and Dobber, T. (2022). You won't believe what they just said! The effects of political deepfakes embedded as vox populi on social media. Social Media plus Society, 8(3). DOI: 10.1177/20563051221116346 Hancock, J. T., and Bailenson, J. N. (2021). The social impact of deepfakes. Cyberpsychology, Behavior, and Social Networking, 24(3), 149-152. DOI: 10.1089/cyber.2021.29208.jth Helmus, T. C. (2022). Artificial Intelligence, Deepfakes, and Disinformation: A Primer. RAND Corporation. Jin, Y., van der Meer, T. G. L. A., Lee, Y. I., and Lu, X. (2020). The effects of corrective communication and employee backup on the effectiveness of fighting crisis misinformation. Public Relations Review, 46(3), 101910. DOI: 10.1016/j.pubrev.2020.101910 Karnouskos, S. (2020). Artificial intelligence in digital media: The era of deepfakes. IEEE Transactions on Technology and Society, 1(3), 138-147. DOI: 10.1109/TTS.2020.3001312 Kietzmann, J., Lee, L. W., McCarthy, I. P., and Kietzmann, T. C. (2020). Deepfakes: Trick or treat? Business Horizons, 63(2), 135-146. DOI: 10.1016/j.bushor.2019.11.006 Kobis, N. C., Dolezalova, B., and Soraperra, I. (2021). Fooled twice: People cannot detect deepfakes but think they can. iScience, 24(11), 103364. DOI: 10.1016/j.isci.2021.103364 Leibowicz, C. R., McGregor, S., and Ovadya, A. (2021). The deepfake detection dilemma: A multistakeholder exploration of adversarial advantage through trust. Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 736-744. DOI: 10.1145/3461702.3462584 Lewandowsky, S., and van der Linden, S. (2021). Countering misinformation and fake news through inoculation and prebunking. European Review of Social Psychology, 32(2), 348-384. DOI: 10.1080/10463283.2021.1876983 Martel, C., and Rand, D. G. (2023). Misinformation warning labels are widely effective: A review of warning effects and their moderating features. Current Opinion in Psychology, 54, 101710. DOI: 10.1016/j.copsyc.2023.101710 Murphy, G., and Flynn, E. (2022). Deepfake false memories. Memory, 30(4), 480-492. DOI: 10.1080/09658211.2021.1919715 Mustak, M., Salminen, J., Mantymaki, M., Rahman, A., and Dwivedi, Y. K. (2023). Deepfakes: Deceptions, mitigations, and opportunities. Journal of Business Research, 154, 113368. DOI: 10.1016/j.jbusres.2022.113368 Nightingale, S. J., and Farid, H. (2022). AI synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences, 119(8), e2120481119. DOI: 10.1073/pnas.2120481119 Pennycook, G., and Rand, D. G. (2021). The psychology of fake news. Trends in Cognitive Sciences, 25(5), 388-402. DOI: 10.1016/j.tics.2021.02.007 Rini, R. (2020). Deepfakes and the epistemic backstop. Philosophers' Imprint, 20(24), 1-16. Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., and Lewandowsky, S. (2022). Psychological inoculation improves resilience against misinformation on social media. Science Advances, 8(34), eabo6254. DOI: 10.1126/sciadv.abo6254 Sellnow, T. L., and Seeger, M. W. (2021). Theorizing Crisis Communication (2nd ed.). Wiley Blackwell. Shin, S. Y., and Lee, J. (2022). The effect of deepfake video on news credibility and corrective influence of cost based knowledge about deepfakes. Digital Journalism, 10(3), 412-432. DOI: 10.1080/21670811.2022.2026797 Sundar, S. S., Molina, M. D., and Cho, E. (2021). Seeing is believing: Is video modality more powerful in spreading fake news via online messaging apps? Journal of Computer Mediated Communication, 26(6), 301-319. DOI: 10.1093/jcmc/zmab010 Ternovski, J., Kalla, J., and Aronow, P. (2022). The negative consequences of informing voters about deepfakes: Evidence from two survey experiments. Journal of Online Trust and Safety, 1(2). DOI: 10.54501/jots.v1i2.28 Traberg, C. S., Roozenbeek, J., and van der Linden, S. (2022). Psychological inoculation against misinformation: Current evidence and future directions. The ANNALS of the American Academy of Political and Social Science, 700(1), 136-151. DOI: 10.1177/00027162221087936 Twomey, J., Ching, D., Aylett, M. P., Quayle, M., Linehan, C., and Murphy, G. (2023). Do deepfake videos undermine our epistemic trust? A thematic analysis of tweets that supplement or disparage a deepfake video. PLoS ONE, 18(10), e0291668. DOI: 10.1371/journal.pone.0291668 Vaccari, C., and Chadwick, A. (2020). Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news. Social Media plus Society, 6(1). DOI: 10.1177/2056305120903408 van der Linden, S. (2022). Misinformation: Susceptibility, spread, and interventions to immunize the public. Nature Medicine, 28(3), 460-467. DOI: 10.1038/s41591-022-01713-6 van der Meer, T. G. L. A., and Jin, Y. (2020). Seeking formula for misinformation treatment in public health crises: The effects of corrective information type and source. Health Communication, 35(5), 560-575. DOI: 10.1080/10410236.2019.1573295 Vizoso, A., Vaz Alvarez, M., and Lopez Garcia, X. (2021). Fighting deepfakes: Media and internet giants' converging and diverging strategies against hi tech misinformation. Media and Communication, 9(1), 291-300. DOI: 10.17645/mac.v9i1.3494 Vraga, E. K., and Bode, L. (2021). Addressing COVID 19 misinformation on social media preemptively and responsively. Emerging Infectious Diseases, 27(2), 396-403. DOI: 10.3201/eid2702.203139 Walter, N., and Tukachinsky, R. (2020). A meta analytic examination of the continued influence of misinformation in the face of correction: How powerful is it, why does it happen, and how to stop it? Communication Research, 47(2), 155-177. DOI: 10.1177/0093650219854600 Weikmann, T., and Lecheler, S. (2023). Visual disinformation in a digital age: A literature synthesis and research agenda. New Media and Society, 25(12), 3696-3713. DOI: 10.1177/14614448221141648 Whittaker, L., Letheren, K., and Mulcahy, R. (2021). The rise of deepfakes: A conceptual framework and research agenda for marketing. Australasian Marketing Journal, 29(3), 204-214. DOI: 10.1177/1839334921999479 Zerback, T., Topfl, F., and Knopfle, M. (2021). The disconcerting potential of online disinformation: Persuasive effects of astroturfing comments and three strategies for inoculation against them. New Media and Society, 23(5), 1080-1098. DOI: 10.1177/1461444820908530

Global organisations now depend on concentrated clusters of internationally mobile professionals who move often, speak several languages, and hold weak long term ties to any single office. Traditional leadership models were built for stable, co-located and mostly monolingual teams, so they explain only part of what happens inside these clusters. This article proposes an ecological reading of executive behaviour. It borrows four organising ideas from #behavioral_ecology in the life sciences, namely niche use, signalling, territory and dispersal, and applies them, as analytical metaphors rather than biological claims, to the work of the #modern_executive who leads mobile, multilingual talent. The study is an integrative conceptual review. It synthesises recent scholarship on global mobility, global talent management, language in international human resource management, and virtual and paradoxical leadership, and organises that scholarship into a single framework with eight testable propositions. The analysis suggests that effective leaders in these settings behave less like commanders of a fixed unit and more like stewards of a shifting habitat. They design the environment, manage the meaning of signals, reduce the cost of linguistic difference, and treat departure as a normal event rather than a failure. Practical implications are drawn for the design of #leadership_paradigms, language policy, mobility policy and performance systems. The article closes with a research agenda for students and early career scholars who wish to test the framework empirically. Keywords: global mobility; expatriate management; multilingual teams; leadership; talent hubs; international human resource management; conceptual framework 1. Introduction Across finance, energy, logistics, consulting, technology and higher education, a familiar organisational form has appeared. A firm concentrates a large share of its regional capability in one city, staffs it with people drawn from twenty or thirty countries, and accepts that a meaningful proportion of those people will leave within three to five years. The office is stable. The population inside it is not. Scholars have begun to describe such settings as #talent_hubs, and they raise a leadership problem that older theory did not anticipate. The problem is not simply cultural difference. It is the combination of three conditions that rarely appeared together in earlier decades. The first is high turnover by design. Many workers in these hubs are on fixed term contracts, sponsored visas or project assignments, and their exit is planned from the start. The second is deep #linguistic_diversity. A single meeting may contain four or five mother tongues, one working language, and several levels of confidence in that working language. The third is the fluidity of the category of "#expatriate_talent" itself. Reviews of #global_mobility research show that the field now includes assigned expatriates, self initiated movers, commuters, short term assignees, international business travellers and virtual assignees, and that these groups differ in motivation, contract and attachment (Bonache, Brewster and Froese, 2021; Caligiuri, Collings, De Cieri and Lazarova, 2024). Most leadership theory that managers learn in business schools was developed for a different animal. Transformational, transactional and servant models assume a leader and a follower who will remain in relationship long enough for influence, identification and reciprocity to accumulate. In a hub where a quarter of the team turns over every year, that assumption is fragile. Vision statements do not have time to sink in. Trust must be built quickly, then rebuilt. Knowledge leaves the building regularly. Managers report that they are running a highly skilled organisation whose members are, in an important sense, passing through. This article argues that the language of ecology helps. In the study of animal behaviour, researchers do not begin by asking what a creature intends. They begin by asking what the environment rewards, what it punishes, what resources are scarce, how information travels, and how individuals decide whether to stay or leave. Translating that habit of thought into management does not require any claim that people are animals or that behaviour is genetically fixed. It requires only a shift of attention, from the personality of the leader to the structure of the setting in which leadership occurs. A leader in a talent hub is, in this reading, a manager of habitat. The article makes three contributions. First, it offers a conceptual vocabulary. It shows how four ecological ideas, namely #niche_construction, #signaling_theory, #territoriality and #dispersal, can be redescribed as management constructs and used to organise findings that are currently scattered across separate literatures. Second, it integrates evidence. Research on language in international human resource management, on expatriate adjustment, on global talent management and on virtual leadership has grown quickly but rarely speaks to itself. This article brings the strands together and identifies where they agree, where they conflict and where they are silent. Third, it produces propositions. Eight propositions are stated in a form that students and researchers can operationalise, measure and challenge. The intention is not to close the topic but to give it a testable shape. The remainder of the article proceeds as follows. Section 2 develops the ecological lens. Section 3 reviews the relevant literature. Section 4 presents the conceptual framework and propositions. Section 5 explains the review method. Section 6 analyses the four ecological domains in detail. Section 7 discusses theoretical, practical and policy implications. Section 8 sets out limitations and a research agenda. Section 9 concludes. A note on scope is necessary. This is a conceptual paper. It does not report new survey or interview data. Its claims are claims about how existing evidence can be organised, and they should be read as invitations to empirical work rather than as findings. 2. Conceptual Background: Ecology as an Analytical Lens 2.1 Why an ecological vocabulary Organisational scholars have long borrowed from biology. Population ecology, resource dependence and evolutionary economics all use metaphors of variation, selection and retention. What is proposed here is narrower and, for that reason, more useful. Rather than treating whole organisations as populations, the article treats the internal environment of a talent hub as a habitat and treats managerial behaviour as adaptation to that habitat. Four cautions apply. The metaphor is a lens, not a law. It does not naturalise inequality, and it must not be used to excuse exploitation by calling it competition. It does not assume that behaviour is fixed. And it must be tested, not merely asserted. With those cautions in place, the lens earns its keep because it directs attention to variables that conventional leadership models under-specify, namely the distribution of resources, the reliability of information, the cost of movement and the structure of the local environment. 2.2 Niche and habitat In ecology, a niche is the set of conditions under which an organism can survive and reproduce. Individuals do not merely occupy niches. They also modify them, a process known as #niche_construction. Beavers build dams. Ants farm fungus. The environment is partly made by those who live in it. The management translation is direct. Every professional in a hub occupies a role niche defined by task, network position, language repertoire, visa status and career horizon. Some of those conditions are given by the organisation. Others are constructed by the individual. A mobile professional who deliberately becomes the person who understands both the regional regulator and the head office finance system has built a niche that did not previously exist. The leadership implication is that a manager who only allocates roles is working at half strength. A manager who shapes the conditions under which useful niches can form is working with the grain of the system. This reframes the executive task as one of #ecosystem_stewardship, that is, the deliberate design of an internal environment in which the behaviours the organisation needs are also the behaviours that pay. Two further constructs follow. The first is #carrying_capacity, the number of people a given structure can support before crowding degrades performance. Hubs frequently exceed it. When six senior managers compete for two regional roles, the resulting behaviour is not a character flaw. It is a predictable response to a scarce #resource_patches distribution. The second is the #fitness_landscape, the map of which behaviours are actually rewarded. Employees read that map far more accurately than they read the mission statement, and they respond to what is rewarded rather than what is announced. 2.3 Signals and information Animals face a persistent problem. They must judge the quality and intention of others on limited information. The theoretical solution, developed in evolutionary biology and later imported into economics, is #signaling_theory. Signals that are cheap to fake are ignored. Signals that are costly, and therefore hard to fake, carry information. Talent hubs are saturated with this problem. New colleagues arrive constantly. There is no long shared history against which to judge them. Competence must be inferred from cues. And here language enters, because in a multilingual workplace the most available cue is fluency in the working language, which is not the same thing as competence. Research on language in international human resource management shows that language skills influence recruitment, promotion, compensation and career mobility, and that they can shape who is heard and who is believed (Presbitero, Froese, Peltokorpi, Pudelko and Tenzer, 2023). This creates a systematic distortion. The fluent are read as capable, and the less fluent are read as less capable, regardless of underlying skill. #costly_signals in these environments include visible technical output, a track record documented in artefacts rather than in conversation, and the willingness to do work that a low performer would not undertake. A central task of the executive is therefore to redesign the signalling environment so that the loudest signal is not simply the most fluent one. 2.4 Territory and status Territorial behaviour appears wherever a resource is valuable, defensible and scarce. In organisations, the resources are budget, headcount, access to decision makers, control of information and proximity to the parent company. #territoriality is therefore not a personality defect of difficult managers. It is a response to structure. Ecological thinking also predicts the formation of stable #status_hierarchies, which reduce the cost of constant conflict but entrench advantage. In expatriate settings, hierarchy often maps onto nationality and passport. Ethnographic work in one of the world's most prominent mobility destinations describes distinct #expatriate_bubbles in which Western expatriates, expatriates from Asian and other Eastern origins, and local nationals occupy separate social worlds that interact at work but rarely mix on equal terms (Primecz, 2023). Related research examines how these bubbles form and how difficult it is for individuals to leave them (van Bakel and Vance, 2023). For the leader, the lesson is uncomfortable but useful. Segregation in a diverse hub is usually not caused by prejudice alone. It is sustained by the ordinary costs of interaction, namely language effort, uncertainty about norms and the risk of embarrassment. Reducing those costs is a design problem before it is a values problem. 2.5 Dispersal, philopatry and the decision to leave Ecology treats movement as a decision with costs and benefits. #dispersal is the departure of an individual from the place where it currently lives. #philopatry is the tendency to remain. Both are strategies, and neither is a moral category. Whether an individual disperses depends on local resource density, competition, the perceived quality of alternatives and the cost of the move itself. This is the single most valuable transfer for the study of mobile talent. Organisations habitually treat resignation as failure, as disloyalty, or as evidence that engagement scores were mismeasured. An ecological reading treats it as a rational response to the local reward structure and to information about better patches elsewhere. Reviews of global talent management note that careers are now less organisational and more individual, and that mobility is increasingly initiated by workers rather than by employers (Caligiuri et al., 2024; Machado, 2022). Self initiated expatriation has become a large field in its own right, precisely because so much movement is no longer arranged by a head office (Andresen, Brewster and Suutari, 2021). The managerial consequence is significant. If departure is normal, then #retention cannot be the only measure of success, and a leadership model built entirely on loyalty will systematically underperform. 3. Literature Review 3.1 The changing shape of global mobility The classic expatriate of the twentieth century was sent by a parent company, was accompanied by a family, was paid a premium, and was expected to return. That figure still exists, but the field has widened considerably. Current scholarship treats mobility as a family of arrangements that includes #assigned_expatriates, #self_initiated_expatriates, #international_business_travelers, short term assignees, commuters, inpatriates and, more recently, remote workers who cross borders without being posted anywhere (Bonache et al., 2021). Two developments deserve emphasis. The first is #virtual_global_mobility, the practice of moving work to people rather than people to work. This was accelerated by the pandemic and has not reversed, and research on the transition from enforced homeworking to voluntary work from anywhere documents its persistence (Smite, Moe, Klotins and Gonzalez-Huerta, 2023). The second is the rise of #digital_nomads, who work remotely while relocating repeatedly, and who now attract attention from management, migration and tourism scholars alike (Kozak, Cetin and Alrawadieh, 2024; Mancinelli and Germann Molz, 2023). Nomads sit awkwardly in existing categories because they are neither posted nor settled, and their relationship to any employer is often thin. These developments matter for leadership because they break the link between physical presence and organisational membership. A hub is no longer only the people in the building. 3.2 Global talent management Global talent management concerns the identification of pivotal positions, the development of pools of high potential people who reflect the global scope of the firm, and the design of human resource architecture to fill those positions. Recent critical review argues that the field must now be rethought for a new organisational reality in which work is distributed, careers are boundaryless and the assumptions of the old assignment model no longer hold (Caligiuri et al., 2024). Multilevel reviews similarly note that talent practices operate simultaneously at individual, unit and organisational levels, and that studies too often examine only one (Aljbour, French and Ali, 2022). Two tensions recur in this body of work. The first concerns exclusivity. Firms that concentrate investment in a small elite risk demotivating the majority, and inclusive approaches to talent development have been advanced as an alternative (Kaliannan, Darmalinggam, Dorasamy and Abraham, 2023). The second concerns implementation. High potential programmes are frequently undermined by the everyday practices of the managers who must run them, producing tensions between corporate design and local reality (Garavan, Morley, Cross, Carbery and Darcy, 2021). Both tensions intensify in hubs, where the population is diverse, the tenure is short and the audience for any perceived unfairness is immediate. The construct of #cultural_agility, understood as the capacity to vary one's response appropriately across contexts, has been proposed as a competence that international experience can build when that experience is well designed (Caligiuri and Caprar, 2023). Alongside it, #cultural_intelligence remains the most extensively validated framework for individual cross cultural capability, and it continues to be refined and extended (Richter, Schlaegel, Taras, Alon and Bird, 2023; Semenov and Randrianasolo, 2024). 3.3 Language, power and inclusion Language is the most under-theorised variable in mainstream leadership research and one of the most consequential in multilingual hubs. A multi-paradigmatic review of language sensitive research in international business demonstrates how varied the field has become, spanning functionalist studies of communication efficiency and critical studies of power and identity (Piekkari, Gaibrois and Johansson, 2022). Several findings are robust enough to build on. A common corporate #language_policy, usually the adoption of English as a #lingua_franca, solves a coordination problem and creates a distribution problem. It reduces translation cost while redistributing influence towards those who are comfortable in the chosen language. Language shapes careers. Evidence links language competence to hiring, advancement and pay, which means that #language_sensitive_recruitment is not a neutral filter (Presbitero et al., 2023). Language can exclude. Work on equality, diversity and inclusion in multinational companies argues explicitly that language can operate as a mechanism of exclusion and that #equity_diversity_inclusion agendas must therefore incorporate the management of linguistic difference rather than treating it as a technical matter (Ciuk, Sliwa and Harzing, 2023). Language affects team process. Studies of multicultural virtual teams show that language diversity interacts with media choice to shape performance, so the same team can function well or badly depending on the channel it uses (Eisenberg, Glikson and Lisak, 2021). Qualitative work in multinational teams documents the emotional load carried by non native speakers and the strategies used to bridge competence gaps (Weinzierl, 2025). Two related phenomena deserve names. #language_ostracism refers to the experience of being excluded from conversation and therefore from information, whether or not exclusion is intended. #code_switching, the movement between languages within an interaction, may serve solidarity for some participants while producing anxiety and suspicion in others. Finally, language competence functions as a form of capital that migrants and mobile workers convert, with difficulty, into social and economic advantage inside multinational firms (Peltokorpi and Xie, 2023). 3.4 Leadership paradigms under strain Leadership theory has responded to distributed work with several partially overlapping constructs. #e_leadership and virtual team leadership research has grown quickly. Reviews report that #transformational_leadership remains the most studied approach in virtual settings, with shared and transactional approaches also prominent, and that effectiveness depends heavily on trust, communication quality and technology use (Bagga, Gera and Haque, 2023; Batirlik, Gencer and Akkucuk, 2022). Broader analyses of the digital transformation of teams describe a shift from conventional to virtual forms that changes the demands placed on leaders (Vuchkovski, Zalaznik, Mitrega and Pfajfar, 2023). #paradoxical_leadership addresses the requirement to hold competing demands at once, such as control and autonomy, or standardisation and localisation. A meta-analytic review reports positive associations between paradoxical leader behaviour and follower and team outcomes, with #psychological_safety and role clarity among the mechanisms, while also noting inconsistency in the evidence base (Lee, Lyubovnikova, Zheng and Li, 2023). Related work examines how such leaders cultivate a paradox mindset in followers (Boemelburg, Zimmermann and Palmie, 2023) and how paradoxical behaviour relates to knowledge sharing and hiding (Devi, 2024). #shared_leadership and #inclusive_leadership are frequently recommended for diverse teams, on the reasoning that distributing influence reduces the dominance of any single cultural or linguistic group. What the literature has not done is integrate these strands with the specific conditions of the multilingual, high turnover hub. Virtual leadership research often holds language constant. Language research often holds leadership constant. Talent management research often treats mobility as a variable to be managed rather than as the defining feature of the habitat. 3.5 The gap Three gaps follow from the review. First, a conceptual gap. There is no shared framework that connects the environmental conditions of a hub to the behaviour of the executive who leads it. Second, an integrative gap. Language, mobility, talent and leadership are studied separately although employees experience them together. Third, a normative gap. Existing prescriptions tend to assume that the goal is stability and retention, which is not obviously the right goal when high turnover is a structural feature rather than a defect. The framework developed next is an attempt to address these gaps. 4. Conceptual Framework and Propositions 4.1 The framework in outline The framework has four domains, each derived from the ecological ideas set out in Section 2, and each linked to a set of executive practices and expected outcomes. Domain 1: Habitat design. The executive shapes the internal environment, including the distribution of resources, the visibility of the reward structure and the opportunities for niche formation. Domain 2: Signal management. The executive shapes how competence, commitment and trustworthiness are demonstrated and read, with particular attention to the distortion introduced by language fluency. Domain 3: Boundary management. The executive shapes the cost of crossing internal boundaries, whether those boundaries are national, linguistic, functional or hierarchical. Domain 4: Flow management. The executive treats arrival and departure as continuous processes to be designed, rather than as exceptions to be minimised. Two cross cutting capabilities support all four domains. The first is #adaptive_capacity, the ability of the leader to change behaviour as conditions change. The second is #sensemaking, the ability to give a coherent and honest account of a confusing environment to people who have little shared history with each other. 4.2 Propositions Proposition 1. In talent hubs, the perceived clarity of the #fitness_landscape, meaning the degree to which employees can accurately predict what behaviour will be rewarded, will be more strongly associated with discretionary effort than the perceived inspirational quality of the leader. The reasoning is that in short tenure environments there is insufficient time for identification with a leader to develop, whereas the reward structure is legible almost immediately. Proposition 2. Where the working language is a second language for most members, the correlation between fluency and evaluated performance will be positive and larger than the correlation between fluency and objectively measured task output. This proposition operationalises the signalling distortion described in Section 2.3 and can be tested with paired supervisor ratings and objective output data. Proposition 3. Leader practices that shift evaluation towards artefacts, meaning written work, code, models and documented decisions, and away from real time verbal display will weaken the association described in Proposition 2. Proposition 4. The strength of #expatriate_bubbles within a hub will be positively associated with the interaction cost of cross group contact, and only weakly associated with measured intercultural attitudes. This proposition is deliberately provocative. It predicts that structure explains segregation better than sentiment does. Proposition 5. Territorial behaviour among managers will increase as the ratio of qualified candidates to senior positions increases, independent of individual personality measures. Proposition 6. Leaders who explicitly legitimate departure, for example by supporting alumni networks and by making exit interviews developmental rather than punitive, will experience lower rates of unplanned #turnover_intention among those who remain, and higher rates of knowledge handover from those who leave. The mechanism is that hidden exit planning consumes attention and suppresses candour, while legitimate exit planning frees both. Proposition 7. #paradoxical_leadership will be more strongly associated with team effectiveness in hubs with high linguistic and national diversity than in comparatively homogeneous units, because the number of competing demands the leader must hold is greater. Proposition 8. Investment in #boundary_spanning roles, meaning individuals formally rewarded for linking language groups, functions and locations, will predict #knowledge_transfer retention across turnover events better than investment in documentation systems alone. 4.3 Summary of the framework Stated plainly, the framework claims that in a mobile multilingual hub the leader's leverage lies in the environment rather than in the relationship. Relationships are short. The environment persists. A leader who improves the environment improves the performance of people they may never manage directly and may never meet again. 5. Method 5.1 Design This study is an integrative conceptual review. This design is appropriate when a field is fragmented across disciplines, when the object of study is changing faster than the theory that describes it, and when the purpose is to synthesise rather than to test. The output of an integrative review is a framework and a research agenda, not a statistical estimate. The review followed four stages. Stage one: scoping. The boundaries of the review were set to include four literatures, namely global mobility and expatriate management, global talent management, language in international business and human resource management, and leadership in virtual, diverse and paradoxical contexts. Stage two: identification. Peer reviewed articles, reviews and scholarly books were identified through database searching and through the reference lists of recent review articles, a technique sometimes called backward snowballing. Preference was given to work published in the last five years, and to review articles, meta-analyses and conceptual pieces, because these summarise a broader evidence base than single empirical studies. Stage three: selection. Sources were retained if they addressed at least one of the following, namely the behaviour or careers of internationally mobile workers, the management of linguistic difference in organisations, the design or evaluation of leadership in distributed or diverse settings, or the strategic management of talent across borders. Sources were excluded if they addressed migration or tourism without an organisational dimension, or if they addressed leadership without any international or distributed component. Stage four: synthesis. Retained sources were coded against the four ecological domains. Where a source spoke to more than one domain it was coded to each. Contradictions between sources were recorded rather than resolved, and they appear in Section 6 as tensions. 5.2 Quality and limitations of the method Three limitations should be stated openly. The review is interpretive. Another researcher, applying the same procedure, might organise the same sources differently. This is a known feature of integrative reviews rather than a defect that can be engineered away. The review is English language dominant. This is a particular irony in an article about #multilingual_workforce management, and it means that scholarship published in other languages is under-represented. The framework is untested. The propositions in Section 4 are derived from theory and from existing empirical findings, but they have not themselves been examined with new data. They are offered as hypotheses. 5.3 Ethical considerations The article uses ecological metaphors to describe human behaviour. There is a real risk in this move. Language about competition, hierarchy and dispersal can be used to make unjust arrangements appear natural. The position taken here is explicit. Describing a pattern is not endorsing it. Where the analysis identifies #power_asymmetries, for example between passport holders of different nationalities or between native and non native speakers, the appropriate response is to change the structure that produces the asymmetry, not to accept it as inevitable. 6. Analysis 6.1 Habitat design 6.1.1 The reward map is the real strategy Employees infer strategy from resource allocation. If promotions consistently go to those who are visible at headquarters, then presence at headquarters is the strategy, whatever the strategy document says. In hubs this inference happens rapidly because newcomers arrive with no loyalty and no history, and therefore examine the environment with fresh and sceptical eyes. The practical implication is that the executive should audit the reward landscape as carefully as they audit the budget. Useful questions include the following. Who was promoted in the last two years, and what did they have in common. What behaviour was quietly tolerated. Which contributions were invisible in the performance system. The answers describe the organisation more accurately than any published value statement. 6.1.2 Constructing niches on purpose Since individuals build niches whether or not managers permit it, the leader's choice is between accidental and deliberate niche construction. Deliberate construction means naming the roles the hub actually needs but has not formally created, for example the person who translates regulatory language into engineering requirements, or the person who can brief the parent company in its own idiom without distorting local reality. These roles carry a risk. They can become single points of failure, and the individuals who hold them accumulate informal power. The countermeasure is rotation and documentation, which brings the discussion back to #knowledge_transfer. 6.1.3 Crowding Hubs are often designed with more talent than the structure can absorb, because the labour market permits it and because leaders like optionality. The predictable result is that highly capable people compete for a small number of visible opportunities. Symptoms include information hoarding, credit contests and an unusual concentration of political behaviour. Reading this as an issue of #carrying_capacity suggests interventions that do not depend on changing anyone's character. These include creating genuinely differentiated career tracks, expanding the definition of advancement beyond line management, and, uncomfortably, hiring fewer people for roles that cannot grow. 6.2 Signal management 6.2.1 The fluency trap The central distortion in a multilingual hub is that fluency is a cheap and highly visible signal that is easily mistaken for competence. Research on language in international human resource management establishes that language ability shapes selection, promotion and pay outcomes, which is exactly what one would expect if fluency operates as a dominant signal (Presbitero et al., 2023). Related work shows that linguistic capital converts into social and economic advantage inside multinationals, and that this conversion is uneven (Peltokorpi and Xie, 2023). The consequences are not confined to individual careers. If the loudest voices in a design review are the fluent ones, then design decisions are being made on a partly linguistic basis, and the organisation is quietly discarding information. 6.2.2 Redesigning the signalling environment Four practices follow. The first is to move evaluation towards artefacts. Written proposals, prototypes, models and documented analysis allow contribution to be assessed without penalising accent or hesitation. The second is to change meeting protocol. Circulating material in advance, inviting written comments, allowing contributions in writing during discussion and slowing the pace of speech all reduce the advantage of fluency without lowering standards. Evidence that language diversity interacts with media choice in multicultural virtual teams supports the general point that channel design is a leadership decision, not an administrative one (Eisenberg et al., 2021). The third is to make competence signals costly in the informative sense. A signal that only a genuinely capable person could produce, such as a correct forecast, a working solution, or an accurate prediction of a problem, is worth more than confident speech. The fourth is to name the distortion aloud. Leaders who state plainly that hesitation in a second language is not evidence of weak thinking change the interpretive habits of the room. 6.2.3 Trust at speed Because tenure is short, #trust_building cannot rely on time. It must rely on reliability made visible. Small commitments, kept and seen to be kept, generate trust faster than large statements. Transparent decision rules do the same, because they allow people to verify that the process is what it claims to be. 6.3 Boundary management 6.3.1 Why bubbles form Ethnographic evidence from a major mobility destination describes distinct social worlds among Western expatriates, expatriates from other regions and local nationals, which coexist at work but seldom mix as equals (Primecz, 2023). Complementary work considers the difficulty of moving beyond such enclaves and connecting with local society (van Bakel and Vance, 2023). The ecological reading is that bubbles are low cost patches. Inside one's own language and cultural group, communication is cheap, norms are known and the risk of embarrassment is low. Outside it, every interaction costs effort. Absent a countervailing force, people will do what is cheap. This does not require anyone to be prejudiced. 6.3.2 Lowering the cost of crossing If cost is the mechanism, then cost reduction is the intervention. Structured cross group work, in which teams are formed deliberately rather than by self selection, removes the need for individuals to bear the social risk of initiating contact. Shared tasks with real stakes generate contact that has a purpose, which is more durable than contact organised for its own sake. Language support that is genuinely two directional matters. If only non native speakers are expected to invest in language learning, the organisation has already decided whose effort counts. Formal recognition of bridging work matters most of all. People who connect groups are performing a service that rarely appears in a job description and almost never appears in a bonus calculation. Where it is unrewarded, it is under-supplied. 6.3.3 The justice question #organizational_justice is not a peripheral concern in hubs. Where compensation, housing, schooling support and job security vary by nationality or visa class, employees perceive it, and the perception shapes behaviour. Scholarship on equality, diversity and inclusion in multinationals argues that these agendas must confront linguistic and structural exclusion directly rather than through general statements of values (Ciuk et al., 2023). The uncomfortable implication for executives is that some inequities are embedded in national #mobility_regimes and #visa_policy and cannot be removed by the firm. What the firm can do is refuse to add to them, be transparent about what is legally fixed and what is discretionary, and avoid the pretence that a hierarchy imposed by immigration law is a hierarchy of merit. 6.4 Flow management 6.4.1 Departure as a normal event If a hub is a habitat in which departure is expected, then a leadership model built on permanence is misaligned with reality. Reviews of global talent management describe a world in which individuals increasingly direct their own careers and in which mobility is frequently self initiated (Caligiuri et al., 2024; Machado, 2022; Andresen et al., 2021). Three reframings follow. Retention should be measured selectively rather than globally. The question is not how many people stayed but whether the people in pivotal positions stayed long enough, and whether their departure was orderly. Exit should be designed. A structured handover, a documented decision log and a genuine developmental conversation convert a loss into a transfer. Alumni are assets. In dense professional markets, former employees become clients, partners, regulators and referrers. Treating departure as betrayal destroys value that costs nothing to preserve. 6.4.2 Arrival as a repeated design problem Where turnover is high, onboarding is not an occasional human resource task but a core operating process. #cross_cultural_adjustment research has long shown that early experience shapes later performance, and the practical inference is that the first weeks deserve the same design attention as a product launch. Arrival design includes explicit instruction in the unwritten rules, early assignment of a task that produces a visible result, and deliberate introduction into more than one social group so that the newcomer does not default to the nearest bubble. 6.4.3 Memory that survives turnover The greatest technical risk in a high flow hub is the silent erosion of organisational memory. Documentation systems help but are insufficient, because much of what matters is tacit, contextual and relational. The ecological analogue is #social_learning, meaning the transmission of behaviour through observation and interaction. Practices that support it include deliberate overlap between departing and arriving staff, pairing across seniority, and forums in which decisions are explained rather than merely recorded. Proposition 8 predicts that these relational mechanisms will outperform documentation alone. 6.5 The executive as steward Pulling the four domains together, the emerging portrait of the effective leader in a mobile multilingual hub has the following features. They think in terms of environments rather than individuals, because the population changes and the environment does not. They accept contradiction. They must standardise and localise, control and release, plan for permanence and plan for departure. This is precisely the terrain of #paradoxical_leadership, and it is why meta-analytic evidence associating such behaviour with team outcomes is relevant here (Lee et al., 2023). They treat language as a leadership variable rather than as a technical or personnel matter. They distribute influence, because a single leader cannot hold the cultural and linguistic range that the hub contains. This is the practical case for #inclusive_leadership and #shared_leadership. They protect #wellbeing deliberately, because mobile professionals often lack local support networks, and because #work_life_balance in a hub is complicated by distance from family, time zone spread and the emotional labour of operating in a second language, a burden documented in qualitative research on multinational teams (Weinzierl, 2025). 6.6 Tensions in the evidence Honesty requires that disagreements be stated rather than smoothed over. The first tension concerns language policy. A single corporate language improves coordination and damages inclusion. There is no configuration that removes the trade off. What varies is who bears the cost. The second tension concerns talent exclusivity. Concentrating investment in a small pool is efficient and demotivating at the same time. Inclusive approaches address the second problem and dilute the first (Kaliannan et al., 2023). The third tension concerns virtuality. Distributed work widens the talent pool and weakens the informal learning that transmits tacit knowledge, a trade off visible in research on the digital transformation of teams (Vuchkovski et al., 2023; Smite et al., 2023). The fourth tension concerns paradoxical leadership itself. The meta-analytic evidence is encouraging but not uniform, and there are legitimate questions about whether behavioural inconsistency is read by followers as flexibility or as unpredictability (Lee et al., 2023). A framework that pretended to dissolve these tensions would be less useful than one that locates them. 7. Discussion 7.1 Theoretical implications The article contributes to theory in three ways. It reframes leadership as environmental design. Most leadership constructs are relational, describing what passes between a leader and a follower. The ecological reading adds a third term, namely the setting, and argues that in high turnover environments the setting carries more of the explanatory weight. It brings language into leadership theory as a structural variable rather than as a communication skill. If fluency distorts the reading of competence, then language is not a channel through which leadership passes. It is part of the mechanism by which status is allocated. It normalises mobility. By treating movement and settling as alternative strategies rather than as success and failure, the framework removes a moral overlay that has distorted both research and practice. #career_capital accumulated by individuals is not stolen from the firm, and #psychological_contract breach in these settings often arises from the pretence that permanence was ever on offer. 7.2 Practical implications for executives Several actions follow directly from the analysis. Audit the reward map annually and publish what it shows. Employees already know the answer. Redesign meetings so that contribution does not depend on fluency, and evaluate performance through artefacts wherever possible. Formalise and reward #boundary_spanning. Make it a role, put it in the objectives, and pay for it. Design arrival and departure with equal care. Treat handover as a deliverable. Distinguish clearly between what is fixed by immigration law and what is chosen by the firm, and remove discretionary inequities. Measure the health of the habitat, not only the satisfaction of the individual. Useful indicators include cross group collaboration rates, the distribution of speaking time, the number of active bridging relationships and the quality of handover documentation, alongside conventional engagement measures. 7.3 Implications for human resource policy Human resource functions in hub organisations should reconsider three defaults. The first is the assumption that lower turnover is always better. Where mobility is structural, the right target is orderly turnover with protected knowledge, not minimal turnover. The second is the assumption that language is a training issue. #language_policy is a distributional decision and should be made with the same seriousness as a compensation decision, drawing on evidence that language shapes careers and inclusion (Presbitero et al., 2023; Ciuk et al., 2023). The third is the assumption that talent identification is culturally neutral. Assessment centres, interviews and potential ratings all reward verbal performance in a dominant language, and their outputs should be examined for the fluency distortion described in Proposition 2. #cultural_intelligence and #cultural_agility frameworks are useful here, but they should be applied to the assessors as well as to the assessed (Richter et al., 2023; Caligiuri and Caprar, 2023). 7.4 Implications for host locations and policy makers Cities that market themselves as #talent_hubs, including major #gulf_labour_markets and comparable destinations elsewhere, have an interest in the internal social structure of the firms they attract, because segregation inside firms reproduces segregation outside them. Policy instruments that affect the cost of movement, including #visa_policy, residence rights and the portability of benefits, shape the strategies that mobile professionals adopt. Research on frictional mobility regimes and on the governance of remote and nomadic work shows that these instruments are actively contested and rapidly evolving (Mancinelli and Germann Molz, 2023; Holleran, 2022; Zhou, Buhalis, Fan, Ladkin and Lian, 2024). 7.5 Implications for students For students who intend to work in or study these environments, three habits of mind follow from the analysis. Read the environment before reading the people. Ask what is scarce, what is rewarded and what is punished, and much of the observed behaviour will become predictable. Notice who speaks and who does not, and treat that distribution as data. Build #career_capital deliberately and expect to move. This is not disloyalty. It is the rational strategy in a habitat structured for movement, and the organisations that understand this will treat it as ordinary. 8. Limitations and Future Research 8.1 Limitations The framework is conceptual and therefore unproven. Its propositions are plausible, derived from existing evidence, and falsifiable, which is the most that can be claimed at this stage. The ecological metaphor may mislead if taken literally. Human behaviour is shaped by institutions, law, history and moral commitment in ways that have no biological analogue, and any application of this lens must remain alert to that difference. The evidence base skews towards large multinationals in particular sectors, and towards research written in English. Small firms, public sector organisations and non English scholarship are under-represented. The framework treats the hub as a relatively bounded unit. In practice hubs overlap with virtual networks, contractor populations and platform labour, and the boundaries are less clean than the analysis implies. 8.2 A research agenda The following studies would test the framework directly and are within reach of doctoral and master's level researchers. Study one: the fluency distortion. Collect supervisor performance ratings and objective output measures for the same individuals in a multilingual unit, alongside independent assessments of second language fluency. Proposition 2 predicts that fluency will correlate more strongly with the ratings than with the output. Study two: the artefact intervention. Introduce artefact based evaluation in one set of teams and not in another, and examine whether the association between fluency and rated performance weakens. This tests Proposition 3. Study three: bubble formation. Map interaction networks in a diverse hub and model tie formation as a function of interaction cost, measured through shared language, shared function and physical proximity, while controlling for measured intercultural attitudes. This tests Proposition 4. Study four: legitimated exit. Compare units that operate alumni programmes and developmental exit interviews with units that do not, examining #turnover_intention among stayers and the completeness of knowledge handover. This tests Proposition 6. Study five: paradoxical leadership and diversity. Test whether the association between paradoxical leader behaviour and team effectiveness is moderated by linguistic and national diversity, extending the meta-analytic evidence to a specifically multilingual setting (Lee et al., 2023). This tests Proposition 7. Study six: virtual hubs. Examine whether the four domains operate differently when the hub is partly or wholly distributed, drawing on work concerning the shift to work from anywhere (Smite et al., 2023) and on the emerging literature concerning #digital_nomads (Kozak et al., 2024). Study seven: longitudinal habitat. Follow a single hub across several turnover cycles and examine whether environmental features, rather than individual leaders, predict performance continuity. This is the strongest available test of the framework's central claim. Methodologically, mixed designs are likely to be the most informative. Network analysis captures boundary structure. Experience sampling captures the daily cost of operating in a second language. Qualitative interviewing remains essential for #sensemaking, because the meaning that mobile professionals attach to their own movement cannot be inferred from behaviour alone. #feedback_loops between individual strategy and environmental structure are best examined over time, which makes longitudinal designs particularly valuable. Finally, #resilience deserves separate attention. Hubs are exposed to shocks, including regulatory change, geopolitical disruption and abrupt shifts in visa rules, and the capacity of a habitat to absorb such shocks without collapse is an open empirical question. #future_research on this point would be of immediate use to practitioners. 9. Conclusion The organisations that concentrate mobile, multilingual talent in a single location have created an environment that leadership theory did not design for. People arrive from everywhere, work in a language that most of them learned as adults, compete for a small number of visible opportunities, and leave in a few years by plan rather than by accident. This article has argued that the most useful description of the leader in such a setting is not commander, coach or visionary, but steward of a habitat. The four domains of habitat design, signal management, boundary management and flow management provide a vocabulary for that role, and eight propositions turn the vocabulary into a research programme. Three claims are worth restating. The environment does more work than the relationship. Where tenure is short, what the leader builds into the setting outlasts what the leader builds with any individual. Language is a leadership variable. A hub that reads fluency as competence is making its decisions on the wrong information, and no amount of goodwill will correct a signal problem that has not been recognised as one. Departure is not failure. In a habitat structured for movement, the organisation that plans for exit will keep more of what matters than the organisation that pretends exit will not happen. The framework is offered in the spirit of a hypothesis. It should be tested, refined and, where it is wrong, discarded. Its value lies not in its elegance but in whether it helps students, researchers and managers see the environments they inhabit a little more clearly, and act in them a little more wisely. Hashtags #Behavioral_Ecology_Of_Leadership #Executive_Adaptation_In_Global_Firms #Expatriate_Talent_Hubs #Multilingual_Management #Global_Mobility_Research #International_Human_Resource_Management #Cross_Cultural_Leadership #Talent_Ecosystems #Highly_Mobile_Workforce #Language_And_Power_At_Work #Adaptive_Leadership_Paradigms #Organizational_Behavior_Research #Management_Studies_For_Students #Expatriate_Leadership_Practice #STULIB_Research References Aljbour, A., French, E., and Ali, M. (2022). An evidence-based multilevel framework of talent management: A systematic review. International Journal of Productivity and Performance Management, 71(8), 3348-3376. https://doi.org/10.1108/IJPPM-02-2020-0065 Andresen, M., Brewster, C., and Suutari, V. (Eds.). (2021). Self-initiated expatriation in context: Recent research and future directions. Routledge. Bagga, S., Gera, S., and Haque, S. (2023). The mediating role of organizational culture: Transformational leadership and change management in virtual teams. Asia Pacific Management Review, 28(2), 120-131. https://doi.org/10.1016/j.apmrv.2022.07.003 Batirlik, S., Gencer, Y., and Akkucuk, U. (2022). Global virtual team leadership scale (GVTLS) development in multinational companies. Sustainability, 14(2). Boemelburg, R., Zimmermann, A., and Palmie, M. (2023). How paradoxical leaders guide their followers to embrace paradox: Cognitive and behavioral mechanisms of paradox mindset development. Long Range Planning, 56(2). Bonache, J., Brewster, C., and Froese, F. J. (Eds.). (2021). Global mobility and the management of expatriates. Cambridge University Press. https://doi.org/10.1017/9781108679220 Caligiuri, P., and Caprar, D. V. (2023). Becoming culturally agile: Effectively varying contextual responses through international experience and cross-cultural competencies. The International Journal of Human Resource Management, 34(12). Caligiuri, P., Collings, D. G., De Cieri, H., and Lazarova, M. B. (2024). Global talent management: A critical review and research agenda for the new organizational reality. Annual Review of Organizational Psychology and Organizational Behavior, 11, 393-421. https://doi.org/10.1146/annurev-orgpsych-111821-033121 Ciuk, S., Sliwa, M., and Harzing, A.-W. (2023). Implementing the equality, diversity, and inclusion agenda in multinational companies: A framework for the management of (linguistic) diversity. Human Resource Management Journal, 33(4), 868-888. https://doi.org/10.1111/1748-8583.12487 Devi, N. C. (2024). Paradoxical leadership and employee creativity: Knowledge sharing and hiding as mediators. Journal of Knowledge Management, 28(2), 312-340. Eisenberg, J., Glikson, E., and Lisak, A. (2021). Multicultural virtual team performance: The impact of media choice and language diversity. Small Group Research, 52(5), 507-534. https://doi.org/10.1177/1046496420985614 Garavan, T. N., Morley, M. J., Cross, C., Carbery, R., and Darcy, C. (2021). Tensions in talent: A micro practice perspective on the implementation of high potential talent development programs in multinational corporations. Human Resource Management, 60(2). Holleran, C. (2022). The nomad's dilemma: Mobility, precariousness, and the failure of the digital nomad outlook to address state-citizen relations. Mobilities, 17(6), 832-847. https://doi.org/10.1080/17450101.2022.2104085 Kaliannan, M., Darmalinggam, D., Dorasamy, M., and Abraham, M. (2023). Inclusive talent development as a key talent management approach: A systematic literature review. Human Resource Management Review, 33(1). Kozak, M., Cetin, G., and Alrawadieh, Z. (2024). Repositioning work and leisure: Digital nomads versus tourists. International Journal of Tourism Research, 26(4), e2732. https://doi.org/10.1002/jtr.2732 Lee, A., Lyubovnikova, J., Zheng, Y., and Li, Z. F. (2023). Paradoxical leadership: A meta-analytical review. Frontiers in Organizational Psychology, 1, 1229543. https://doi.org/10.3389/forgp.2023.1229543 Machado, C. F. (2022). Self-initiated expatriation: A new management challenge. An analysis based on bibliometrics. Administrative Sciences, 12(4), 136. Mancinelli, F., and Germann Molz, J. (2023). Moving with and against the state: Digital nomads and frictional mobility regimes. Mobilities, 19(2), 189-207. Peltokorpi, V., and Xie, J. (2023). When little things make a big difference: A Bourdieusian perspective on skilled migrants' linguistic, social, and economic capital in multinational corporations. Journal of International Business Studies. Piekkari, R., Gaibrois, C., and Johansson, M. (2022). A review of language-sensitive research in international business: A multi-paradigmatic reading. Journal of Comparative International Management, 25(1), 144-174. https://doi.org/10.55482/jcim.2022.32906 Presbitero, A., Froese, F. J., Peltokorpi, V., Pudelko, M., and Tenzer, H. (2023). Language in international human resource management: Current state of research and future research directions. The International Journal of Human Resource Management, 34(21), 4027-4045. https://doi.org/10.1080/09585192.2023.2274659 Primecz, H. (2023). Expatriate bubbles in Dubai: Expatriates from West and East cohabitating with locals. Journal of Global Mobility, 11(1), 6-20. https://doi.org/10.1108/JGM-06-2022-0032 Richter, N. F., Schlaegel, C., Taras, V., Alon, I., and Bird, A. (2023). Reviewing half a century of measuring cross-cultural competence: Aligning theoretical constructs and empirical measures. International Business Review, 32(4). Semenov, A. V., and Randrianasolo, A. (2024). An extended conceptualization of cultural intelligence. International Journal of Cross Cultural Management. https://doi.org/10.1177/14705958241268289 Smite, D., Moe, N. B., Klotins, E., and Gonzalez-Huerta, J. (2023). From forced working-from-home to voluntary working-from-anywhere: Two revolutions in telework. Journal of Systems and Software, 195, 111509. van Bakel, M., and Vance, C. M. (2023). Breaking out of the expatriate bubble in Denmark: Insights from the challenge of making connections with local Danes. Journal of Global Mobility, 11(1). Vuchkovski, D., Zalaznik, M., Mitrega, M., and Pfajfar, G. (2023). A look at the future of work: The digital transformation of teams from conventional to virtual. Journal of Business Research, 163, 113912. Weinzierl, M.-L. (2025). Bridging the boundaries of corporate language competence in multinational teams. International Journal of Business Communication. https://doi.org/10.1177/23294906231221135 Zhou, L., Buhalis, D., Fan, D. X. F., Ladkin, A., and Lian, X. (2024). Attracting digital nomads: Smart destination strategies, innovation and competitiveness. Journal of Destination Marketing and Management, 31, 100850.

Artificial intelligence has become a structural condition of academic work rather than a specialist tool used by a small number of computer science departments. Yet the ability of a university to train, adapt, audit or even reliably access advanced AI systems depends on physical and financial infrastructures that are distributed with extreme unevenness across the world. This article applies #world_systems analysis, in the tradition established by Immanuel Wallerstein and extended by later scholars of dependency and coloniality, to the global landscape of #AI_infrastructure in higher education. It argues that the worldwide university system is being restructured into a recognisable #core_periphery hierarchy defined not by curriculum or ranking position alone but by control over eight infrastructural layers: semiconductors, computing capacity, energy and water, data, human capital, knowledge circulation, procurement and finance, and standards and governance. Drawing on recent empirical and critical scholarship, including work mapping the global distribution of cloud computing capacity into what has been called the Compute North, the Compute South and the compute desert, the article develops a layered model of #unequal_exchange in academic AI. Five mechanisms are identified: infrastructural rent, data transfer without value return, credential and talent drainage, epistemic standard setting, and dependency lock-in. The article then examines the role of the #semi_periphery, where states such as China, India, Brazil, the Gulf countries and parts of Southeast Asia are attempting selective upgrading through sovereign compute programmes. Finally it evaluates counter-strategies available to universities in peripheral and semi-peripheral positions, including regional compute consortia, open weight model adoption, small and efficient model research, data governance regimes, and curriculum reform. The article concludes that debates about equity in AI and education cannot be resolved at the level of classroom policy or academic integrity rules, because the decisive inequalities are infrastructural, capital intensive, and geopolitical in nature. Keywords: artificial intelligence infrastructure; world-systems analysis; higher education; compute divide; digital colonialism; global south; technological sovereignty; resource allocation 1. Introduction Universities have always been embedded in an international division of intellectual labour. Some institutions produce theories, journals, standards and graduates that are exported globally. Others import those products, adapt them locally, and send their most talented students and researchers to the exporters. This pattern predates the computer, and it has been described in various vocabularies: dependency, academic imperialism, centre and margin, the geopolitics of knowledge. What is new in the current period is that this hierarchy is being hardened into physical infrastructure that is expensive, geographically concentrated, and controlled by a very small number of firms and states. The rise of large-scale machine learning has made computing capacity a condition of participation in entire fields of research. A doctoral student who wishes to study protein folding, climate downscaling, medical imaging, natural language processing, or econometrics with modern methods now requires access to accelerated computing hardware. That hardware is manufactured through a supply chain with only a handful of viable nodes, is sold at prices that exceed the annual research budgets of many universities, is hosted in data centres that are concentrated in a limited number of countries, and is increasingly rented rather than owned. The result is that the practical capacity to do frontier research is being redistributed away from the university sector as a whole and toward corporate laboratories, and within the university sector it is being redistributed toward a small group of wealthy institutions in wealthy states. Empirical work supports this reading. Besiroglu and colleagues document a pronounced and growing gap between the computing resources available to industrial laboratories and those available to academic ones, and they show that academic-only research teams have become less well represented in compute intensive areas such as foundation models, with academia shifting toward using pre-trained models produced by industry rather than producing such models itself (Besiroglu et al., 2024). Lehdonvirta, Wu and Hawkins map the geography of public cloud computing capacity and find a world divided into countries that host compute suitable for model training, countries that host only compute suitable for running existing models, and countries that host no relevant public cloud AI capacity at all, a category they describe as a #compute_desert (Lehdonvirta, Wu and Hawkins, 2024). The vocabulary of #Compute_North and #Compute_South emerges directly from this mapping, and it maps closely onto older economic geographies of industrial concentration. For #higher_education this is not a peripheral technical matter. It determines which universities can train models rather than merely consume them, which can audit systems that are deployed on their own populations, which can build tools in their own languages, which can negotiate with vendors from a position of strength, and which must accept whatever terms are offered. It shapes hiring, doctoral training, publication, ranking, and ultimately the intellectual sovereignty of national academic systems. This article makes three contributions. First, it adapts world-systems analysis, which was developed to explain the structure of the capitalist world economy, into a framework for analysing #AI_infrastructure in academic contexts, and specifies what the categories of core, semi-periphery and periphery mean when applied to computing rather than to cotton, sugar or steel. Second, it disaggregates AI infrastructure into eight layers and traces how each layer generates a distinct form of dependency for universities outside the core. Third, it evaluates the realistic strategies available to institutions and states that occupy subordinate positions, distinguishing between strategies that produce genuine upgrading and strategies that merely reproduce dependency in a more comfortable form. The argument is not that AI is uniquely oppressive, nor that universities in the #global_south are passive victims. The argument is structural. Positions in a system generate pressures and opportunities, and actors respond strategically within them. Some peripheral universities are pursuing sophisticated strategies of selective engagement, and some semi-peripheral states are executing deliberate upgrading programmes with substantial success. The purpose of a structural analysis is to make the terrain visible so that such strategies can be designed with clear eyes rather than with borrowed enthusiasm. The article proceeds as follows. Section 2 sets out the conceptual framework. Section 3 reviews the relevant literature across critical AI studies, the political economy of higher education, and infrastructure studies. Section 4 explains the method and its limits. Section 5 presents the layered analysis. Section 6 synthesises the mechanisms of #unequal_exchange. Section 7 examines counter-strategies. Section 8 offers implications for policy. Section 9 states limitations. Section 10 concludes. 2. Conceptual Framework: World-Systems Analysis Applied to Computing 2.1 The classical framework World-systems analysis, most closely associated with Immanuel Wallerstein, treats the modern world as a single integrated economic system rather than as a collection of separate national societies that happen to trade with one another (Wallerstein, 2004). The system contains a single division of labour and multiple political units. Within that division of labour, production processes can be arranged along a spectrum from core-like to peripheral. Core-like processes are those that are relatively monopolised, capital intensive, high skill, and capable of capturing a large share of the total surplus generated in a commodity chain. Peripheral processes are those that are competitive, low margin, and capable of capturing only a small share of that surplus. Three features of the framework matter for the present argument. The first is that core and periphery are properties of processes, not of countries. A country is core-like when core-like processes are concentrated within it. This is important because it allows the analysis to describe a single country that contains both a world-leading research university with substantial computing capacity and a set of teaching institutions with almost none. Structural position is not uniform inside national borders. The second is that the #semi_periphery is a real and stable category, not simply a stage on the way to the core. Semi-peripheral zones combine core-like and peripheral processes. They are exploited by the core and simultaneously exploit the periphery. They tend to be the sites of the most aggressive state-led development strategies, because their position is unstable and their elites are acutely aware of the risk of downgrading. The third is the concept of #unequal_exchange. Trade between core-like and peripheral processes systematically transfers surplus toward the core, not because of fraud or coercion in each individual transaction, but because of the structural difference in monopoly power between the two sides. Each transaction may be voluntary and legally sound, and the aggregate effect may still be a persistent transfer of value. 2.2 Translating the framework to AI infrastructure Applying this framework to AI requires specifying what counts as a core-like process in the production of intelligent systems. The relevant #commodity_chain runs approximately as follows: raw materials and minerals, semiconductor design, semiconductor fabrication, accelerator manufacture and assembly, data centre construction and energy provision, cloud service operation, model pre-training, model fine-tuning and adaptation, application development, deployment and user interface, and finally data generated by users which flows back into the chain. The distribution of monopoly power across this chain is severe. Design of advanced accelerators is dominated by a small number of firms. Fabrication at leading process nodes is concentrated in an extremely small number of facilities. Lithography equipment has effectively one dominant supplier. Cloud hosting at hyperscale is dominated by a handful of providers. Pre-training of the largest models is performed by a small number of laboratories, most of them corporate. By contrast, the peripheral segments of this chain are highly competitive and low margin: mineral extraction, hardware assembly, content moderation and data annotation, and application layer development. Universities are located at multiple points in this chain, and their position determines their fate. A university that can pre-train a large model occupies a core-like position. A university that can only fine-tune models produced elsewhere occupies an intermediate position. A university that can only consume models through a commercial interface, paying per query, occupies a peripheral position. A university whose main contribution is to supply data, students and labour to institutions elsewhere occupies a deeply peripheral position, regardless of the wealth of the country in which it happens to sit. 2.3 Four propositions From this translation, four propositions follow, and they organise the empirical sections that follow. Proposition one: infrastructural monopoly generates rent. Because the upstream segments of the AI chain are monopolised, the owners of those segments can charge prices well above the cost of production. Universities that must rent access pay this rent continuously, and it is extracted from budgets that would otherwise fund salaries, libraries, scholarships and locally defined research. This is a transfer from academic budgets in the periphery to corporate balance sheets in the core. Proposition two: value flows toward the layer with the highest capital intensity. In the AI chain, the highest #capital_intensity is found in fabrication and in large scale pre-training. These are also the layers with the strongest #scale_economies. Consequently the surplus generated by the entire chain, including the surplus generated by peripheral universities using AI to produce useful local research, tends to accumulate at those layers. Proposition three: dependency reproduces itself through skills and standards. Once an institution builds its teaching, its research pipelines and its administrative systems around a particular vendor's tools, the cost of switching rises. Graduates are trained on those tools. Papers are benchmarked against those models. Grant applications assume their availability. This is #dependency_theory in a contemporary form: the very act of catching up in a dependent manner deepens the dependence. Proposition four: the semi-periphery pursues sovereignty projects. States in intermediate positions respond by attempting to build national capacity in one or more layers, usually compute and models, sometimes chips. These #technological_sovereignty projects are expensive, politically driven, and only partially successful, but they alter the structure of the system and cannot be dismissed as symbolic. 2.4 Relationship to decolonial and dependency scholarship World-systems analysis is not the only available critical lens, and it is strengthened by dialogue with decolonial approaches. Muldoon and Wu argue that the global power imbalances in AI production are continuous with historical colonialism, and they describe a colonial #supply_chain of AI running from mineral extraction through data labour to knowledge production (Muldoon and Wu, 2023). Tacheva and Ramasubramanian analyse the generative AI order as an interlocking system of extraction, automation, surveillance and containment that operates transnationally (Tacheva and Ramasubramanian, 2023). Ricaurte insists that ethical frameworks designed in the core cannot simply be exported to the majority world, because they abstract away from the material violence of the systems they claim to regulate (Ricaurte, 2022). The value of combining these traditions is that world-systems analysis contributes a precise account of how surplus moves through a chain, while decolonial scholarship contributes a precise account of whose knowledge counts and whose bodies bear the cost. Both are necessary. A purely economic account cannot explain why certain languages are absent from training data. A purely cultural account cannot explain why a university cannot buy an accelerator. 3. Literature Review 3.1 From digital divide to compute divide The concept of the #digital_divide, developed in the 1990s, initially described unequal access to computers and connectivity, and was later refined into a multi-level construct covering access, skills, usage and outcomes. This literature was largely adequate for a period in which the marginal cost of participation in computing was falling. Personal computers became cheaper. Connectivity spread. Open source software reduced licensing barriers. Cloud services allowed small actors to rent capacity that they could not afford to own. The trajectory appeared to be one of gradual convergence, punctuated by persistent but narrowing gaps. Large-scale machine learning reversed that trajectory in one crucial respect. The capital requirement for participation at the frontier began to rise sharply rather than fall. Training runs that were once feasible in a well-funded university laboratory moved beyond the reach of any university. The resulting #compute_divide has been documented most systematically by Besiroglu and colleagues, who show a decline in the representation of academic-only teams in the most compute intensive areas of machine learning, and a corresponding shift in academic work toward the adaptation of industrially produced pre-trained models (Besiroglu et al., 2024). This is a qualitative change in the structure of the field, not merely a quantitative gap. It converts a large part of the academic community from producers into users. The geographic dimension of this divide is addressed by Lehdonvirta, Wu and Hawkins, whose census of hyperscale cloud regions demonstrates that the capacity relevant to model development is concentrated in a small number of jurisdictions, while a large number of countries host no relevant public cloud AI capacity whatsoever (Lehdonvirta, Wu and Hawkins, 2024). Their tripartite division into Compute North, Compute South and compute desert provides an empirical foundation that maps with striking directness onto the core, semi-periphery and periphery of world-systems theory. 3.2 Political economy of AI and the concentration of power A parallel literature addresses the concentration of power in AI production. Verdegem analyses AI as a general purpose technology whose development is shaped by the commodification of data, the concentration of talent, and the accumulation of computing capacity within a small number of firms, producing an enclosure dynamic that forecloses public alternatives, and proposes the commons as a counter-model (Verdegem, 2024). Whittaker, writing from experience inside both industry and academia, describes the capture of AI research by industry interests and warns that the field's research agenda, its methods and even its critical vocabulary are shaped by the entities that fund and host it (Whittaker, 2021). Ahmed, Wahed and Thompson document the growing influence of industry across AI research inputs and outputs, including compute, data, talent and publications (Ahmed, Wahed and Thompson, 2023). The literature on openness is directly relevant to universities, because open models are frequently proposed as the escape route from dependency. Widder, Whittaker and West argue that systems described as open in AI are typically open in limited respects only, and that openness in weights does not by itself redistribute the power that flows from control over compute, data and distribution (Widder, Whittaker and West, 2024). This is an important corrective to a comfortable assumption in higher education policy, namely that the availability of downloadable model weights solves the infrastructural problem. It does not, although it substantially changes its shape. 3.3 Environmental and material foundations AI infrastructure is physical. Crawford's account traces the material substrate of machine learning through minerals, energy, labour and land, and insists that the apparent immateriality of the cloud is an illusion sustained by distance (Crawford, 2021). Quantitative work has followed. De Vries analyses the trajectory of AI-related electricity consumption and argues that the expansion of AI servers places significant new demands on data centre electricity use (de Vries, 2023). Li and colleagues document the water footprint of AI models, drawing attention to freshwater consumption for server cooling and for electricity generation, a dimension that had previously remained largely outside public discussion (Li, Yang, Islam and Ren, 2025). For the present argument, the significance of this literature is distributional. The #energy_footprint and #water_footprint of AI are borne in particular places, and those places are frequently not the places where the resulting models are designed, monetised, or primarily used. This is a classical pattern of ecologically unequal exchange, and it now runs directly through the infrastructure on which universities depend. 3.4 AI and higher education The higher education literature on AI has expanded rapidly, but much of it remains focused on pedagogy, assessment and academic integrity. Bearman, Ryan and Ajjawi conducted a critical review of references to artificial intelligence in leading higher education journals and found that definitions were scarce and inconsistent, and that two dominant discourses recur: a discourse of imperative change, in which AI is presented as an inevitable transformation to which institutions must respond, and a discourse of altering authority, in which authority is redistributed among teachers, students, machines and corporations (Bearman, Ryan and Ajjawi, 2023). Their finding that AI is rarely treated as a research object in its own right is telling, and it helps explain why infrastructural questions have received so little attention in the field. Williamson argues that AI in education must be understood through its social, political and economic embedding, including the role of venture capital, the strategies of the #edtech_industry, and the policy enthusiasm that legitimises deployment (Williamson, 2024). Selwyn cautions against the assumption that AI will straightforwardly improve education, and draws attention to the distributional and labour implications of automation in teaching (Selwyn, 2022). Bozkurt and colleagues, writing collectively across many countries, capture the ambivalence of the global educational community toward generative AI and stress the risk that the technology will widen existing gaps rather than close them (Bozkurt et al., 2023). Work grounded in the global south is particularly important here. Eke, Wakunuma and Akintoye assemble analyses of responsible AI in African contexts that foreground infrastructural constraints, governance capacity and the danger of importing frameworks designed elsewhere (Eke, Wakunuma and Akintoye, 2023). Png analyses the structural difficulties faced by global south stakeholders in AI governance forums, including limited representation, resource constraints, and agenda setting by better resourced actors (Png, 2022). 3.5 Language, data and benchmarks A final strand concerns the representation of the world in the data on which models are trained. Joshi and colleagues show a severe skew in the availability of language resources, with a small number of languages dominating the data and tooling that make natural language processing possible, and a very large number of languages effectively excluded (Joshi et al., 2020). Blasi, Anastasopoulos and Neubig quantify systematic inequalities in the performance of language technologies across the world's languages, demonstrating that these gaps are not random but structured (Blasi, Anastasopoulos and Neubig, 2022). Bender and colleagues warn that scale in language models brings not only performance but also the encoding of dominant perspectives and the marginalisation of others, alongside significant environmental and financial costs (Bender, Gebru, McMillan-Major and Shmitchell, 2021). The higher education implication is severe. A university teaching in a #low_resource_language faces not only a compute constraint but a data constraint, and the two compound. It cannot easily build strong models in its own language because the data does not exist in the required form, and it cannot easily create that data because doing so requires funding, tooling and compute that it does not control. 3.6 The gap this article addresses Three literatures thus exist in parallel: a technical and economic literature on the compute divide, a critical and decolonial literature on AI power, and an educational literature on AI in universities. They rarely meet. The educational literature mostly discusses classrooms; the compute literature mostly discusses laboratories and firms; the decolonial literature mostly discusses labour and epistemology. What is missing is an integrated structural account of how the global university system as a whole is being stratified by AI infrastructure, and of what this implies for #resource_allocation, research agendas and institutional strategy. That is the gap this article addresses. 4. Method 4.1 Approach This is a conceptual and structural analysis supported by a synthesis of secondary evidence. It does not present new primary data. Its purpose is theory building and integration, and its claims should be assessed by the standards appropriate to that purpose: internal coherence, consistency with available evidence, and usefulness in generating researchable propositions. The method proceeds in four steps. First, a framework is specified by translating the categories of world-systems analysis into terms appropriate to computing infrastructure, as set out in Section 2. This translation is disciplined by the requirement that each category must be operationalisable, at least in principle, by observable indicators. Second, the AI infrastructure relevant to universities is disaggregated into eight layers. The layers were derived by tracing the production chain of a deployed AI system from raw material to end use, and by asking at each point which actors control the layer and what a university must do to obtain access to it. Third, for each layer, recent scholarly literature is reviewed to establish the degree of concentration, the mechanism by which access is allocated, and the consequences for universities in different structural positions. Priority was given to peer-reviewed work published within roughly the last five years, supplemented where necessary by well-established earlier work whose findings remain uncontested. Fourth, the layer analyses are synthesised into a model of #unequal_exchange specifying five mechanisms, and the model is used to evaluate counter-strategies. 4.2 Operationalising structural position For empirical work that may follow, structural position in the academic AI system can be approximated by a composite of indicators. A university may be treated as core-like to the extent that it satisfies conditions such as the following: it owns or has guaranteed institutional access to a substantial cluster of AI accelerators rather than renting capacity per hour; it hosts researchers who publish work involving the pre-training or substantial architectural modification of large models; it participates in the governance bodies and standards processes that define acceptable practice in the field; it produces graduates who are recruited into the laboratories that build frontier systems; it holds intellectual property or open assets that other institutions depend upon; and it negotiates with vendors on terms it substantially influences. A university may be treated as peripheral to the extent that it satisfies the inverse conditions: it accesses AI only through commercial interfaces on standard terms; its researchers apply externally produced models without modifying them; it exports its strongest doctoral candidates and receives no compensating inflow; it contributes data or annotation labour to systems it does not control; and its curriculum is defined by the tools that vendors make available to it. The semi-periphery holds a mixed profile: sovereign or national compute programmes exist, substantial fine-tuning and adaptation work is undertaken, models in national languages are produced, and yet the underlying hardware, the leading architectures, and the standards remain externally determined. 4.3 Limitations of method Three limits should be stated at the outset. The framework is deliberately structural and therefore under-describes agency; individual institutions and individual scholars regularly outperform their structural position. The evidence base is uneven, since infrastructure data is often commercially confidential and national reporting is inconsistent. And the empirical situation changes quickly, which means that specific configurations described here may be superseded even while the structural logic persists. Section 9 returns to these points. 5. The Eight Layers of AI Infrastructure in Higher Education 5.1 Layer one: semiconductors and the physical substrate The foundation of the entire system is the advanced #semiconductors industry. Leading edge fabrication is concentrated in a very small number of facilities, the equipment required to operate those facilities has effectively one dominant supplier, and the design of the accelerators used for machine learning is dominated by a small number of firms whose software ecosystems raise further barriers to substitution. Universities interact with this layer in three ways. They purchase hardware, they teach the engineering that underlies it, and they are affected by the policy instruments that govern its trade. In recent years, #export_controls have made the last of these decisive. When states restrict the sale of advanced accelerators to particular jurisdictions, the effect is not confined to military or commercial actors. Universities in the restricted jurisdiction find their research capacity constrained, their collaborations complicated, and their graduate students exposed to visa and compliance frictions. A university may thus find its research agenda shaped by trade policy decisions made in a country in which it has no representation and no vote. Crawford's insistence on the materiality of AI is relevant here in a second sense (Crawford, 2021). Mineral extraction for electronics occurs disproportionately in countries whose universities have almost no access to the resulting technology. The chain begins in the periphery and terminates in the core, which is precisely the pattern that world-systems analysis was developed to describe. A university located near a mining region may study the environmental consequences of extraction while being unable to afford a single accelerator built from the extracted materials. For most institutions, the practical conclusion is that this layer is inaccessible in the short term. Attempts to enter it require sustained state investment over decades, and only a small number of semi-peripheral states are plausibly positioned to try. 5.2 Layer two: computing capacity and the shape of the compute divide The second layer is the one that has attracted the most attention, and rightly so, because it is the layer at which the divide becomes operational for researchers. The evidence is clear. Industrial laboratories command computing resources that academic laboratories do not, and the gap has widened over the period in which large models became central to the field (Besiroglu et al., 2024). Meanwhile the geographic distribution of the cloud capacity through which most institutions would obtain access is itself highly concentrated, producing a division between countries hosting training-capable capacity, countries hosting only inference-capable capacity, and countries hosting none (Lehdonvirta, Wu and Hawkins, 2024). Three consequences follow for universities. The first is a shift in the type of research that is feasible. When #GPU_access is scarce, researchers rationally select problems that fit the available budget. Over time this produces a systematic bias in the questions asked. Work that requires large-scale experimentation migrates to institutions that can afford it, and work that requires only modest computation remains elsewhere. The scientific record then reflects the distribution of hardware as much as the distribution of curiosity. Klinger, Mateos-Garcia and Stathoulopoulos raise a related concern about the narrowing of research directions as the field concentrates around compute intensive paradigms (Klinger, Mateos-Garcia and Stathoulopoulos, 2024). The second is a shift in the type of expertise that is cultivated. Where compute is abundant, doctoral students learn to design and train systems. Where compute is scarce, they learn to use systems designed and trained by others. Both skills are valuable, but only one of them places its holder in a position to shape the field. Over a generation, this asymmetry becomes an entrenched difference in national research capability. The third is the emergence of #cloud_dependency as the default mode of access. Because owning hardware requires capital, power, cooling, staff and continuous replacement, most institutions rent. Renting is often rational at the margin and dangerous in aggregate, because it converts a capital asset into an operating expense that must be renewed indefinitely, and it places the terms of access in the hands of the provider. 5.3 Layer three: energy, water and environmental burden Computing capacity is meaningless without power. Data centres require large, stable, and increasingly enormous electricity supplies, and their expansion places measurable pressure on national grids (de Vries, 2023). They also require water, both directly for cooling and indirectly through electricity generation, a footprint that has been quantified only recently and that varies considerably by location and by season (Li et al., 2025). This layer produces a specific injustice that is often invisible in university sustainability reporting. A university may procure AI services from a provider whose data centre is located in a water-stressed region, and may count the resulting emissions as belonging to someone else. The infrastructure is elsewhere, therefore the burden appears to be elsewhere. In reality the burden has been displaced, not eliminated. #environmental_justice analysis requires that the burden be traced. There is also a strategic dimension. Countries with abundant and cheap energy have a comparative advantage in hosting AI infrastructure, and several are pursuing this advantage deliberately. This creates an unusual possibility: a state that is peripheral in semiconductor design may become significant in compute hosting because of its energy endowment. Whether this constitutes genuine upgrading or merely a new form of resource-based dependency is an open empirical question, and it depends on whether hosting is accompanied by domestic capability building in models, data and skills, or whether it remains an enclave that exports processed compute to foreign clients. For universities, the practical implication is that energy policy is now research policy. An institution that cannot secure reliable power cannot operate a cluster, regardless of whether it can afford one. In many peripheral contexts, the binding constraint is not the price of hardware but the reliability of the grid. 5.4 Layer four: data, language and the representation of the world Models are shaped by their training data, and the world is very unevenly represented in that data. Joshi and colleagues document a severe skew in the availability of linguistic resources, with a small group of languages richly supported and the majority effectively unsupported (Joshi et al., 2020). Blasi and colleagues demonstrate that this translates into systematically unequal performance of language technologies across the world's languages (Blasi et al., 2022). The result is that a student writing in a widely supported language receives a capable assistant, while a student writing in a #low_resource_language receives a degraded one, or none at all. This is #epistemic_inequality in its most concrete form. It affects teaching, since assessment tools and tutoring systems perform unevenly. It affects research, since text mining, translation and information retrieval are less effective. It affects administration, since student support systems fail for some populations. And it affects the historical record, since digitised archives in dominant languages are absorbed into models while others are not. The layer also generates a distinctive form of extraction. Universities produce enormous quantities of high quality text, code, imagery and structured data. Much of this is published openly, in the name of #open_science, and is absorbed into commercial training corpora without compensation, attribution at the individual level, or any mechanism by which the producing institution obtains a return. The knowledge flows out; the model comes back as a product to be rented. This is a textbook instance of #unequal_exchange, and it is one of the few points in the chain where peripheral universities hold an asset with real value. Whether they can convert that asset into leverage is a central question of #data_sovereignty. Behind the data sits labour. Muldoon and Wu foreground the global regimes of data labour and knowledge extraction that make AI systems possible while remaining invisible in the discourse of technological progress (Muldoon and Wu, 2023). #data_labour and #ghost_work are performed disproportionately in the periphery and semi-periphery, frequently by graduates and students of the very institutions that cannot afford the systems their labour helps to create. 5.5 Layer five: human capital and the direction of talent flows The fifth layer is people. Academic AI depends on a small number of highly trained researchers whose skills are in extreme demand. The consequences are twofold. Within national systems, universities lose staff to industry. Whittaker describes the capture of the field that results when the same small set of firms fund the research, employ the researchers, host the infrastructure and define the problems (Whittaker, 2021). #academic_capture is not primarily a matter of individual corruption; it is a structural consequence of an asymmetry in resources so large that the university cannot compete for its own graduates. Between national systems, the flow is directional. Talented students move from peripheral to core institutions for doctoral training, and a substantial share remains in the core. This #brain_drain is not new, but AI intensifies it, because the compute constraint gives ambitious researchers a concrete and non-ideological reason to leave: the work they wish to do cannot be done at home. A peripheral university may therefore fund a student's undergraduate education, discover their talent, and then effectively transfer that talent to a wealthier system at no cost to the recipient. In world-systems terms this is a transfer of embodied surplus, and it is one of the largest and least measured flows in the entire system. There is a partial counter-flow. Diaspora networks, remote collaboration, returning scholars, and internationally distributed research teams do return value. Whether the counter-flow approaches the outflow is doubtful, and it should be studied rather than assumed. 5.6 Layer six: knowledge circulation, publication and benchmarks The sixth layer concerns how knowledge about AI is produced, validated and circulated. Several mechanisms interact. Conference geography matters. The dominant venues of machine learning research are concentrated in a small number of countries, and physical attendance requires visas, funding and time. Scholars from countries facing visa barriers experience a structural disadvantage in networking, collaboration and visibility that is invisible in the published record but decisive in careers. #conference_access is a boundary of the field. Reviewing and agenda setting matter. #peer_review committees are drawn disproportionately from institutions with abundant compute, and the standards they apply reflect what is achievable with such resources. A paper reporting a careful study on modest hardware may be judged against work that used vastly greater resources, and be found wanting on grounds that have nothing to do with the quality of the reasoning. Benchmarks matter. The evaluation suites that define progress in the field encode particular tasks, particular languages and particular assumptions about what capability means. #benchmark_bias means that a model optimised for the priorities of the core will look strong on the metrics of the core, while a model optimised for local priorities will look weak, even if it is more useful in its context. Because benchmarks determine publication, funding and prestige, this is a powerful mechanism of #standards_setting. Publication economics matter. The transition to open access has shifted costs from readers to authors in many venues, and article processing charges represent a substantial barrier for authors in institutions without dedicated funds. The intention was to widen access to reading; one consequence has been to narrow access to publishing. #publication_geography is thus shaped by fee structures as well as by research capacity. Taken together, these mechanisms mean that the periphery can produce excellent work and still fail to accumulate the symbolic capital that would allow it to influence the direction of the field. This is why #knowledge_production must be treated as an infrastructural layer in its own right. 5.7 Layer seven: procurement, licensing and institutional finance The seventh layer is the one that university administrators actually experience, and it deserves closer attention than it usually receives in critical scholarship. Universities acquire AI capability through #procurement. They sign agreements with cloud providers, licence software, purchase seats in productivity suites that now embed generative assistants, and pay for API access measured in tokens. Each of these decisions appears technical and reversible. In aggregate they are neither. Three dynamics deserve emphasis. The first is #vendor_lock_in. Once teaching, assessment, research pipelines, identity management and administrative workflows are built around a particular provider's ecosystem, migration becomes expensive and risky. The provider's pricing power increases accordingly. Introductory discounts and free educational tiers, which are frequently offered, function as an investment in future lock-in rather than as philanthropy. This is not an accusation of bad faith; it is a description of a rational commercial strategy that universities should anticipate. The second is the conversion of capital budgets into recurrent obligations. Owning a cluster produces a depreciating asset that the institution controls. Renting capacity produces a permanent liability that the institution does not control. Public universities operating under multi-year budget frameworks are structurally vulnerable to price increases in the latter arrangement, and they have limited ability to negotiate because the market is concentrated. The third is the asymmetry of information. Universities negotiating with hyperscale providers typically face counterparties with vastly greater legal, commercial and technical capacity. Consortium purchasing, which is common in library acquisitions, remains underdeveloped in AI procurement, and this is one of the clearest and most immediately actionable weaknesses in current institutional practice. #licensing_costs negotiated collectively behave very differently from licensing costs negotiated individually. Williamson's analysis of the political economy of educational technology is directly relevant, since it shows how investor expectations and market structures shape what is offered to educational institutions and on what terms (Williamson, 2024). The products universities buy are not neutral tools; they are designed to produce returns for their owners, and their design reflects that purpose. 5.8 Layer eight: standards, governance and the making of rules The final layer is regulatory and normative. Rules about safety, evaluation, data protection, model release and permissible use are being written now, and they will shape what universities everywhere may do. Png demonstrates that global south stakeholders face structural constraints in AI governance processes, including resource limits, representational deficits and agendas set elsewhere (Png, 2022). The consequence is that regulatory regimes designed in the core are exported to the periphery through trade agreements, procurement conditions, publication norms and the compliance requirements of the platforms themselves. A university in a peripheral country may be bound by a data protection standard it did not write, evaluated against a safety benchmark it did not design, and required to use a model licence whose terms it cannot alter. There is a second-order effect on curriculum and research ethics. If the definition of responsible AI is settled in the core, then peripheral institutions teach that definition, examine on it, and produce graduates who reproduce it. Ricaurte's argument that ethical frameworks must be reconstructed from the position of the majority world, rather than translated into it, applies directly to the university curriculum (Ricaurte, 2022). #policy_governance is therefore not downstream of the technical layers; it is one of the mechanisms by which the technical hierarchy is stabilised. 6. Synthesis: Five Mechanisms of Unequal Exchange in Academic AI The layer analysis can now be synthesised. Five mechanisms transfer value from peripheral and semi-peripheral universities toward the core of the AI system. They are analytically distinct but empirically entangled. Mechanism one: infrastructural rent. Because the upstream layers are monopolised, access must be rented, and the rent is paid continuously. Every token processed, every hour of accelerator time, every seat licence is a small transfer. Individually these payments are modest; in aggregate, across thousands of institutions and millions of users, they constitute a substantial and permanent flow from public education budgets to a small number of firms. Crucially, the rent is paid regardless of whether the institution's own research succeeds. It is a charge on participation, not on outcome. This is the purest expression of #infrastructure_power. Mechanism two: data transfer without value return. Universities generate data of exceptional quality: peer-reviewed text, curated datasets, code, lecture material, structured records. Much of this flows into training corpora. The models built from it are then sold back to the institutions that produced the inputs. There is no accounting mechanism that returns value to the producing institution, and in most cases no mechanism by which it could even be measured. #data_extraction here does not require deception; it requires only the absence of an institutional mechanism for capturing the value of an asset that the institution has been trained to give away for free. Mechanism three: talent drainage. Peripheral systems bear the cost of primary, secondary and undergraduate education, and core systems capture the return. The compute divide converts this from a general tendency into a specific and rational choice for individual researchers. The transfer is large, embodied, and largely unmeasured. Mechanism four: epistemic standard setting. Benchmarks, review norms, venue prestige, ethical frameworks and safety standards are defined in the core and adopted globally. Compliance is a condition of participation. The effect is that peripheral institutions expend resources meeting standards that were designed around the capabilities and priorities of core institutions, and they are systematically disadvantaged in doing so. This mechanism converts material advantage into symbolic authority, and symbolic authority back into material advantage, since funding follows prestige. Mechanism five: dependency lock-in. Each of the four preceding mechanisms raises the cost of exit. A university that has trained its staff on a vendor's tools, built its systems on that vendor's cloud, published work benchmarked against that vendor's models, and shaped its curriculum around that vendor's assumptions cannot easily leave. Lock-in is the mechanism that converts a set of individually reasonable decisions into a structural condition. Table 1 summarises the structural positions that result. Attribute Core position Semi-peripheral position Peripheral position Compute access Owned clusters, guaranteed capacity National or consortium compute, partial Rented per use, or none Model relationship Pre-trains and designs Fine-tunes and adapts Consumes through interface Data relationship Aggregates data from elsewhere Builds national datasets Supplies data and annotation labour Talent flow Net importer Mixed, active retention policy Net exporter Standards role Sets benchmarks and norms Contests some norms, adopts most Adopts norms Procurement power Negotiates terms Negotiates collectively, partially Accepts standard terms Principal risk Regulatory constraint Downgrading, stranded investment Permanent dependence The table is a heuristic, not a classification scheme. Individual institutions frequently occupy different positions on different rows, and one of the most useful diagnostic exercises an institution can perform is to locate itself honestly on each row separately. 7. Responses from the Semi-Periphery and the Periphery Structures generate responses. This section evaluates the main strategies currently visible, assessing each by a single criterion: does it change the institution's structural position, or does it merely make a subordinate position more comfortable? 7.1 Sovereign and national compute programmes Several states have launched national computing initiatives intended to provide researchers with capacity that the market will not supply on acceptable terms. These #national_compute programmes vary enormously in scale and seriousness, from genuine multi-year investments in owned infrastructure to rebranded cloud credit schemes. The distinction matters. A programme that purchases capacity from a hyperscale provider and distributes it to researchers relieves the immediate constraint while deepening the dependency, because the capacity remains under external control and the payments continue indefinitely. A programme that builds owned infrastructure, trains the staff to operate it, and secures the energy to power it creates a durable asset. The first is consumption; the second is investment. Both are described in policy documents using the language of sovereignty, and the two should not be confused. Semi-peripheral states are the principal actors here, and their programmes should be evaluated on whether they include the complementary elements without which hardware is inert: skilled operators, sustained energy, software ecosystems, and a research community capable of using the capacity productively. A cluster without engineers is an expensive monument. 7.2 Regional consortia and pooled infrastructure For most individual universities, national programmes are out of reach and individual purchase is impossible. #regional_consortia offer a middle path. By pooling demand across institutions and across borders, universities can achieve three things: they can afford hardware that none could afford alone; they can negotiate procurement terms with a credibility that individual institutions lack; and they can share the scarce human expertise required to operate advanced systems. The model is not speculative. Research library consortia, shared telescope facilities, particle physics collaborations and national research and education networks all demonstrate that universities can build and govern shared infrastructure across institutional and national boundaries when the incentives are aligned. #south_south_cooperation in compute is a natural extension of this tradition, and it has the additional advantage that it builds governance capacity that is not dependent on the core. The obstacles are political rather than technical. Consortia require agreement on allocation rules, cost sharing, governance and dispute resolution. They require institutions to accept that they will sometimes queue behind others. And they require stable multi-year funding, which is precisely what many national systems find hardest to provide. Nevertheless, judged against the criterion stated above, pooled infrastructure genuinely changes structural position, and it deserves far more attention than it currently receives. 7.3 Open weight models and their limits The availability of models whose weights can be downloaded and run locally is the most significant recent development for institutions outside the core, and it should be neither dismissed nor overstated. The benefits are real. #open_weight_models can be run on institutionally controlled hardware, which removes the per query rent, permits work with sensitive data that cannot leave the institution, allows genuine inspection and auditing, and supports fine-tuning for local languages and local tasks. For teaching, they permit students to examine a system rather than merely to prompt one. For research, they permit reproducibility. The limits are equally real. Widder, Whittaker and West argue that openness in AI is frequently partial, and that the release of weights does not transfer the power that flows from control over the compute, data and distribution channels required to produce and deploy such systems at scale (Widder, Whittaker and West, 2024). The models are produced by firms whose strategic interests are served by their release; the licences frequently contain restrictions; the training data is usually undisclosed; and running a large open model still requires accelerators that many institutions cannot obtain. Openness lowers the barrier; it does not remove it. The realistic conclusion is that open weights convert a relationship of pure consumption into one of adaptation. That is a meaningful upgrade in structural position, comparable to the difference between importing finished goods and importing components for local assembly. It is not the same as building the components. 7.4 Small models, efficiency and frugal research A distinct strategy accepts the compute constraint and works within it. Research on smaller models, on parameter efficient fine-tuning, on quantisation, on distillation, and on data efficient training produces systems that are cheaper to build and to run. #small_language_models tuned for specific domains and specific languages frequently outperform far larger general models on the tasks that matter locally, at a fraction of the cost. This is #frugal_AI, and it has three strategic virtues. It is feasible with available resources. It aligns research effort with local need rather than with the benchmark priorities of the core. And it produces expertise that is genuinely scarce, since the ability to achieve strong results under constraint is a skill that abundant institutions have little incentive to develop. It also has a strategic risk. If peripheral institutions specialise exclusively in efficiency while core institutions retain the frontier, the division of labour is reproduced in a more sophisticated form, with the periphery performing the optimisation work whose results are absorbed by the core. Frugal research is therefore best understood as a necessary component of a strategy rather than as a complete strategy in itself. 7.5 Data governance and the leverage of the archive Universities hold assets they systematically undervalue: archives, corpora, medical records, agricultural data, linguistic material, curated collections, and the ongoing output of their scholars. In many cases these assets are unique and irreplaceable, and in some cases they are precisely what would be required to build models that work in the local context. A coherent #data_sovereignty strategy would treat these holdings as strategic rather than incidental. This does not require enclosure, and it should not be confused with a retreat from open scholarship. It requires that institutions decide deliberately, and preferably collectively, on the terms under which their data is used to build commercial systems, and that they seek a return, whether in the form of compute access, model access, revenue, or governance rights. Eke, Wakunuma and Akintoye's work on responsible AI in African contexts is instructive, since it emphasises that governance capacity must be built alongside technical capacity, and that frameworks imported wholesale from elsewhere frequently fail to address local realities (Eke, Wakunuma and Akintoye, 2023). Data governance is where the periphery has the most leverage and where it has been slowest to organise. 7.6 Curriculum, literacy and the formation of judgement Finally, there is the layer over which universities retain almost complete control: what they teach. #AI_literacy that consists only of prompting technique trains students to be efficient users of systems owned by others. #curriculum_reform that takes infrastructure seriously would teach students where models come from, what they cost, who owns them, what they consume, whose data they were built from, whose languages they serve, and how they are evaluated. It would treat the political economy of AI as core content rather than as an optional ethics module appended at the end of a technical course. Bearman, Ryan and Ajjawi's identification of a discourse of imperative change is important here, because it names the rhetorical move by which infrastructural questions are foreclosed (Bearman, Ryan and Ajjawi, 2023). If AI adoption is presented as inevitable, then the only remaining questions are procedural. If it is presented as a set of choices with distributional consequences, then a university can deliberate. Teaching students to make that distinction is itself a form of infrastructural resistance, and it costs nothing but intellectual seriousness. 8. Implications for Policy and Institutional Strategy The analysis supports a set of concrete recommendations, addressed to different actors. For individual universities. Conduct an infrastructural audit before adopting an AI strategy. The audit should locate the institution honestly on each row of Table 1, and should include the total recurring cost of current AI commitments, the switching cost of current vendor relationships, the institution's data holdings and their governance status, and the destination of its graduating researchers. Most institutions have never assembled this information in one place. Strategy without it is guesswork. Negotiate collectively. Individual negotiation with concentrated suppliers is a structurally losing position. Sector-wide procurement bodies exist for library content and should exist for compute and model access. #procurement is the most immediately available lever, and it is underused. Retain optionality. Prefer architectures and contracts that permit migration. Insist on data portability. Avoid building core academic functions on interfaces that cannot be replicated elsewhere. The cost of optionality is real and should be paid consciously, as insurance. For national systems. Distinguish between compute consumption and compute investment, and fund the latter. A national programme that only distributes credits creates neither capability nor sovereignty. Fund the complements: energy, engineers, software, and the research community that will use the capacity. Invest in the assets that cannot be bought elsewhere: national language corpora, curated public datasets, and the legal and institutional capacity to govern them. These are the foundations of any credible sovereignty strategy, and they are cheap relative to hardware. Address the #brain_drain directly rather than rhetorically. Retention requires that the work be possible at home, which is a compute question, and that the career be viable, which is a funding question. For regional and international bodies. Support #regional_consortia with multi-year funding and with the legal instruments required to operate shared facilities across borders. Support the participation of peripheral scholars in the governance and standards processes where the rules are being written, and recognise that participation without resources is decorative. Reform evaluation. Funders, journals and conferences should recognise that a compute-normalised assessment of research contribution would produce a different and more accurate picture of scientific merit than the current implicit standard, which rewards resources as though they were insight. For the research community. Treat infrastructure as a legitimate object of scholarship in education research. The field has produced a large literature on assessment and integrity and a very small literature on procurement, energy, compute and dependency. The imbalance reflects the disciplinary habits of the field rather than the importance of the questions. 9. Limitations and Directions for Future Research Four limitations should be stated plainly. First, the framework is structural, and structural analysis systematically under-weights agency. Individual universities in peripheral positions have built strong research groups, and individual scholars have made contributions that no account of their institution's resources would predict. The framework explains pressures and probabilities, not destinies. Second, the evidence base is uneven. Data on institutional compute holdings, procurement terms, and total AI expenditure is rarely public. Much of what would be required to test the propositions advanced here is held privately by vendors and by institutions themselves. The composite indicators proposed in Section 4 are offered as a research programme, not as a completed measurement. Third, the situation is changing rapidly. Model efficiency is improving, which lowers the compute cost of a given capability. Open weight releases have altered the landscape substantially within a short period. Energy-rich states are entering compute hosting. Any specific configuration described here may be superseded. The structural logic, which concerns concentration, rent and dependency, is likely to be more durable than any particular arrangement. Fourth, world-systems analysis was developed for a particular purpose and carries the risks of any macro-structural theory: it can become a template applied mechanically, and it can flatten the differences between contexts that require careful ethnographic attention. The framework should be tested against cases, not imposed on them. Four research directions follow. Comparative case studies of national compute programmes, assessing which produced durable capability and which produced dependency in a new form. Empirical measurement of the total recurring AI cost borne by university systems, disaggregated by country. Studies of consortium procurement in higher education, examining whether collective negotiation produces measurably better terms. And longitudinal work on research agendas, testing whether the distribution of compute predicts the distribution of research questions, which is the strongest single test of the argument advanced in Section 5.2. 10. Conclusion The debate about artificial intelligence in higher education has been conducted largely in the vocabulary of the classroom. Institutions have written policies on academic integrity, guidance on the disclosure of AI use, and rubrics for assessment in a world of generative tools. This work is necessary. It is also, in a structural sense, downstream. The decisive inequalities in academic AI are not pedagogical. They are infrastructural. They concern who owns the accelerators, who hosts the data centres, who supplies the energy, whose data trains the models, whose languages are represented, whose graduates staff the frontier laboratories, whose benchmarks define progress, whose standards define responsibility, and who pays rent to whom. These are questions of political economy, and world-systems analysis offers a mature framework for asking them. The picture that emerges is of a global university system being restructured into a #core_periphery hierarchy along infrastructural lines. At the core, a small number of institutions retain the capacity to build and shape AI systems, though even they are increasingly outmatched by corporate laboratories. In the semi-periphery, states pursue sovereignty projects of varying seriousness, achieving genuine but partial upgrading. In the periphery, universities consume systems they cannot inspect, in languages that may not be their own, on terms they did not negotiate, paying rent from budgets that are already inadequate, while their most talented graduates depart for the institutions that build the systems they consume. This outcome is not inevitable. Structures constrain; they do not determine. The strategies available to institutions outside the core are real: pooled infrastructure, collective procurement, open weight adoption, efficient model research, serious data governance, and a curriculum that treats infrastructure as a subject rather than as a background condition. What these strategies share is that they require institutions to act collectively and to think in decades. What they require, above all, is that universities recognise the nature of the problem. A university that believes its AI challenge is a plagiarism problem will solve the plagiarism problem and lose its autonomy. A university that understands its AI challenge as an infrastructural and geopolitical one may still lose, but it will at least be playing the right game. The oldest insight of world-systems analysis is that positions in a division of labour are reproduced by the ordinary, rational, day-to-day decisions of the actors within it. Each individual decision to rent rather than to build, to consume rather than to construct, to adopt rather than to adapt, is defensible on its own terms. It is the aggregate that constitutes the structure. Recognising this is the beginning of the possibility of choosing otherwise. Hashtags #WorldSystemsAnalysis #AIInfrastructure #CoreAndPeriphery #GlobalHigherEducation #ComputeInequality #DigitalColonialism #AIGovernance #TechnologicalSovereignty #ResourceAllocation #ScholarlyEquity #AIandEducation #GlobalSouthResearch #CriticalDataStudies #InfrastructureStudies References Ahmed, N., Wahed, M. and Thompson, N. C. (2023) The growing influence of industry in AI research. Science, 379(6635), pp. 884-886. Bearman, M., Ryan, J. and Ajjawi, R. (2023) Discourses of artificial intelligence in higher education: a critical literature review. Higher Education, 86(2), pp. 369-385. Bender, E. M., Gebru, T., McMillan-Major, A. and Shmitchell, S. (2021) On the dangers of stochastic parrots: can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 610-623. Besiroglu, T., Bergerson, S. A., Michael, A., Heim, L., Luo, X. and Thompson, N. (2024) The compute divide in machine learning: a threat to academic contribution and scrutiny? arXiv preprint, 2401.02452. Blasi, D., Anastasopoulos, A. and Neubig, G. (2022) Systematic inequalities in language technology performance across the world's languages. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pp. 5486-5505. Bozkurt, A., Xiao, J., Lambert, S., Pazurek, A., Crompton, H., Koseoglu, S., Farrow, R., Bond, M., Nerantzi, C., Honeychurch, S., Bali, M., Dron, J., Mir, K., Stewart, B., Costello, E., Mason, J., Stracke, C. M., Romero-Hall, E., Koutropoulos, A., Toquero, C. M., Singh, L., Tlili, A., Lee, K., Nichols, M., Ossiannilsson, E., Brown, M., Irvine, V., Raffaghelli, J. E., Santos-Hermosa, G., Farrell, O., Adam, T., Thong, Y. L., Sani-Bozkurt, S., Sharma, R. C., Hrastinski, S. and Jandric, P. (2023) Speculative futures on ChatGPT and generative artificial intelligence: a collective reflection from the educational landscape. Asian Journal of Distance Education, 18(1), pp. 53-130. Crawford, K. (2021) Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press. De Vries, A. (2023) The growing energy footprint of artificial intelligence. Joule, 7(10), pp. 2191-2194. Eke, D. O., Wakunuma, K. and Akintoye, S. (eds.) (2023) Responsible AI in Africa: Challenges and Opportunities. Cham: Palgrave Macmillan. Joshi, P., Santy, S., Budhiraja, A., Bali, K. and Choudhury, M. (2020) The state and fate of linguistic diversity and inclusion in the NLP world. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 6282-6293. Klinger, J., Mateos-Garcia, J. and Stathoulopoulos, K. (2024) A narrowing of AI research? Technological Forecasting and Social Change, 200, 123145. Lehdonvirta, V., Wu, B. and Hawkins, Z. (2024) Compute North vs. Compute South: the uneven possibilities of compute-based AI governance around the globe. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), pp. 828-838. Li, P., Yang, J., Islam, M. A. and Ren, S. (2025) Making AI less thirsty: uncovering and addressing the secret water footprint of AI models. Communications of the ACM, 68(7), pp. 54-63. Muldoon, J. and Wu, B. A. (2023) Artificial intelligence in the colonial matrix of power. Philosophy and Technology, 36(4), article 80. Png, M.-T. (2022) At the tensions of South and North: critical roles of Global South stakeholders in AI governance. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pp. 1434-1445. Ricaurte, P. (2022) Ethics for the majority world: AI and the question of violence at scale. Media, Culture and Society, 44(4), pp. 726-745. Selwyn, N. (2022) The future of AI and education: some cautionary notes. European Journal of Education, 57(4), pp. 620-631. Tacheva, J. and Ramasubramanian, S. (2023) AI Empire: unravelling the interlocking systems of oppression in generative AI's global order. Big Data and Society, 10(2), pp. 1-13. Verdegem, P. (2024) Dismantling AI capitalism: the commons as an alternative to the power concentration of Big Tech. AI and Society, 39(2), pp. 727-737. Wallerstein, I. (2004) World-Systems Analysis: An Introduction. Durham: Duke University Press. Whittaker, M. (2021) The steep cost of capture. Interactions, 28(6), pp. 50-55. Widder, D. G., Whittaker, M. and West, S. M. (2024) Why 'open' AI systems are actually closed, and why this matters. Nature, 635, pp. 827-833. Williamson, B. (2024) The social life of AI in education. International Journal of Artificial Intelligence in Education, 34(1), pp. 97-104.

Senior leaders increasingly make consequential choices alongside machine learning systems that recommend, rank, score, forecast, and sometimes decide. Most research on this shift is framed around individual cognition, such as whether a manager trusts or resists a model, or around technology adoption. This article argues that both framings miss the deeper social process at work. Drawing on Pierre Bourdieu's theory of social practice, it develops the concept of the #algorithmic_habitus of leadership: a durable set of embodied dispositions through which executives perceive, evaluate, and act on machine outputs without conscious deliberation. The article reconstructs four Bourdieusian tools for the context of #hybrid_teams. First, it defines the hybrid field as a structured space of positions in which humans and machines are unequally endowed with resources. Second, it proposes #computational_capital as a species of capital that is convertible into authority, budget, and promotion. Third, it identifies #algorithmic_doxa as the taken-for-granted belief that quantified outputs deserve epistemic priority. Fourth, it applies the idea of #hysteresis to explain why experienced leaders, whose instincts were formed in a pre-algorithmic world, sometimes act in ways that no longer fit the field they occupy. The article synthesises recent organisational and sociological research published between 2020 and 2025 and derives twelve propositions and five mechanisms that link field position, capital endowment, and dispositional formation to observable patterns of #executive_decision_making. The contribution is threefold. Theoretically, it moves the conversation from trust to disposition, and from adoption to position. Methodologically, it offers a relational research agenda that studies leaders and models as co-located points in the same field rather than as separate agents in a dyad. Practically, it reframes leadership development as the cultivation of #reflexivity about the social conditions that make some machine outputs feel obviously right and others feel obviously wrong. The article concludes that the central question for governance is not whether leaders trust algorithms, but which leaders, positioned where, come to find algorithmic reasoning natural, and what is lost when that naturalness goes unexamined. Keywords Bourdieu, habitus, field theory, algorithmic decision-making, human-machine hybrid teams, executive leadership, computational capital, symbolic power, organisational practice, artificial intelligence in management 1. Introduction A regional director at a logistics company opens a planning dashboard on a Monday morning. The system has ranked twelve depots by predicted risk of missed delivery windows. She glances at the list, agrees with nine of the rankings, feels a flicker of unease about the tenth, and overrides it. She does not run a calculation. She does not consult a decision matrix. She simply knows, in a way she would struggle to explain, that the model has misread something about that particular depot. Three weeks later she is proved right. When asked how she knew, she says she had a feeling. That feeling is the subject of this article. Research on #artificial_intelligence in organisations has grown quickly, and much of it is valuable. We now know a great deal about when people accept or reject machine advice, about the design of #decision_support_systems, and about the effects of automated evaluation on frontline workers. Yet the dominant framings share a hidden assumption. They treat the human decision-maker as a rational, bounded, calculating individual who weighs the machine's suggestion against private information and then chooses. The vocabulary is that of #trust_calibration, #algorithm_aversion, and #algorithm_appreciation. In this vocabulary, the director's feeling is noise, a bias to be corrected, or at best an unexplained residual called intuition. Pierre Bourdieu spent a career arguing that such residuals are not residuals at all. They are the visible tip of a social structure that has been deposited in the body over time. What we call intuition is, in his account, #practical_sense: a feel for the game acquired through prolonged exposure to the conditions of a particular social world. This article takes that argument seriously and applies it to the newest and least examined of social worlds, the #hybrid_teams in which senior humans and learning machines now share the work of judgment. 1.1 The problem Three observations motivate this study. The first is empirical. Executives do not treat all machine outputs the same way. The same leader may accept a demand forecast without question and dismiss a talent-attrition score with contempt. Individual-level explanations struggle here, because the individual is constant and the behaviour varies. Something about the relationship between the leader's position, the domain, and the model must be doing the work. The second is theoretical. Existing accounts of human-machine collaboration tend to be dyadic. They describe an agent and a tool, or two agents, and study the interaction between them. But leadership does not happen in a dyad. It happens in a #hybrid_field crowded with peers, rivals, boards, analysts, consultants, regulators, and the reputational stakes that bind them together. A manager who publicly overrides a model is not simply making a judgment. She is making a claim about who is entitled to know. The third is normative. If the disposition to defer to machines is socially produced rather than individually chosen, then the standard remedies of AI #governance, which centre on training, disclosure, and explanation, are aimed at the wrong target. You cannot debias a #habitus with a slide deck. 1.2 Research questions This article asks: How can Bourdieu's concepts of habitus, field, capital, and doxa be reconstructed to explain executive practice in human-machine hybrid teams? Through what mechanisms does prolonged exposure to algorithmic systems restructure the perceptual and evaluative dispositions of senior leaders? What are the consequences for #legitimacy, #accountability, and the distribution of #power_asymmetry inside organisations? 1.3 Contribution and structure The article contributes a relational theory of algorithmic leadership. Section 2 reviews the state of research on hybrid teams and on the sociology of algorithms, and shows where each stops short. Section 3 sets out the Bourdieusian apparatus and reconstructs it for the algorithmic case, introducing #computational_capital, #algorithmic_doxa, and the hybrid field. Section 4 presents a conceptual model with twelve propositions. Section 5 explains the method, which is an integrative conceptual synthesis illustrated with published empirical cases. Section 6 develops five mechanisms of dispositional formation. Section 7 discusses theoretical implications, Section 8 practical ones, and Section 9 sets out limitations and a research agenda. Section 10 concludes. 2. Literature Review 2.1 Human-machine hybrid teams and the augmentation debate The management literature has largely organised itself around a single distinction: does the machine replace the human or extend her? Raisch and Krakowski (2021) name this the automation-augmentation paradox and show that the two are not alternatives but a loop. Firms pursue #augmentation, which produces data, which enables #automation, which reduces the human learning that made augmentation valuable. The paradox is structural, not a design error. Murray, Rhymer, and Sirmon (2021) push further with the idea of conjoined agency. They argue that in many settings we cannot cleanly separate the human contribution from the machine contribution, because the two are constituted together in the act of deciding. Puranam (2021) treats the resulting arrangement as an organisation design problem: the question is not who is smarter but how tasks, information, and authority should be partitioned between agents with different error profiles. Empirical studies have added texture. Lebovitz, Levina, and Lifshitz-Assaf (2021) show that the #ground_truth on which models are trained is often itself a professional construction, contested among experts, so that a model's apparent objectivity conceals a settled dispute. In a companion study, Lebovitz, Lifshitz-Assaf, and Levina (2022) find that professionals who engage seriously with AI outputs for high-stakes judgments experience a loss of confidence rather than a gain, because engagement exposes the gap between the model's confidence and its explicability. Anthony (2021) demonstrates that when analysts are shielded from a system's inner workings, they develop a form of confident practice built on the very opacity that ought to trouble them, a phenomenon she calls black boxing. Anthony, Bechky, and Fayard (2023) argue that the term collaboration itself misleads, because it imports assumptions of mutual intent that do not hold, and they call instead for a systems view of the whole configuration of people, tools, and data. Waardenburg, Huysman, and Sergeeva (2022) show that in a police setting a new occupational role emerged, the algorithmic knowledge broker, whose #knowledge_brokerage work of interpreting and translating machine output became indispensable and, over time, a source of influence. Grønsund and Aanestad (2020) describe similar human-in-the-loop configurations forming around algorithmic systems, with new tasks of auditing and correcting. This body of work is rich, but it is overwhelmingly focused on professionals and frontline workers. The executive level is comparatively neglected. Von Quaquebeke and Gerpott (2023) are among the few to consider what AI does to leadership as such, arguing that many classical leadership functions, from goal setting to feedback, are becoming machine-mediated. Bankins, Ocampo, Marrone, Restubog, and Woo (2024) provide a multilevel review and note explicitly that the strategic apex remains understudied. 2.2 Algorithmic management and control A second stream examines algorithms as instruments of control. Kellogg, Valentine, and Christin (2020) map #algorithmic_management onto six mechanisms of direction, evaluation, and discipline, and argue that algorithms constitute a new contested terrain. Rahman (2021) shows that opaque evaluation produces an invisible cage: workers cannot see the criteria, so they engage in anxious, generalised effort rather than targeted improvement. Curchod, Patriotta, Cohen, and Neysen (2020) document the power asymmetries that arise when sellers on a platform cannot contest the ratings that determine their fate. Christin (2020) shows how #metrics travel across occupational settings and are reinterpreted according to local values, so that the same number means different things in a newsroom and a courtroom. Leonardi and Treem (2020) introduce behavioural visibility, the idea that digital traces make work observable in unprecedented ways, and that this observability itself changes behaviour. Vaast and Pinsonneault (2021) show that when digital technologies both enable and threaten #occupational_identity, professionals engage in identity work to reconcile the two. Again, the empirical gaze is directed downward. The literature has documented what algorithms do to those who are managed. It has said far less about what algorithms do to those who manage, and almost nothing about what they do to those who manage the managers. 2.3 The sociology of algorithms and the Bourdieusian opening The most promising theoretical resources lie outside management. Burrell and Fourcade (2021) describe the emergence of a society of algorithms in which a coding elite accumulates the authority to define what counts as knowledge, while a cybertariat is subject to that definition. Fourcade and Johns (2020) analyse the recursive loops by which social learning and machine learning shape one another. Crawford (2021) insists that AI is neither artificial nor intelligent but a material and political formation. Amoore (2020) argues that algorithms do not merely apply rules but generate a form of ethics through the way they cut the world into categories. Seaver (2022) shows, through ethnography of recommender developers, that #taste is not discovered by machines but manufactured through the assumptions engineers embed in them. The decisive contribution is Airoldi's (2022) concept of #machine_habitus. Airoldi argues that algorithms are not neutral calculators but socialised agents: they are formed by training data that carry the traces of past social structures, and they then act back on the social world, reproducing those structures. This is a genuinely Bourdieusian move, and it inverts the usual question. Instead of asking how humans respond to machines, it asks how machines and humans jointly constitute a field of practice. Airoldi and Rokka (2022) extend this to consumption, showing how algorithmic mediation shapes cultural classification. Romele (2021) offers a parallel philosophical move, proposing technological capital as a Bourdieusian species of resource, and arguing that the philosophy of technology needs Bourdieu to explain why some technical competences confer authority and others do not. Ignatow and Robinson (2017), writing earlier, made an initial case for theorising the digital with Bourdieu and identified the danger of treating digital skill as mere human capital rather than as a positional resource. 2.4 The gap Bringing these streams together reveals a specific and addressable gap. Management scholarship has studied human-machine hybrid work extensively but has framed the human side in cognitive and attitudinal terms. Sociology has developed a relational and structural account of algorithms but has applied it mainly to platforms, consumers, and street-level professionals. Nobody has yet applied the full Bourdieusian apparatus, habitus and field and capital together, to the specific case of #executive_decision_making, where the stakes for #strategic_judgment and #corporate_governance are highest and where the dispositions in question are those of the people who decide what the organisation will do. This article addresses that gap. 3. Theoretical Framework: Bourdieu Reconstructed for the Hybrid Field 3.1 The core apparatus Bourdieu's theory rests on a formula he offered half in jest: practice is the product of the relation between habitus and capital within a field. It is worth taking each term slowly, because the argument that follows depends on precision. Habitus. The #habitus is a system of durable, transposable #dispositions. It is acquired through experience, especially early and repeated experience, and it operates below the level of conscious calculation. Bourdieu (1990) describes it as a feel for the game. A skilled player does not compute where to run; she finds herself already running. The habitus generates practices that are regular without being the product of obedience to rules, and adjusted to circumstances without being the product of deliberate strategy. Crucially, the habitus is a structured structure that functions as a structuring structure: it is shaped by social conditions and it shapes perception of those conditions in turn. Field. A field is a structured space of positions defined by the distribution of resources relevant to that space. Fields have their own logic and their own stakes. What counts as success in the field of investment banking is not what counts as success in the field of academic physics. Every field has dominant and dominated positions, and the struggle between them is a struggle over the very definition of what is valuable. Capital. Capital is any resource that is effective in a given field. Bourdieu distinguished economic capital, #cultural_capital in its embodied, objectified, and institutionalised forms, #social_capital as durable networks of relationship, and #symbolic_capital, which is any of the others when recognised as legitimate rather than as mere advantage. The convertibility of capitals is central: economic capital can buy education, which becomes cultural capital, which can be converted into credentials, which confer symbolic capital. Doxa and illusio. Doxa is the set of assumptions so basic to a field that they are not stated because they are not noticed. #Illusio is the investment in the game, the belief that the stakes are worth pursuing. Both are conditions of participation. Symbolic violence. #symbolic_violence occurs when the dominated perceive the arbitrary hierarchies of the field as natural and legitimate, and so participate in their own subordination without coercion. The power is invisible precisely because it is misrecognised. Hysteresis. Because the habitus is formed slowly and lags behind changing conditions, agents whose dispositions were shaped under one set of field conditions may act inappropriately when conditions change. Bourdieu called this the Don Quixote effect. #hysteresis is the mismatch between an old habitus and a new field. Emirbayer and Johnson (2008) argued that organisational research had adopted Bourdieu's vocabulary while ignoring his relational method, borrowing capital and habitus as variables while discarding the field analysis that gives them meaning. This article tries not to repeat that error. 3.2 The hybrid field The first reconstruction is the field itself. A #hybrid_field is a structured space of positions in which both human and non-human agents hold positions, and in which the distribution of capital across those positions structures practice. Three claims follow. First, the machine occupies a position, not merely a role. This is not a mystical claim about machine consciousness. It is a structural claim. A forecasting system that consistently produces outputs which are cited in board papers, which shape budget allocation, and which nobody can afford to contradict without a written justification, occupies a dominant position in the field, in exactly the way an accounting standard or a credit rating agency occupies a position. Its power is relational, not psychological. Second, the machine's position is not fixed but is itself the outcome of struggle. Pachidi, Berends, Faraj, and Huysman (2021) show how the introduction of an analytics system into a sales organisation triggered a contest between two regimes of knowing, one grounded in relational expertise and one in data. The eventual dominance of the analytic regime was not a technical inevitability but a political outcome, achieved partly through symbolic actions. Third, the field is hierarchical in a specific way: the closer an agent stands to the definition of what counts as evidence, the more dominant its position. This is why data science teams, which are often junior in formal hierarchy, can wield disproportionate influence. They stand near the doxa. 3.3 Computational capital The second reconstruction concerns capital. I propose #computational_capital as a distinct species of capital operative in hybrid fields. Following Romele (2021), but specifying it for organisations, computational capital exists in three states, mirroring Bourdieu's treatment of cultural capital. Embodied. The fluency to read a model card, to sense when a distribution has shifted, to ask why the confidence interval is that wide. This is a bodily competence, acquired slowly, and it shows in the speed with which a leader can look at a dashboard and find the anomaly. Like all embodied capital, it cannot be transmitted instantaneously and it carries the marks of its conditions of acquisition. An executive who learned to read data in a quantitative hedge fund carries a different embodied capital from one who learned it in a public health agency. Objectified. The infrastructure itself: the pipelines, feature stores, model registries, and the #data_infrastructure that a leader can command. A leader who controls a data platform controls the conditions under which evidence can be produced. Objectified computational capital is only effective in combination with the embodied capital to use it, which is precisely Bourdieu's point about books in a library. Institutionalised. Credentials, certifications, formal titles such as chief data officer, and membership of the AI steering committee. These convert competence into recognised authority and make it portable across organisations. The critical property is convertibility. Computational capital converts into #symbolic_capital when an executive's data fluency is recognised as a mark of the modern leader rather than as a technical specialism. It converts into economic capital through promotion and budget. And, in the other direction, economic capital purchases computational capital in the form of consultants, platforms, and hires. The distribution of computational capital is not random. It follows existing lines of advantage: age, function, education, and the prestige of prior employers. This is why the arrival of #artificial_intelligence in the executive suite functions less as a levelling force than as a reshuffling of an already ordered pack. 3.4 Algorithmic doxa The third reconstruction concerns doxa. #algorithmic_doxa is the pre-reflective assumption, shared by participants in a hybrid field, that quantified, model-generated representations of reality carry a stronger claim to truth than other representations, and that the burden of justification therefore falls on whoever departs from them. Doxa is recognisable by an asymmetry of justification. In a field governed by algorithmic doxa, a leader who follows the model's recommendation need say nothing. A leader who overrides it must produce reasons, document them, and accept the risk of later being shown to have been wrong. The model's recommendation is the default state of the world; the human judgment is the deviation. Nobody decided this. It is simply how the meeting works. Three features deserve emphasis. First, algorithmic doxa is not the same as trust. A leader may distrust a model in the sense of doubting its accuracy and still operate wholly within algorithmic doxa, because her doubt takes the form of demanding a better model rather than a different kind of evidence. Doxa concerns the terms of the argument, not the position taken within it. Second, algorithmic doxa produces #symbolic_violence in a precise sense. Those whose knowledge is not machine-readable, which typically means those whose expertise is relational, tacit, or embodied, come to experience their own knowledge as soft, anecdotal, and merely intuitive. They adopt the classification that devalues them. Rahman's (2021) invisible cage describes this at the worker level; the same logic operates in the boardroom, where the head of a care service or a customer relationship function learns to preface her contributions with an apology for not having the numbers. Third, doxa is sustained by the very opacity that critics attack. Anthony (2021) shows that black boxing does not reduce confidence but relocates it: unable to interrogate the system, users invest their confidence in it as a whole. The #black_box is not an obstacle to authority; under certain conditions it is a condition of authority, in the way that the mystery of a rite sustains the standing of the priest. 3.5 Algorithmic habitus We can now define the central construct. The #algorithmic_habitus of leadership is the system of durable dispositions, acquired through repeated exposure to algorithmically mediated decision environments and structured by the leader's position in the hybrid field, that generates her perception, appreciation, and action with respect to machine outputs without passing through conscious deliberation. It has three components, following Bourdieu's own tripartite characterisation of habitus as schemes of perception, appreciation, and action. Schemes of perception. What the leader notices. An executive with a formed algorithmic habitus does not see a dashboard as a wall of numbers. She sees, immediately, the one series that is behaving oddly. Equally, and more consequentially, she does not see what the dashboard omits. The formation of perception is simultaneously a formation of blindness. What is not instrumented does not appear, and what does not appear is not managed. Schemes of appreciation. What the leader values. This is the domain of taste, and Bourdieu's insight in Distinction applies with uncomfortable force. Taste is a marker of position and a mechanism of #distinction. In many contemporary firms, a preference for evidence over anecdote, for the experiment over the hunch, for the dashboard over the site visit, functions exactly as a taste for classical music once did: it distinguishes those who possess it, and it disqualifies those who do not, while presenting itself as merely a matter of rigour. Schemes of action. What the leader does, fluently and without hesitation. Whether she opens the model before the meeting or after. Whether her first question is about the confidence interval or about the customer. Whether her override is written as a challenge to the model or as an addition to it. The algorithmic habitus explains the logistics director in the introduction. Her feeling about the tenth depot was not a mystical faculty. It was the sedimented product of years in which she had watched that particular model fail in that particular way, and the perception of the anomaly had become automatic. Her practical sense was, in Bourdieu's phrase, history turned into nature. 3.6 Machine hysteresis Finally, #hysteresis. When a field changes faster than the dispositions of those who occupy it, agents misplay the game. Two forms are relevant here, and they are mirror images. Legacy hysteresis. The senior leader whose habitus was formed in a pre-algorithmic field continues to trust the site visit, the long lunch, and the relationship. In the new field these dispositions, once the very definition of judgment, are reclassified as sentimentality. She is not less competent than she was; the field has moved beneath her. Her practice remains coherent but has ceased to be legitimate. Anticipatory hysteresis. The leader whose habitus was formed entirely inside data-rich environments carries dispositions that fit the instrumented parts of the field beautifully and the uninstrumented parts not at all. In a crisis, when the data distribution breaks and the model's training set no longer describes the world, she has no practical sense to fall back on. She has never developed the feel for the game that only the uninstrumented world can teach. This is arguably the more dangerous condition, because it is invisible in normal conditions and catastrophic in abnormal ones. Hysteresis is not a personal failing. It is the ordinary consequence of the fact that dispositions are formed slowly and fields can change quickly. 4. Conceptual Model and Propositions The framework can be stated as a model with three linked stages. Field position and capital endowment shape exposure. Exposure, over time, forms disposition. Disposition generates practice, and practice reproduces the field. The loop is closed, which is what makes it a theory of reproduction rather than a theory of adoption. 4.1 Propositions on field and position P1. The greater an executive's dependence on machine-generated evidence for the legitimacy of her decisions in the eyes of superiors, the more strongly her perceptual dispositions will align with the categories encoded in the model, independent of her personal beliefs about the model's accuracy. P2. The position of an algorithmic system in the hybrid field is determined not by its technical performance but by the volume of capital held by the actors who sponsor it. A mediocre model sponsored by the chief executive occupies a more dominant position than an excellent model sponsored by a middle manager. P3. In fields where the criteria of evaluation are contested, the introduction of an algorithmic system will not resolve the contest but will displace it into a struggle over the definition of the target variable, the training population, and the metric. 4.2 Propositions on capital P4. #computational_capital is convertible into symbolic capital at a rate that varies by field. In fields whose doxa is already quantitative, such as finance, the conversion rate is high and the capital is quickly taken for granted. In fields whose doxa is relational or aesthetic, the conversion rate is initially low, and early holders of computational capital act as challengers rather than incumbents. P5. Executives who lack embodied computational capital but possess objectified computational capital, meaning they command the infrastructure without being able to read it, will exhibit systematically higher deference to machine outputs than those who possess both, because they lack the practical means to locate the point at which a model becomes questionable. P6. The conversion of computational capital into authority will be contested most fiercely by holders of the capital it threatens to devalue, and this contest will be conducted in the language of risk, ethics, and prudence rather than in the language of position. 4.3 Propositions on doxa and symbolic power P7. Where #algorithmic_doxa is established, the cost of overriding a model is asymmetric: the reputational penalty for an unsuccessful override exceeds the reward for a successful one. Executives will therefore under-override relative to what their private information would justify. P8. The asymmetry in P7 increases with the visibility of the decision and with the formality of the audit trail. Systems designed to increase #accountability by logging overrides will, by that very design, reduce the number of overrides, including the correct ones. P9. Executives whose expertise is not representable in the model's feature space will progressively reclassify their own knowledge as subjective, and will voluntarily reduce their contributions in mixed forums. This is #symbolic_violence in the strict sense, requiring no coercion. 4.4 Propositions on habitus formation and hysteresis P10. The algorithmic habitus is formed most rapidly under conditions of high-frequency, low-stakes feedback, and most slowly under conditions of low-frequency, high-stakes feedback. Executives therefore develop reliable practical sense about operational models and unreliable practical sense about strategic ones, while the confidence generated by the former transfers to the latter. P11. Legacy hysteresis manifests as an over-reliance on relational evidence in instrumented domains; anticipatory hysteresis manifests as an inability to act in the absence of data. Both increase organisational fragility, but only the first is visible to the organisation, and only the first is punished. P12. When a model fails publicly, the attribution of responsibility follows field position rather than causal contribution. The human closest to the point of output, typically a middle manager, absorbs the blame, functioning as a #moral_crumple_zone, while the actors who defined the target variable and sponsored the system remain unmarked. 5. Method 5.1 Approach This is a conceptual article. Its purpose is theory construction, not hypothesis testing, and its method is integrative conceptual synthesis. The procedure had four steps. First, the Bourdieusian apparatus was reconstructed from the primary texts, with attention to the internal logic of the concepts rather than to their common borrowings. The point of returning to the primary sources is that habitus and capital are routinely detached from field analysis and used as freestanding variables, a practice Emirbayer and Johnson (2008) identified as fatal to their explanatory power. Second, a body of recent empirical and theoretical research on AI in organisations, published mainly between 2020 and 2025, was assembled across management, information systems, and sociology. Studies were selected for their capacity to speak to the relations among position, resource, and practice, rather than for coverage of a defined population. Third, each Bourdieusian concept was tested against this material by asking what it would have to mean in order to apply, and whether, so specified, it explained anything that existing concepts did not. Concepts that failed this test were discarded. Concepts that survived were retained and specified as constructs. Fourth, the resulting constructs were assembled into the model and propositions set out above. 5.2 The status of the illustrations The vignettes and examples in this article, including the logistics director, are composite illustrations constructed to make abstract relations legible. They are not data. They do not report observed events, and no claim is made that they describe any particular firm or person. Where empirical claims are made, they rest on the cited published studies, and readers are directed to those studies for evidence. This distinction matters, and it is stated plainly because conceptual articles in this area sometimes blur it. The empirical foundation of this article is the work of others. Its contribution is the frame. 5.3 Boundary conditions The framework is developed for organisations in which senior leaders routinely encounter machine outputs that carry institutional weight. It assumes a field with a degree of stability, sufficient for dispositions to form. It is less applicable to very small firms without data infrastructure, to settings where AI use is purely experimental, and to fields in such rapid flux that no habitus can settle. These boundaries are themselves interesting, since a field too unstable to produce habitus is also a field in which no one can develop practical sense, and everyone is a novice. 6. Analysis: Five Mechanisms of Dispositional Formation The propositions describe relations. This section specifies the mechanisms that produce them: the actual processes by which a leader comes to have an algorithmic habitus. Each mechanism is stated, illustrated, and connected to the empirical literature. 6.1 Mechanism one: perceptual entrainment through instrumentation The first mechanism operates through the eye. An executive's attention is a scarce resource, and it is allocated by the environment as much as by the will. A dashboard is not a neutral window onto the organisation. It is a selection, and the selection was made by someone. Leonardi and Treem (2020) show that behavioural visibility restructures what is available to be known. What is instrumented becomes visible; what is visible becomes discussable; what is discussable becomes actionable; and what is actionable becomes, in practice, the whole of the manageable world. The mechanism is entrainment. Repeated exposure to a particular #quantification of the organisation trains the perceptual schemes to expect that form, to find it easy, and to find its absence uncomfortable. After sufficient exposure, the leader does not experience the dashboard as a representation of the business. She experiences it as the business. The map has become the territory not through error but through habituation. The consequence for practice is systematic and quiet. Domains that resist instrumentation, and these are typically those involving relationships, morale, craft, and long time horizons, do not merely receive less attention. They come to feel less real. A leader asked about them experiences a peculiar vagueness, which she interprets as a lack of information, and which is in fact a lack of perceptual training. This is the most under-recognised harm of #dashboard_culture. It is not that leaders are given wrong numbers. It is that they are gradually rendered unable to think in anything but numbers, and this incapacity presents itself as rigour. 6.2 Mechanism two: evaluative alignment through recognition The second mechanism operates through the desire to be respected. Bourdieu's account of #illusio holds that agents invest in a field's stakes because they have been formed to find those stakes worth pursuing. In the contemporary corporate field, the stakes have shifted. Being seen as data-driven is now a form of #symbolic_capital, and being seen as instinct-driven is a liability. The executive who says she has a feeling about the depot is taking a risk that the executive who cites the model is not. Recognition therefore does the work of persuasion. A leader does not need to be convinced that models are superior. She needs only to notice, repeatedly and pre-consciously, which contributions are received well in the room. Over time her evaluative schemes align with the pattern of recognition. She begins, without deciding to, to find quantified arguments more satisfying, and she begins to find her own unquantified intuitions embarrassing. Vaast and Pinsonneault (2021) document the identity work that this provokes when a technology both enables and threatens professional identity. Pachidi and colleagues (2021) document the symbolic contest that accompanies it at the collective level. What the Bourdieusian frame adds is the observation that the winner of such contests need not persuade anybody of anything. The winner needs only to define what a good contribution looks like, and let the ordinary human wish to be esteemed do the rest. This is why exhortations to keep human judgment in the loop so often fail. They ask leaders to spend #symbolic_capital they cannot afford. 6.3 Mechanism three: capital conversion and the sorting of careers The third mechanism operates through promotion. If #computational_capital converts into authority, then over time the population of executives will change. Those whose dispositions fit the new field will rise; those whose dispositions do not will plateau. Nobody need be fired. The mechanism is selection, not conversion, and it is far more powerful than persuasion because it operates on the composition of the group rather than on the beliefs of individuals. Burrell and Fourcade (2021) describe precisely this at the societal level, identifying the rise of a coding elite whose control over the means of classification confers a durable advantage. Within the firm, the same logic produces a slow shift in the social profile of the senior team: more quantitative training, more experience in data-rich sectors, more comfort with abstraction, less time spent in operational or customer-facing roles where tacit knowledge accumulates. The organisational consequence is a loss of variety in the habitus of the senior team, and therefore a loss of variety in what the senior team can perceive. Balasubramanian, Ye, and Xu (2022) argue that substituting machine learning for human decision-making degrades #organizational_learning, because humans stop generating the varied experience from which learning derives. The Bourdieusian account adds a second, slower degradation: the organisation stops recruiting and promoting the kinds of people who would have had the dispositions to notice what the models miss. The result is an executive team that agrees with the model not because it has been persuaded but because it has been assembled. 6.4 Mechanism four: doxic closure through the audit trail The fourth mechanism operates through the record. Governance regimes for AI almost universally recommend that human overrides be documented. The intent is accountability. The effect, in a field with #algorithmic_doxa, is closure. Consider the asymmetry described in Proposition 7. To follow the model is to leave no trace, since agreement generates no artefact. To override it is to create a document, in one's own name, that will be read in the event of failure and will not be read in the event of success. The rational response of a career-conscious executive is obvious. She will override only when she is very confident, which means she will override less often than her information warrants, which means the model will appear to be right more often than it is, which will strengthen the doxa, which will raise the cost of the next override. This is doxic closure: a self-sealing loop in which the apparent performance of the system is partly an artefact of the reluctance to contradict it. The loop is invisible from inside, because the only evidence of the errors not caught is a counterfactual that nobody generates. Elish (2019) named the endpoint of this logic the #moral_crumple_zone: when the system fails, the human operator, who had neither the information nor the standing to intervene, absorbs the blame, and the structure that placed her in that position remains unexamined. Proposition 12 restates this in field terms. Blame flows to the position with the least capital, not to the position with the most causal responsibility. The practical implication is sharp and counterintuitive. #accountability mechanisms that focus exclusively on the human at the point of output will reliably produce the appearance of human control and the reality of machine authority. 6.5 Mechanism five: hysteresis and the crisis The fifth mechanism operates through time, and it becomes visible only when the world breaks. A model is a compression of the past. It performs well when the future resembles its training distribution and poorly when it does not. Crises, by definition, are the moments when the future does not resemble the past. They are therefore precisely the moments when the algorithmic habitus is least useful and when the organisation most needs the dispositions that the previous four mechanisms have been quietly eliminating. Consider what a leader with a fully formed algorithmic habitus does when the data goes strange. Her perceptual schemes register anomaly but supply no interpretation. Her evaluative schemes tell her that unquantified claims are weak. Her action schemes reach for the dashboard. And the dashboard, which is a compression of a world that no longer exists, tells her calmly that everything is within range. Lebovitz and colleagues (2022) found that deep engagement with AI in high-stakes judgment produced not confidence but uncertainty, because engagement revealed that the model could not explain itself. That finding is normally read as a problem. Read through Bourdieu, it is a resource: uncertainty is the beginning of #reflexivity, the moment at which doxa becomes visible as doxa and can be argued about. The organisational task, then, is not to eliminate the discomfort of the leader who cannot justify her hunch. It is to preserve it. An organisation that has trained all such discomfort out of its senior team has optimised itself for the world of the training set and has disarmed itself for every other world. 7. Discussion 7.1 From trust to disposition The dominant frame in this literature is trust. Glikson and Woolley (2020) review the empirical research on human trust in AI, and Mahmud, Islam, Ahmed, and Smolander (2022) synthesise the work on #algorithm_aversion. This body of work is careful and useful, but it rests on an image of the decision-maker as an individual who forms a belief about a tool and calibrates her reliance accordingly. The Bourdieusian frame proposes a different image. Reliance is not primarily a belief. It is a disposition, formed in a field, sustained by recognition, and reproduced through selection. This reframing has consequences. It explains variance that the trust frame struggles with, such as why the same executive treats two models differently, or why leaders who explicitly say they distrust a model nonetheless follow it. It explains why interventions aimed at belief, such as training and explanation, have modest effects. And it dissolves the puzzle of aversion and appreciation, which the individual-level literature treats as competing findings. Logg, Minson, and Moore (2019) found appreciation; earlier work found aversion. In field terms these are not competing findings but observations of different fields with different doxa. The question is not whether people prefer algorithms but under what conditions of position and capital the preference is produced. 7.2 From adoption to position The second reframing concerns the unit of analysis. Adoption research asks whether an individual or organisation takes up a technology. Field analysis asks how the arrival of the technology redistributes capital, and who gains and who loses by the redistribution. This turns a technical question into a political one, and rightly so. Berente, Gu, Recker, and Santhanam (2021) frame the management of AI in terms of the tension between autonomy, inscrutability, and learning. Faraj and Pachidi (2021) argue that technology and organising are co-constituted, so that neither can be treated as the independent variable. The field frame extends both by giving a precise account of the currency in which the political contest is conducted, which is capital, and of the terrain over which it is fought, which is the definition of legitimate knowledge. Csaszar and Steinberger (2022) note that organisation theory has long used AI as an analogy for the firm. The relation runs in both directions. The firm is also becoming a machine for producing the categories through which it will subsequently see itself. 7.3 Symbolic violence and the quiet losses The most uncomfortable implication concerns those who lose. In every organisation there are forms of knowledge that do not fit a feature vector. The nurse manager who can tell that a ward is about to become unsafe. The engineer who does not like the sound the machine is making. The regional director who knows the depot. These knowledges are real, they are hard-won, and they are, in the strict sense, expert. They are also unquantified, and in a field governed by #algorithmic_doxa, unquantified means unserious. The violence here is symbolic, in Bourdieu's precise sense: it is exercised with the complicity of those who suffer it. Nobody tells the nurse manager that her judgment does not count. She learns it from the shape of the meeting, and she comes to agree. She begins to say, apologetically, that she has nothing more than a feeling. She has been taught to devalue the very capacity that made her worth employing. Trittin-Ulbrich, Scherer, Munro, and Whelan (2021) call attention to the dark and unexpected sides of digitalisation. Bankins and Formosa (2023) examine what AI does to the meaningfulness of work. The Bourdieusian contribution is to locate the harm not in job loss or surveillance, which are visible, but in the redistribution of epistemic standing, which is not. 7.4 The machine as agent in the field Airoldi's (2022) argument that algorithms possess a #machine_habitus is worth taking further than he takes it. If a model is trained on the historical decisions of an organisation, then it has internalised the dispositions of that organisation's past leaders, including the dispositions that were the product of their positions and interests. When the model then recommends, it does not merely predict. It reproduces. It is a habitus made explicit, exported into code, and returned to the field with the authority of objectivity. This is why the language of bias, though not wrong, is inadequate. Bias implies a deviation from a neutral baseline. What models actually encode is not a deviation but a position: a particular way of cutting up the world, which was the way the dominant agents of a previous period cut it up. Amoore (2020) makes a related argument about how algorithms generate rather than apply ethics. Fourcade and Johns (2020) describe the recursive loop by which this becomes self-reinforcing. The strategic implication for leaders is severe. To adopt a model trained on your own history is to bind your organisation to the dispositions of its own past at exactly the moment you may most need to escape them. 7.5 What the framework does not claim Three disclaimers are necessary. First, the framework is not a claim that algorithmic decision support is bad. Models frequently outperform unaided human judgment, and the deference that this article describes is, in many domains, correct. The argument is not that leaders defer too much. It is that the deference is not produced by the reasons that leaders give for it, and therefore that it will not adjust correctly when those reasons cease to hold. Second, the framework does not deny agency. Bourdieu was consistently misread as a determinist, and he consistently protested. Habitus generates practice, but it does so in an open way, and reflexivity, especially the collective reflexivity generated by field crisis, can make dispositions available for revision. Third, the framework is not anti-quantitative. Bourdieu himself was an enthusiastic user of statistics. The critique is not of measurement but of the misrecognition of measurement as a view from nowhere. 8. Practical Implications The framework yields recommendations that differ sharply from the standard playbook, and the difference is instructive. Standard recommendations target beliefs and skills. These target positions and conditions. 8.1 Do not train the belief, restructure the field Literacy programmes teach executives what a model is. They rarely change practice, because practice is not generated by knowledge. If the field rewards the citation of models and penalises the citation of intuition, executives will cite models regardless of what they have learned in a workshop. What changes practice is a change in the conditions of recognition. If a board wants leaders to exercise judgment, it must make the exercise of judgment a source of #symbolic_capital rather than a liability, and it must do so in the room, visibly, and repeatedly. This is a matter of what the chair praises, not of what the policy states. 8.2 Make agreement as costly to record as disagreement Given the asymmetry of Proposition 8, the design implication is direct. If overrides must be documented, then acceptances must be documented too. A leader who follows a model's recommendation should be required to state, briefly, why the recommendation is appropriate to this case, and that statement should enter the same record. This equalises the reputational cost of the two choices and restores the possibility of a genuine decision. The objection will be that this creates friction. The objection is correct. That is the purpose. Frictionless acceptance is not #augmented_judgment. It is automation with a human signature on it. 8.3 Preserve heterogeneity of habitus If the mechanism of capital conversion sorts the senior team toward homogeneity, then the countermeasure is deliberate heterogeneity of dispositions at the top. This means valuing, protecting, and promoting leaders whose practical sense was formed in uninstrumented conditions, not as a concession to tradition but as a hedge against the model's distributional assumptions. An organisation whose entire senior team has an identical algorithmic habitus has, in effect, a single point of perceptual failure. 8.4 Audit the target variable, not the algorithm Most #ethics_of_AI audits inspect model performance. The Bourdieusian analysis suggests the decisive choices were made earlier, when someone decided what the model would predict, on what population, against what definition of success. Lebovitz and colleagues (2021) show that the #ground_truth is itself a contested professional artefact. The audit that matters asks who chose the label, whose interests it served, and what alternative labels were rejected. 8.5 Rethink #leadership_development Development programmes should aim at #reflexivity: the capacity to see that one's own sense of what is obvious is a product of one's position. This is teachable, but not by lecture. It is taught by exposure to fields with different doxa, by structured encounters with those whose knowledge the organisation devalues, and by the deliberate practice of deciding without the dashboard, in low-stakes settings, so that the capacity does not atrophy before the crisis in which it is needed. Parker and Grote (2022) argue that work design must be treated as a first-order variable in the age of algorithms rather than as a consequence of technical choices. The same holds at the executive level. The design of the decision environment is the design of the future leader. 9. Limitations and Future Research 9.1 Limitations The article is conceptual, and its propositions are untested. The constructs, particularly #computational_capital and #algorithmic_doxa, require operationalisation before they can be examined empirically, and operationalising a doxa is intrinsically difficult, since its defining feature is that participants do not remark on it. The framework is also developed largely from research in Western, English-language, and often North American settings. Bourdieu's own work insisted that fields are historically and nationally specific. The dispositions that constitute an algorithmic habitus in a Silicon Valley firm are unlikely to be identical to those in a state-owned enterprise or a family conglomerate, and the article does not attempt cross-national theorisation. Finally, the treatment of machines as position-holders in a field, while defensible as a structural claim, will not satisfy readers who hold stronger or weaker views about machine agency. The article takes an agnostic position and is content to be judged on whether that position is useful. 9.2 A research agenda Five directions follow. Ethnography of the executive suite. The most valuable next step is close observation of senior decision-making forums: what is said, what is not said, what is apologised for. Christin (2020) and Seaver (2022) demonstrate what ethnographic attention yields in adjacent settings. The boardroom remains largely closed to it, which is itself a finding about power. Correspondence analysis of hybrid fields. Bourdieu's preferred technique was multiple correspondence analysis, which maps positions in a field according to the volume and composition of capital. Applying it to senior teams, using measures of computational and other capitals, would allow the field structure to be represented empirically rather than asserted. Longitudinal study of dispositional drift. Because habitus forms slowly, the mechanisms proposed here can only be observed over years. Panel research tracking the perceptual and evaluative habits of cohorts of managers as their organisations instrument themselves would provide the strongest test. The override as a unit of analysis. Proposition 7 predicts systematic under-override. This is testable wherever override logs exist, by comparing override rates against the rate that private information would justify, and by examining whether the asymmetry increases with visibility. Comparative field studies. Proposition 4 predicts that the conversion rate of computational capital varies by field doxa. Comparing sectors with quantitative doxa, such as insurance, against those with relational or aesthetic doxa, such as design or care, would test this directly. 10. Conclusion The question that organisations keep asking is whether leaders should trust algorithms. This article has argued that the question is badly posed, because it locates in the individual something that lives in the relation between a position and a field. Executives do not confront machines as free minds weighing evidence. They confront them as agents with a history, occupying a place in a structured space, holding some resources and lacking others, and equipped with dispositions that were formed by exactly the conditions those resources produced. The feeling of obviousness that attends a leader's acceptance of a model, and the feeling of embarrassment that attends her intuition, are not psychological facts. They are social facts, and they are the trace of a structure at work in the body. Bourdieu's contribution to this problem is not a set of variables to be added to an existing model. It is a demand that we stop treating the natural as natural. The #algorithmic_habitus feels like clarity. It presents itself as rigour, as objectivity, as the simple preference of a serious person for evidence over anecdote. And it may often be all of those things. But it is also, and always, a position, and positions have interests, and interests have losers. The organisations that will handle the coming decades well are unlikely to be those that trust their models most, or least. They will be the ones that retain, at the top, enough people who can still feel the difference between a world their instruments describe and a world their instruments do not, and enough courage in the room to say so before the numbers agree. Hashtags #algorithmic_habitus #Bourdieu_in_management #hybrid_teams #executive_decision_making #computational_capital #algorithmic_doxa #symbolic_power_in_organisations #field_theory #machine_habitus #artificial_intelligence_governance #future_of_work #critical_algorithm_studies #organisational_sociology #augmented_judgment #leadership_theory References Airoldi, M. (2022). Machine habitus: Toward a sociology of algorithms. Cambridge: Polity Press. Airoldi, M., and Rokka, J. (2022). Algorithmic consumer culture. Consumption Markets and Culture, 25(5), 411-428. Amoore, L. (2020). Cloud ethics: Algorithms and the attributes of ourselves and others. Durham, NC: Duke University Press. Anthony, C. (2021). When knowledge work and analytical technologies collide: The practices and consequences of black boxing algorithmic technologies. Administrative Science Quarterly, 66(4), 1173-1212. Anthony, C., Bechky, B. A., and Fayard, A.-L. (2023). Collaborating with AI: Taking a system view to explore the future of work. Organization Science, 34(5), 1672-1694. Balasubramanian, N., Ye, Y., and Xu, M. (2022). Substituting human decision-making with machine learning: Implications for organizational learning. Academy of Management Review, 47(3), 448-465. Bailey, D. E., Faraj, S., Hinds, P. J., Leonardi, P. M., and von Krogh, G. (2022). We are all theorists of technology now: A relational perspective on emerging technology and organizing. Information Systems Research, 33(1), 1-18. Bankins, S., and Formosa, P. (2023). The ethical implications of artificial intelligence for meaningful work. Journal of Business Ethics, 185(4), 725-740. Bankins, S., Ocampo, A. C., Marrone, M., Restubog, S. L. D., and Woo, S. E. (2024). A multilevel review of artificial intelligence in organizations: Implications for organizational behavior research and practice. Journal of Organizational Behavior, 45(2), 159-182. Benbya, H., Pachidi, S., and Jarvenpaa, S. L. (2021). Special issue editorial: Artificial intelligence in organizations: Implications for information systems research. Journal of the Association for Information Systems, 22(2), 281-303. Berente, N., Gu, B., Recker, J., and Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433-1450. Bourdieu, P. (1977). Outline of a theory of practice. Cambridge: Cambridge University Press. Bourdieu, P. (1984). Distinction: A social critique of the judgement of taste. London: Routledge. Bourdieu, P. (1990). The logic of practice. Stanford, CA: Stanford University Press. Bourdieu, P., and Wacquant, L. J. D. (1992). An invitation to reflexive sociology. Chicago: University of Chicago Press. Burrell, J., and Fourcade, M. (2021). The society of algorithms. Annual Review of Sociology, 47, 213-237. Christin, A. (2020). Metrics at work: Journalism and the contested meaning of algorithms. Princeton, NJ: Princeton University Press. Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. New Haven, CT: Yale University Press. Csaszar, F. A., and Steinberger, T. (2022). Organizations as artificial intelligences: The use of artificial intelligence analogies in organization theory. Academy of Management Annals, 16(1), 1-37. Curchod, C., Patriotta, G., Cohen, L., and Neysen, N. (2020). Working for an algorithm: Power asymmetries and agency in online work settings. Administrative Science Quarterly, 65(3), 644-676. Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. Engaging Science, Technology, and Society, 5, 40-60. Emirbayer, M., and Johnson, V. (2008). Bourdieu and organizational analysis. Theory and Society, 37(1), 1-44. Faraj, S., and Pachidi, S. (2021). Beyond Uberization: The co-constitution of technology and organizing. Organization Theory, 2(1), 1-14. Fourcade, M., and Johns, F. (2020). Loops, ladders and links: The recursivity of social and machine learning. Theory and Society, 49(5-6), 803-832. Glikson, E., and Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 14(2), 627-660. Grønsund, T., and Aanestad, M. (2020). Augmenting the algorithm: Emerging human-in-the-loop work configurations. Journal of Strategic Information Systems, 29(2), 101614. Hallett, T., and Hawbaker, A. (2021). The case for an inhabited institutionalism in organizational research: Interaction, coupling, and change reconsidered. Theory and Society, 50(1), 1-32. Ignatow, G., and Robinson, L. (2017). Pierre Bourdieu: Theorizing the digital. Information, Communication and Society, 20(7), 950-966. Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410. Lebovitz, S., Levina, N., and Lifshitz-Assaf, H. (2021). Is AI ground truth really true? The dangers of training and evaluating AI tools based on experts know-how. MIS Quarterly, 45(3), 1501-1526. Lebovitz, S., Lifshitz-Assaf, H., and Levina, N. (2022). To engage or not to engage with AI for critical judgments: How professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), 126-148. Leonardi, P. M., and Treem, J. W. (2020). Behavioral visibility: A new paradigm for organization studies in the age of digitization, digitalization and datafication. Organization Studies, 41(12), 1601-1625. Logg, J. M., Minson, J. A., and Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90-103. Mahmud, H., Islam, A. K. M. N., Ahmed, S. I., and Smolander, K. (2022). What influences algorithmic decision-making? A systematic literature review on algorithm aversion. Technological Forecasting and Social Change, 175, 121390. Murray, A., Rhymer, J., and Sirmon, D. G. (2021). Humans and technology: Forms of conjoined agency in organizations. Academy of Management Review, 46(3), 552-571. Pachidi, S., Berends, H., Faraj, S., and Huysman, M. (2021). Make way for the algorithms: Symbolic actions and change in a regime of knowing. Organization Science, 32(1), 18-41. Parker, S. K., and Grote, G. (2022). Automation, algorithms, and beyond: Why work design matters more than ever in a digital world. Applied Psychology: An International Review, 71(4), 1171-1204. Puranam, P. (2021). Human-AI collaborative decision-making as an organization design problem. Journal of Organization Design, 10(2), 75-80. Rahman, H. A. (2021). The invisible cage: Workers reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988. Raisch, S., and Krakowski, S. (2021). Artificial intelligence and management: The automation-augmentation paradox. Academy of Management Review, 46(1), 192-210. Romele, A. (2021). Technological capital: Bourdieu, postphenomenology, and the philosophy of technology beyond the empirical turn. Philosophy and Technology, 34(3), 483-505. Seaver, N. (2022). Computing taste: Algorithms and the makers of music recommendation. Chicago: University of Chicago Press. Tong, S., Jia, N., Luo, X., and Fang, Z. (2021). The Janus face of artificial intelligence feedback: Deployment versus disclosure effects on employee performance. Strategic Management Journal, 42(9), 1600-1631. Trittin-Ulbrich, H., Scherer, A. G., Munro, I., and Whelan, G. (2021). Exploring the dark and unexpected sides of digitalization: Toward a critical agenda. Organization, 28(1), 8-25. Vaast, E., and Pinsonneault, A. (2021). When digital technologies enable and threaten occupational identity: The delicate balancing act of data scientists. MIS Quarterly, 45(3), 1087-1112. von Quaquebeke, N., and Gerpott, F. H. (2023). The now, new, and next of digital leadership: How artificial intelligence will take over and change leadership as we know it. Journal of Leadership and Organizational Studies, 30(3), 265-275. Waardenburg, L., Huysman, M., and Sergeeva, A. V. (2022). In the land of the blind, the one-eyed man is king: Knowledge brokerage in the age of learning algorithms. Organization Science, 33(1), 59-82.

Transnational education has grown from a small side activity of a few universities into a permanent feature of the world higher education map. Branch campuses, joint institutes, franchised degrees, validated programmes and online cross border delivery now operate in more than eighty host countries. Wherever these arrangements appear, they arrive with a set of external rules attached: the quality codes of the exporting country, the licensing conditions of the host country, and the standards of international or professional accreditation bodies. This article uses the concept of #coercive_isomorphism, taken from neo institutional theory, to explain why transnational providers across very different societies end up looking so similar to one another, and why locally rooted knowledge, language and pedagogy so often lose ground in the process. Drawing on an integrative review of recent scholarship published between 2021 and 2026, the article maps four channels through which coercive pressure travels: legal licensing, exporting country quality assurance, professional and disciplinary accreditation, and funding or ranking dependence. It then examines the counter pressures that push toward #localisation, including host state cultural policy, student expectations, staff agency and the growing decolonial critique of the curriculum. The analysis argues that the tension between #standardisation and #cultural_preservation is not a problem to be solved once, but a permanent condition to be governed. It proposes a working model, described here as negotiated compliance, in which institutions treat accreditation standards as a floor rather than a ceiling, distinguish between technical requirements and cultural assumptions hidden inside those requirements, and build local evidence into the audit trail rather than around it. The article ends with practical guidance for institutions, accreditation agencies and national regulators, and with an agenda for future research that gives greater weight to host country voices. Keywords: transnational education; coercive isomorphism; accreditation; quality assurance; cultural preservation; glocalisation; institutional theory; higher education governance 1. Introduction A student sitting in a lecture theatre in Kuala Lumpur, Dubai, Ningbo or Tashkent may be studying a degree that was designed in Manchester, Melbourne or Boston. The reading list, the assessment rubric, the module descriptor, the #learning_outcomes and the external examiner may all have been produced somewhere else. The lecturer may hold a passport from a fourth country. The building sits on local soil, the students grew up with local languages and local histories, and the graduates will mostly work in local labour markets. Yet the academic architecture around them was largely imported. This everyday scene is the subject of the present article. It is the visible face of what scholars call #transnational_education, a broad term covering any arrangement in which learners are located in a country different from the one where the awarding institution is based. The category includes international #branch_campuses, franchised and validated programmes, joint and dual degrees, articulation arrangements and cross border online provision. Recent counts place the number of international #branch_campuses at more than three hundred and thirty across more than eighty host countries, with Asia and the Middle East hosting the largest share (Xu, 2025). Research output on the topic has grown quickly since the 2010s, with the United Kingdom, Australia and the United States leading as producers of scholarship, and with China, Singapore, Malaysia, Hong Kong and Vietnam rising as major sites of activity (Nguyen et al., 2025). What interests this article is not the growth itself but the shape of it. Transnational programmes in very different societies tend to look remarkably alike. They use similar module structures, similar credit systems, similar learning outcome statements, similar student satisfaction surveys, similar programme review cycles and similar committee architectures. A quality manual written for a campus in one region can often be lifted, with light editing, into a campus in another. This #convergence is not an accident. It is produced by rules, and those rules carry consequences. Neo institutional theory offers a vocabulary for this. DiMaggio and Powell (1983) argued that organisations in the same field come to resemble one another through three mechanisms: coercive, mimetic and normative. Coercive pressure comes from formal and informal demands made by organisations on which an institution depends, and from the cultural expectations of the wider society. Mimetic pressure arises when organisations copy others under conditions of uncertainty. Normative pressure flows from professional training and professional networks. Meyer and Rowan (1977) had earlier proposed that organisations often adopt formal structures to gain ##legitimacy rather than to become more efficient, and that a gap can open between the structures they display and the work they actually do. That gap is known as #decoupling. These two papers are old, but they remain the reference points for the current literature, and recent reviews continue to organise the field around them (Guil Gorostidi and Rubio Arostegui, 2026). Coercive pressure is the sharpest of the three mechanisms in transnational education, because dependence is unusually direct. An international branch campus normally cannot open without a licence from the host state. It cannot award a degree without the approval of the home system #regulator. Its graduates cannot enter regulated professions without recognition from a professional body. Its business model may depend on a partner university that itself faces audit at home. As Coutet (2022) observed in the school sector, the more an organisation depends on another for survival, the more it will come to resemble that other, and ##accreditation_bodies hold one of the strongest forms of leverage available because they can withdraw the very #legitimacy on which enrolment rests. The tension this creates is not abstract. Host countries rarely invite foreign providers in simply to import a foreign society. They invite them to build capacity, to raise domestic quality, to keep talented students at home and to serve national development goals. Those goals usually include a cultural dimension: the maintenance of national languages, religious values, legal traditions, professional customs and historical narratives. The same host state that demands #compliance with local content requirements may also demand that the foreign provider maintain full #equivalence with its home campus. The provider is then asked to be two things at once. It must be recognisably the same as its parent institution, and it must be recognisably at home in its host society. This article asks three questions. First, through what specific channels does coercive pressure travel in transnational education, and what does each channel actually demand? Second, what happens to local cultural content, local languages and local ways of knowing when those demands are met? Third, what strategies allow institutions to satisfy accreditation requirements without hollowing out local relevance? The contribution is conceptual rather than empirical. The article does not report new fieldwork. It builds an integrative account from recent literature and offers an analytical framework that students, researchers and practitioners can apply to their own settings. Its central claim is that #standardisation and localisation are not opposites on a single line, where more of one means less of the other. They operate on different dimensions, and the practical task is to identify where a standard genuinely requires a specific practice and where it merely appears to, because a cultural assumption has been folded quietly into a technical rule. The article proceeds as follows. Section 2 reviews the recent literature on transnational education, #quality_assurance and institutional theory. Section 3 sets out the theoretical framework. Section 4 explains the review method. Section 5 presents the analysis, organised around the channels of coercive pressure and the counter pressures that resist them. Section 6 discusses what the analysis means for theory. Section 7 offers practical implications. Sections 8 and 9 address limitations and future research, and Section 10 concludes. 2. Literature Review 2.1 The growth and shape of transnational education Scholarship on transnational education has matured over the last decade. Bibliometric work by Nguyen et al. (2025), covering more than sixteen hundred Scopus indexed publications from 1972 to 2023, shows steady growth in the field with a sharp acceleration after 2010. Their analysis also reveals a split in the field's attention. Work on higher education tends to cluster around economic, institutional and regulatory themes, with #quality_assurance, capacity building, education hubs, #branch_campuses and partnerships among the most frequent keywords. Work on school level provision clusters instead around #identity, language acquisition and socio #cultural_adaptation. This divide matters for the present article, because it suggests that the cultural questions have been treated as somebody else's problem within higher education research, even though they arise there with equal force. Definitional work continues. Wilkins and Huisman (2024) have argued for a tighter definition of transnational education, noting that the category has been stretched to cover arrangements with little in common. Precision here is not pedantry. Regulatory obligations differ sharply between a wholly owned branch campus, a franchised programme delivered by a private local college and a validated award where the local partner designs the teaching. Each configuration produces a different distribution of coercive pressure, and each leaves a different amount of room for local adaptation. The geography of provision is also shifting. Wilkins and Muratov (2024) report that Uzbekistan has become one of the largest hosts of international branch campuses in the world, a development that would have surprised most observers a decade ago. Malaysia continues to combine an #education_hub strategy with strong national cultural policy (Xu, 2025). China's system, built on a regulatory requirement that partnerships involve both a local and a foreign institution, produces an unusually structured form of cooperation (Yang, 2023; Mok, 2021). These are not variations on a single Western template. They are distinct institutional environments with distinct coercive profiles. 2.2 Quality assurance across borders The literature on cross border #quality_assurance has expanded in parallel. Carvalho, Rosa and Amaral (2023), in a systematic review, mapped the ways in which #quality_assurance and cross border provision have been linked in scholarship. They identified recurring themes: the role of national legislation in both promoting and regulating foreign provision, the responsibilities of exporting and importing countries, the need for mutual recognition, and the persistent difficulty of building a genuinely transnational quality system. One of their most relevant observations for this article is that international #accreditation, based on transnational guidelines and international agencies, tends to use broad standards designed to cover large and diverse populations of institutions, and therefore does not easily accommodate national specificities. Hou and colleagues have documented how quality assurance agencies in Asia are responding to a changing landscape. Their study of the pandemic period found that universities in the region were pushed into hybrid and virtual delivery of joint and dual degree programmes, and that quality assurance and qualification recognition frameworks struggled to keep pace with modes of delivery they were never designed to cover (Hou et al., 2024a). A companion study examined how Asian agencies are responding to the revised international guidelines issued by the International Network for Quality Assurance Agencies in Higher Education in 2022, which added new modules in response to changes in the sector (Hou et al., 2024b). Both studies show a field in which the standard setters sit at some distance from the classrooms being standardised. 2.3 Isomorphism in higher education Institutional theory has been applied to higher education for many years, and the recent literature is substantial. The most useful recent synthesis is the scoping review by Guil Gorostidi and Rubio Arostegui (2026), which analysed one hundred and thirteen studies on quality management in higher education through a neo institutional lens. Their findings are worth stating carefully, because they anticipate much of what this article develops. First, they found that quality systems act as vectors of #isomorphism in all three of DiMaggio and Powell's forms: coercive through regulatory requirements, mimetic through imitation of leading institutions, and normative through the professionalisation of management teams. Second, they found that these pressures are especially visible in young or peripheral institutions, which tend to replicate dominant models in order to gain legitimacy. Third, they identified a set of recurring challenges: regulatory homogenisation that reduces institutional diversity, over reliance on quantitative indicators that produces a reductionist view of quality, bureaucratisation that creates administrative burden, institutional resistance that produces #symbolic_compliance rather than change, and a functionalist focus on #employability that strains the critical and humanistic purposes of the university. Fourth, and importantly, they found that institutions are not passive. They negotiate, adapt, translate and sometimes resist. Fifth, they observed that the literature itself is skewed toward Europe and North America, and that including the perspective of the Global South is necessary to avoid a false picture of uniform institutional change. Specific studies fill in the picture. Lu and Wang (2023) examined how sudden coercive pressure, in their case the abrupt demand for digital transformation in colleges, is experienced by teachers rather than by managers, and found that the pressure reshapes practice in ways that policy documents do not anticipate. Núñez Valdés, Quirós Alpera and Cerdá Suárez (2021) traced how digital transformation becomes institutionalised through professional networks and shared standards, a clear case of normative pressure operating alongside coercive rules. The most direct precedent for the present argument comes from the school sector. Coutet (2022) analysed the handbook, standards and lesson observation instrument of one large international school accreditor and reconstructed the picture of the ideal school implied by those documents. The conclusion was blunt. Accreditation restricted how the school could be managed, narrowed its learning philosophy and prescribed what happened in classrooms. Space for creativity survived only at the level of individual lesson design, and even there the activities had to align with prescribed standards. If this is what accreditation does to a school, it is reasonable to ask what it does to a branch campus. 2.4 Culture, glocalisation and the local response A parallel literature has grown around what happens to culture in these settings. Miani (2026) uses the concept of #glocalisation to examine a foundation year delivered through a transnational partnership in China. The study shows that the programme is not a copy of its United Kingdom equivalent. Tensions appear around English as the medium of instruction, around attendance monitoring, around the reception of equity and inclusion ideals, and around policies on artificial intelligence. The conclusion is that individual agency and creativity mediate these tensions, and that the local arrangements which emerge are not failures of fidelity but original solutions to a real problem. Xu (2025) approaches the same question through the idea of cultural discount, borrowed from media economics, and argues that educational programmes lose value when they cross cultural boundaries unless they are adapted. Studying branch campuses in Malaysia, a country that combines an education hub strategy with firm #cultural_preservation policies, the study finds that strategic adaptation is what allows campuses to survive, not what threatens their survival. Zhan and Marginson (2024) describe the problem of institutional dual identity in branch campuses, using research capacity building at a campus in China as their case. The campus must be legible to its parent institution and to its host system at the same time, and those two audiences want different things. Wilkins and Annabi (2023), surveying academic staff at offshore campuses across ten countries, add the human dimension. Staff report autonomy, motivated students and smaller classes on the positive side, but also insecurity, limited research support and constraints on #academic_freedom. These conditions shape how much local adaptation staff are willing or able to attempt. 2.5 The critical frame A final strand of literature reframes the whole discussion. Marginson (2022) argues that the dominant model of institution and system in global higher education policy is the Anglo American research university operating in a competitive market, and that this model exercises varying but real influence across the world. Marginson and Xu (2023) go further, criticising the centre and periphery model itself while acknowledging the hierarchies that make it plausible. De Wit and Altbach (2021) observe that #internationalisation has moved from a marginal activity to a central item on the reform agenda, and warn that it has become entangled with competition, reputation and revenue in ways that distort its stated purposes. Thondhlana et al. (2021) assemble perspectives from the Global South that push against the assumption that #internationalisation means adopting Northern forms. Taken together, this literature establishes three points. Coercive pressure in transnational education is real, well documented and strongest where dependence is greatest. Cultural adaptation is not a nice addition but a condition of survival and relevance. And institutions retain agency, which means the outcome is not determined in advance. 3. Theoretical Framework 3.1 The three mechanisms, briefly restated DiMaggio and Powell (1983) distinguished three routes to organisational similarity. #Coercive_isomorphism results from pressure exerted by organisations on which an institution depends and from cultural expectations in the society in which it operates. The pressure may be experienced as force, as persuasion or as an invitation to join a club. #Mimetic_isomorphism results from uncertainty. When goals are ambiguous and the environment is hard to read, organisations copy peers they regard as successful. #Normative_isomorphism results from professionalisation. As staff are trained in the same programmes, join the same associations and read the same handbooks, they carry shared assumptions about what a proper university looks like. The three mechanisms are analytically separate but practically entangled. A rule may begin as a professional norm, harden into an accreditation standard, and then be written into national law. By the time it reaches a branch campus in a #host_country, it arrives as a legal obligation with a professional history that nobody in the room can see. 3.2 Legitimacy and decoupling Meyer and Rowan (1977) supply the second element. Organisations adopt structures that signal conformity to the expectations of their environment, and these structures may have little to do with the technical core of the work. A university may build an elaborate #quality_management system, complete with committees, indicators and annual reports, and yet the teaching may continue much as before. The recent review by Guil Gorostidi and Rubio Arostegui (2026) confirms that decoupling remains a live phenomenon, and that #symbolic_compliance is one of the most common institutional responses to external pressure. Decoupling matters for the cultural argument in two opposite ways. It can be a survival strategy that protects local practice: the institution shows the auditors what they want to see and goes on teaching in the way that makes sense locally. It can also be a trap: energy is consumed by paperwork that changes nothing, staff become cynical, and the genuine improvements that accreditation might have delivered never arrive. 3.3 The framework proposed here This article builds on those foundations with a framework of four coercive channels and three counter pressures. The four channels are: Channel one, host state licensing. The legal permission to operate. Conditions may include ownership rules, land use, staff visas, fee regulation, compulsory subjects, language of instruction and content restrictions. Channel two, home system quality assurance. The exporting country's #regulator or the parent institution's own internal processes, which usually require that the offshore award be equivalent to the onshore award. Channel three, professional and disciplinary accreditation. Engineering, medicine, nursing, law, accounting, architecture, psychology and business #accreditation_bodies, whose standards travel across borders and which control access to regulated professions. Channel four, funding and reputational dependence. Rankings, research assessment, donor conditions, employer expectations and student recruitment markets. This channel is formally voluntary and practically not. The three counter pressures are: Counter pressure one, host society expectation. National language policy, religious and civic values, #curriculum mandates, local labour market needs and public opinion. Counter pressure two, professional and staff agency. The judgement of teachers, programme leaders and local managers, who translate rules into practice and who decide, in the space between the rule and the classroom, what actually happens (Miani, 2026). Counter pressure three, epistemic critique. The argument, advanced most forcefully in #decolonial_scholarship, that imported curricula carry a hidden hierarchy of knowledge, and that local and indigenous knowledge systems deserve standing in their own right. The interaction of the four channels and the three counter pressures produces the range of outcomes observed in practice: full replication, cosmetic localisation, #negotiated_compliance, deep #hybridisation and, occasionally, open conflict and withdrawal. 3.4 Why the standard often looks neutral One further theoretical point is needed. Accreditation standards are usually written in the language of technical requirement. A standard may say that students must receive timely feedback, that programmes must have measurable #learning_outcomes, that staff must hold terminal degrees, that class sizes must permit interaction, or that assessment must be moderated by a second marker. Nothing in these sentences mentions culture. They read as universal statements about good practice. Yet each contains assumptions. Timely feedback assumes a particular rhythm of study. Measurable learning outcomes assume a behaviourist tradition of #curriculum design that is not universal. Terminal degrees privilege one route to expertise and exclude others, including craft, professional and community based knowledge. Interaction norms assume a classroom relationship between teacher and student that varies enormously across societies. Second marking assumes distrust of the individual examiner, which is itself a cultural stance. The framework therefore requires a distinction between what may be called the hard core of a standard, which protects students from real harm and can be defended anywhere, and the soft shell, which encodes the professional habits of the system that wrote it. Confusing the two is the central error in transnational quality practice. Institutions defend the shell as if it were the core, and abandon local practice that the core never required them to abandon. 4. Methodology 4.1 Design This is a conceptual and integrative review. It does not report primary data. The purpose of an integrative review is to synthesise a body of scholarship in order to generate a framework rather than to measure an effect. That purpose fits the present question, which is analytical rather than statistical. 4.2 Sources and selection Literature was identified through searches of indexed journal databases and publisher platforms using combinations of terms relating to transnational education, cross border higher education, international branch campuses, quality assurance, accreditation, institutional isomorphism, neo institutionalism, glocalisation and #cultural_adaptation. Priority was given to peer reviewed journal articles and scholarly books published between 2021 and 2026, on the grounds that the regulatory and market environment of transnational education changed substantially during and after the pandemic period. Two older works, Meyer and Rowan (1977) and DiMaggio and Powell (1983), are retained because they are the theoretical foundation of the entire field and because every recent study reviewed here builds directly on them. Studies were included when they addressed at least one of the following: the operation of transnational or cross border higher education; the effect of quality assurance or accreditation on institutional behaviour; the application of institutional theory to education; or the cultural, linguistic and epistemic consequences of imported curricula. Studies were excluded when they dealt only with student mobility, only with domestic higher education policy in a single country with no cross border dimension, or when they were purely technical treatments of assessment methods. 4.3 Analysis The selected literature was read and coded against the three mechanisms of isomorphism, then re coded against the sites at which pressure is applied, which produced the four channel structure used in Section 5. Counter pressures were identified inductively from studies that reported adaptation, resistance, negotiation or agency. The resulting framework was then tested against the case material described in the reviewed studies to check that it could account for the outcomes those studies reported. 4.4 Limits of the design An integrative review inherits the biases of the literature it reviews. As Guil Gorostidi and Rubio Arostegui (2026) note, scholarship on isomorphism and quality is heavily weighted toward Europe and North America, and English language search terms compound that weighting. The framework offered here should therefore be treated as a hypothesis to be tested in #host_country settings, not as a finding established across them. Section 8 returns to this point. 5. Analysis 5.1 Channel one: the host state licence The first and most obvious source of coercive pressure is the host government. No foreign provider operates without permission, and permission comes with conditions. The conditions vary. Some are structural, such as requirements on ownership, capital, campus facilities and staff qualifications. Some are economic, such as fee caps or quotas on local student enrolment. Some are cultural in a direct sense. A host state may require that certain national subjects be taught to all students regardless of programme, that the national language be used in specified courses, that religious or civic education be included, or that certain content be excluded from the #curriculum. Yang (2023) shows how China's regulatory framework structures transnational provision by requiring cooperation between a domestic and a foreign institution, which places a local partner permanently inside the #governance of every programme. Mok (2021) traces how this framework developed alongside the expansion of non state provision, and Mok and Marginson (2021) place it within the wider story of massification and internationalisation. This channel is interesting because it is the one channel where coercive pressure pushes toward the local rather than away from it. The host state is often the strongest advocate of #cultural_preservation in the system. Malaysia's combination of education hub ambition with firm national cultural policy is the clearest example, and Xu (2025) shows that campuses which adapt strategically to that policy environment perform better than those which attempt to transplant a programme unchanged. The pressure is nonetheless coercive. A compulsory national subject inserted into an imported degree is a legal requirement, not a pedagogical choice, and it may sit awkwardly in a curriculum whose credit structure was designed elsewhere. Programme teams then face a familiar problem. The imported degree is full. Something must be removed to make room. What is removed is usually an elective, and electives are where local content most often lives. The result is that a rule intended to protect local culture can, through the arithmetic of credit systems, squeeze out other local content. 5.2 Channel two: home system quality assurance and the #equivalence rule The second channel is the exporting system. The awarding institution must satisfy its own regulator that the degree taught offshore is equivalent to the degree taught at home. This requirement is entirely reasonable in its purpose. It exists to prevent a well known abuse, in which a reputable university lends its name to a low quality operation abroad and collects the fees. The difficulty lies in how #equivalence is interpreted. Strictly interpreted, it means sameness. Same modules, same assessments, same marking criteria, same standards of achievement. Loosely interpreted, it means comparability of outcome, which permits different routes to the same standard. The strict interpretation is administratively easier, easier to audit and easier to defend. It is therefore the interpretation most institutions adopt. Strict equivalence is the single most powerful engine of #homogenisation in transnational education. It converts every proposed local adaptation into a risk. A programme leader who wants to replace a case study drawn from the home economy with one drawn from the host economy must now argue that the change does not threaten equivalence. The burden of proof sits on the person proposing the local adaptation, never on the person proposing the imported default. Over a few audit cycles, the incentive is obvious. Adaptation costs time, generates paperwork and attracts scrutiny. Replication costs nothing. This is coercive isomorphism working exactly as DiMaggio and Powell described it: an organisation that depends on another for legitimacy comes to resemble it, not because resemblance is better, but because resemblance is safe. Carvalho, Rosa and Amaral (2023) capture the wider version of this problem when they note that transnational standards are written broadly to cover many institutions and therefore do not take national specificities into account. 5.3 Channel three: professional and disciplinary accreditation The third channel is the most technically demanding and the least discussed in cultural terms. Professional accreditation bodies in engineering, medicine, nursing, law, accounting, psychology, architecture and business set standards that graduates must meet in order to enter regulated occupations. These standards travel across borders, sometimes through mutual recognition agreements and sometimes through the direct accreditation of foreign programmes. Professional accreditation carries unusually high stakes for students. A degree that is not professionally accredited may be worth very little in the labour market, whatever its academic merit. Students and their families know this. Enrolment therefore depends on it, and the institution's dependence on the accreditor is close to total. The cultural consequences run deep because professional standards are not merely technical. A nursing curriculum encodes assumptions about the relationship between patient and clinician, about family involvement in care, about consent, about end of life decisions and about the role of religion in health. A law curriculum encodes a legal tradition. An accounting curriculum encodes a reporting framework. A business curriculum encodes assumptions about the firm, the market and the employee that are not neutral across societies. When a professional standard written in one legal and cultural tradition is applied in another, it does not simply raise quality. It also transports a worldview. Coutet (2022) demonstrated the mechanism at school level with unusual clarity. By reading the accreditor's own documents, the study reconstructed the ideal school those documents describe, and found that accreditation narrowed the learning philosophy of the institution and prescribed classroom practice, leaving creativity only at the margins of lesson design. The same reading exercise applied to #professional_accreditation handbooks in higher education would, one suspects, produce similar results. This is a research gap and Section 9 returns to it. 5.4 Channel four: funding, #rankings and reputational dependence The fourth channel is formally voluntary. No law requires a university to pursue a ranking position. Yet the pressure is real, and the recent review literature treats it as one of the strongest isomorphic forces in the sector (Guil Gorostidi and Rubio Arostegui, 2026). Rankings reward research output in indexed journals, international staff and student ratios, and reputation surveys among academics who are themselves distributed unevenly across the world. For a branch campus, this channel produces a specific distortion. The campus is judged, internally and externally, by criteria set at the parent institution, which are in turn shaped by a global competition it did not design. Zhan and Marginson (2024) describe the resulting problem of dual identity in research capacity building. The campus is expected to produce research that counts in the parent's system, which usually means publication in international journals in English, while also being expected to serve the host society, which may need applied work in the local language on questions that no indexed journal will publish. Marginson (2022) situates this within a wider argument. The dominant model in global higher education policy is the Anglo American research university in a competitive market, and its influence spreads well beyond the countries where it originated. Marginson and Xu (2023) complicate the picture by criticising simple centre and periphery accounts, but the hierarchy they describe remains. De Wit and Altbach (2021) reach a similar conclusion from a different direction, warning that internationalisation has become tied to competition and revenue in ways that pull it away from its educational purposes. The effect on #local_knowledge is quiet but severe. A researcher at a branch campus who wants promotion learns quickly which kind of work counts. Work on local problems, in local languages, for local audiences, does not count. Over time the research culture of the campus detaches from the society that hosts it. 5.5 Counter pressure one: host society expectation Against these four channels stand three counter pressures. The first is the host society. Governments, employers, families and students all have expectations, and those expectations are not simply that the foreign degree be delivered unchanged. Students at branch campuses often choose them precisely because they want an international qualification without leaving home, but they also expect the institution to understand where home is. Xu (2025) frames this as cultural discount: a programme that ignores the host context loses value in the eyes of the very students it was designed to attract. Employers apply their own pressure. A graduate who has learned accounting under one reporting framework, or law under one legal tradition, or management under one set of workplace assumptions, may be less employable locally than a graduate from a domestic institution. This creates a commercial incentive for adaptation that runs directly against the equivalence rule described in Section 5.2. The institution is caught between two forms of dependence, and it must choose which one to disappoint. 5.6 Counter pressure two: staff and professional agency The second counter pressure is the professional judgement of the people who actually teach. Rules do not implement themselves. Between the module descriptor and the classroom there is always a person, and that person makes decisions. Miani (2026) provides the clearest recent evidence. In a foundation year delivered through a Sino foreign partnership, tensions appeared around English as the medium of instruction, around attendance monitoring, around equity and inclusion ideals and around policies on artificial intelligence. The programme diverged from its United Kingdom equivalent, not through failure but through the accumulated judgements of staff who were solving real problems in a real place. The study concludes that individual agency and creativity mediate these tensions, and that the local arrangements which result are original solutions rather than compromises. Yet agency has conditions. Wilkins and Annabi (2023), surveying academic staff at offshore campuses in ten countries, found that staff reported greater autonomy and more motivated students than they expected, but also job insecurity, weak research support and limits on #academic_freedom. Staff on short contracts, far from the centre of power in the parent institution, and uncertain of renewal, are not well placed to argue with an accreditor. The capacity to adapt is unevenly distributed, and it correlates with security. Lu and Wang (2023) make a related point from the teacher's perspective. When coercive pressure arrives suddenly, teachers absorb it in ways that the policy never anticipated, and the gap between the official change and the experienced change can be wide. The lesson for transnational education is that the person best placed to localise a curriculum is often the person least protected in the institution. 5.7 Counter pressure three: epistemic critique The third counter pressure is intellectual. A growing body of scholarship argues that imported curricula carry an implicit hierarchy of knowledge, in which theory produced in one part of the world is treated as universal and knowledge produced elsewhere is treated as local colour. Thondhlana et al. (2021) gather perspectives from the Global South that press this case within the internationalisation debate itself. Marginson and Xu (2023) examine how hierarchy operates in global science. This critique bears directly on accreditation. If a standard requires that a programme be informed by current research, and if current research is defined by indexed journals in one language, then the standard has quietly determined whose knowledge counts. If a standard requires that staff hold doctorates from recognised institutions, and recognition follows the same hierarchy, then the standard has determined who may teach. Neither requirement mentions culture, and both have cultural effects. The practical difficulty is that the epistemic critique is easier to state than to operationalise. An #accreditation_panel will not accept the argument that a standard is culturally loaded as a reason for non compliance. What it will accept is evidence that the institution has met the underlying purpose of the standard by a different route. That distinction, between the purpose of a standard and its default implementation, is where the epistemic critique becomes usable rather than merely correct. 5.8 The outcomes: a typology The interaction of channels and counter pressures produces a range of institutional responses. Five are visible in the literature. Full replication. The programme is delivered as designed at home, with no adaptation beyond examples and administrative arrangements. This is the safest response to the equivalence rule and the most common in #franchised_provision. It maximises audit safety and minimises local relevance. Cosmetic localisation. Local names, local examples and local imagery are added to a curriculum whose structure, assumptions and assessment remain unchanged. The institution can point to adaptation without having made any. This is #decoupling in its cultural form. Negotiated compliance. The institution distinguishes between the purposes of standards and their default implementations, meets the purposes by locally appropriate means, and documents the reasoning. This requires effort, confidence and a good relationship with the accreditor. It is the response this article recommends and Section 7 develops it. Deep #hybridisation. A genuinely new institutional form emerges, drawing on both traditions without being reducible to either. The Sino foreign cooperative university, with its mandatory dual #governance, is the clearest structural example (Yang, 2023; Miani, 2026). Hybridisation is not automatically good. It can produce incoherence as easily as innovation. But it is the only route by which transnational education produces something the world did not already have. Conflict and withdrawal. Where the demands of the host state and the demands of the home system cannot be reconciled, the partnership ends. This outcome is under studied, partly because failed ventures do not publish about themselves. 6. Discussion 6.1 Coercive pressure is not one thing The first implication of this analysis is that coercive pressure in transnational education is plural and sometimes self contradictory. The literature frequently treats coercion as a single force pushing institutions toward a global standard. The four channel model shows that it is not. Channel one, host state licensing, often pushes toward the local. Channels two, three and four push away from it. An institution is not being pulled in one direction with varying strength. It is being pulled in different directions at once. This reframing matters because it changes what counts as a good institutional strategy. If coercion were a single force, the only strategies available would be compliance, resistance or symbolic decoupling. Because coercion is plural, a fourth strategy becomes available: playing the channels against one another. An institution can use a host state content requirement as a justification, in an equivalence audit, for a curriculum change it wanted anyway. It can use #professional_accreditation requirements for local practice placements as leverage to secure local research funding. Skilled transnational managers do this routinely, and the literature has not paid it enough attention. 6.2 Standardisation and localisation are not a zero sum trade The second implication concerns the framing of the problem itself. The dominant framing, visible in the title of this article and in much of the policy discussion, treats standardisation and cultural preservation as ends of a single line. More standard means less local. Less standard means more local. The analysis suggests this framing is misleading. The hard core of a good accreditation standard, the part that protects students from being cheated, protects them equally in every society. Requiring that teaching staff be qualified, that assessment be fair, that students receive feedback, that complaints be heard and that awards mean what they claim to mean is not a cultural imposition. It is what a student anywhere is entitled to. Weakening those requirements in the name of cultural specificity would not preserve local culture. It would expose local students to poor provision, which is the outcome the whole quality assurance apparatus was built to prevent. What sits in tension with local culture is not the core but the shell: the accumulated professional habits of the exporting system, encoded in standards as though they were technical necessities. Fifteen week semesters, particular assessment formats, particular ratios, particular committee structures, particular styles of classroom interaction, particular definitions of research output. None of these is required by the underlying purpose. All of them are enforced as though they were. The practical task is therefore not to find a point of balance on a line between standard and local. It is to separate the core from the shell, standard by standard, and to defend the first while negotiating the second. This is slow, unglamorous work. It is also the only version of the task that can actually be done. 6.3 Decoupling is a symptom, not a strategy The third implication concerns decoupling. Guil Gorostidi and Rubio Arostegui (2026) confirm that symbolic compliance is one of the most common institutional responses to #isomorphic_pressure. In transnational education, this often takes the form of an elaborate quality apparatus that satisfies the auditors while the teaching continues along its own path. It is tempting to treat this as a clever solution. It is not. Decoupling burns institutional energy on documentation that changes nothing, teaches staff that quality processes are theatre, and leaves genuinely poor practice undetected because everybody has learned to stop looking. It also fails on its own terms, because the practices it protects are not usually locally grounded alternatives. They are simply the habits that happened to survive the last audit. The alternative is to make local adaptation visible and defensible rather than hidden. This is harder in the short run and safer in the long run, because a documented, reasoned adaptation survives a change of auditor and a hidden one does not. 6.4 Agency is real but unequally distributed The fourth implication concerns the people. Miani (2026) shows that staff agency shapes what transnational programmes actually become. Wilkins and Annabi (2023) show that many of those staff are insecure, under supported and constrained in their academic freedom. Lu and Wang (2023) show that teachers absorb coercive pressure in ways that policy documents do not capture. Put those findings together and a conclusion follows that the literature has not stated plainly enough. The capacity of a transnational institution to adapt its curriculum to its host society depends on the employment conditions of its academic staff. An institution staffed by short contract expatriates with no security and no research time cannot localise, whatever its strategy documents say, because localisation requires #local_knowledge, professional confidence and the willingness to argue with an auditor. None of those things grow in insecure soil. This is an uncomfortable finding for institutions that have built their offshore business model on flexible staffing, and it deserves direct empirical testing. 6.5 The theoretical contribution For institutional theory, this analysis offers a modest refinement. DiMaggio and Powell's three mechanisms were developed for organisational fields located within a single society. Transnational education places an organisation inside two or more fields at once, each with its own coercive authorities, its own professional networks and its own models worth imitating. The result is not stronger #isomorphic_pressure but competing isomorphic pressures, which is a different phenomenon. Zhan and Marginson (2024) capture one aspect of this with the idea of dual identity. The framework proposed here suggests that dual field membership is the structural condition, and dual identity is its psychological and organisational expression. Where the two fields make compatible demands, isomorphism proceeds as the classical theory predicts. Where they conflict, the organisation must choose, and the choice reveals which dependence it judges to be more dangerous to defy. That choice, rather than the resemblance itself, is where the interesting variation lies. 7. Practical Implications 7.1 For institutions Audit your standards for hidden culture. Take each accreditation standard the institution must meet and ask two questions. What harm is this standard preventing? Is the practice it prescribes the only way to prevent that harm? Where the answer to the second question is no, the institution has found negotiating room. Document it before the audit, not during it. Build the local evidence base first. An #accreditation_panel will not accept a cultural argument. It will accept evidence. If a locally designed assessment produces equivalent or better learning outcomes than the imported one, and the institution can show the data, the argument is over. Institutions that want to localise should invest in the evaluation capacity to prove that their local version works, because the burden of proof will always fall on them. Treat equivalence as comparability of outcome, and argue for it early. The strict reading of equivalence, meaning sameness of input, is a choice made by administrators who want an easy audit. It is worth contesting at the point where the partnership agreement is written, because once it is embedded in the contract it will govern every decision for a decade. Protect the people who do the localising. Staff on insecure contracts will not risk an argument with a validating partner. If localisation matters to the institution, it must appear in workload models, in promotion criteria and in contract length. Otherwise it will remain a paragraph in the strategic plan. Distinguish cosmetic from structural adaptation, honestly. Changing the names in a case study is not localisation. Ask whether the assumptions of the discipline, the structure of the assessment and the definition of good work have been examined. Usually they have not. 7.2 For accreditation and quality assurance agencies Write purposes, not practices. A standard that states the purpose to be achieved, and offers examples of acceptable practice rather than a single prescribed one, is more robust across contexts and no weaker in protecting students. Hou et al. (2024b) show that international guidelines are already being revised in response to a changing sector, which suggests the appetite for this exists. Train panels in context. A panel composed entirely of reviewers from the exporting system will read local variation as deficiency, because deficiency is the only category it has for difference. Mixed panels with genuine host country representation are the minimum condition for fair review, and they are still not standard practice. Recognise local scholarship. Where standards require currency of research, they should recognise research published in host country languages and outlets. Otherwise the standard functions as a language policy without saying so. Watch for the burden of proof. If every proposed local adaptation must be justified while every imported default is accepted without question, the system is not neutral, whatever its handbook says. 7.3 For national regulators in host countries Coordinate the channels. When the host licence and the exporting regulator make incompatible demands, the institution resolves the conflict privately, and it usually resolves it in favour of the party that can shut it down fastest. Regulators who want cultural content protected should negotiate directly with exporting regulators rather than leaving individual campuses to arbitrate. Be specific about what preservation means. A requirement to teach national subjects is easy to satisfy cosmetically. A requirement that graduates be able to practise competently within the national legal, professional and linguistic environment is harder to satisfy and much more valuable. The second is worth more than the first. Build domestic quality capacity. The literature on cross border quality assurance is consistent that importing countries which rely entirely on the exporting country's oversight lose control of the terms (Carvalho et al., 2023; Hou et al., 2024a). Domestic capacity is what makes negotiation possible. 7.4 For students Students are not passive in this system. They are the source of the revenue that makes transnational provision viable, and their choices carry weight. A student who asks, before enrolling, whether the programme is professionally recognised in the country where they intend to work, whether local content is taught by staff who know the local context, and whether the qualification is recognised by the host state as well as the home state, is applying exactly the pressure the system responds to. 8. Limitations Three limitations should be stated plainly. First, this is a conceptual review and not an empirical study. The four channel framework is derived from published scholarship and has not been tested against primary data. It should be read as a proposal for investigation rather than a finding. Second, the literature on which it draws is skewed. As Guil Gorostidi and Rubio Arostegui (2026) note directly of their own review, scholarship on isomorphism and quality management is dominated by European and Anglophone sources, and English language searching reinforces that dominance. Work published in Arabic, Chinese, Malay, Vietnamese and other languages of major host countries is under represented here. This is a serious weakness in an article about cultural preservation, and the author does not wish to minimise it. Third, the framework treats the host society as more unified than it is. Host countries contain competing constituencies with different views about what should be preserved and what should be imported. National elites who send their own children abroad may want a foreign curriculum delivered without adaptation. Cultural and religious institutions may want strong content requirements. Employers may want practical skills and care little about either. Speaking of the host society as a single actor, as this article has sometimes done for the sake of clarity, flattens a real politics. 9. Directions for Future Research Five lines of enquiry follow from this analysis. Document analysis of #professional_accreditation standards. Coutet (2022) read one school accreditor's handbook, standards and observation instrument and reconstructed the ideal school they imply. The same method applied to the standards of engineering, nursing, accounting and business accreditation bodies operating transnationally would show precisely which cultural assumptions are embedded in which requirements. This is achievable work and nobody appears to have done it. The employment conditions hypothesis. Section 6.4 argued that an institution's capacity to localise depends on the security and status of its academic staff. This is testable. A study comparing degree of curricular localisation against staff contract type, tenure and research support across a sample of branch campuses would either support or refute it. Failed partnerships. The literature studies survivors. Partnerships that collapsed because the demands of host and home systems could not be reconciled would tell us more about the limits of #negotiated_compliance than any successful case can. Student and graduate outcomes. Does cultural adaptation improve learning, employability and civic contribution, or does it only feel better? Xu (2025) argues from cultural discount theory that adaptation sustains the institution. Whether it also benefits the student is a separate question that deserves direct measurement. Host country scholarship. The most valuable single contribution to this field would be sustained research led by scholars in host countries, published in host country outlets as well as international ones, describing how these pressures are experienced from inside. Thondhlana et al. (2021) point in this direction. The field needs much more of it. 10. Conclusion Transnational education carries a contradiction at its centre. It exists because a foreign qualification is worth more than a domestic one in the eyes of students, employers and often governments. That extra value comes from the reputation of the exporting institution, and that reputation is protected by rules which require the offshore provision to resemble the onshore original. The more faithfully the provision is replicated, the more valuable the qualification and the less it belongs to the place where it is taught. This article has used the idea of #coercive_isomorphism to show how that contradiction is enforced in practice, through four channels: host state licensing, home system quality assurance and its equivalence rule, professional accreditation, and dependence on funding and reputation. It has shown that these channels do not all pull in the same direction, and that host state licensing in particular can be a powerful ally of cultural preservation rather than an enemy of it. It has identified three counter pressures, host society expectation, staff agency and epistemic critique, and it has argued that the second of these, the professional judgement of teaching staff, is both the most effective and the most fragile. The article's central argument is that the tension between global standards and local culture is not a trade off along a single line. Good accreditation standards contain a hard core that protects students everywhere and deserves defending everywhere. They also contain a soft shell of professional habit, inherited from the system that wrote them, which is enforced as though it were essential and is not. Separating the two, standard by standard, and negotiating the shell while defending the core, is the practical work of transnational quality management. It is what this article has called negotiated compliance. That work will not be done by writing better strategy documents. It will be done by teaching staff who know both the discipline and the place, who have the security to argue, and who are supported by institutions willing to spend the effort to make local adaptation visible and defensible rather than hidden. Whether transnational education becomes a genuine meeting of knowledge systems or remains an efficient export industry will depend, in the end, on whether anyone is willing to pay for that. References Carvalho, N., Rosa, M. J., and Amaral, A. (2023). Cross border higher education and quality assurance: results from a systematic literature review. Journal of Studies in International Education. https://doi.org/10.1177/10283153221076900 Coutet, K. (2022). International school accreditation: an isomorphic force against creativity in a growing competitive market. Journal of Research in International Education, 21(2), 105 to 122. https://doi.org/10.1177/14752409221117252 de Wit, H., and Altbach, P. G. (2021). Internationalization in higher education: global trends and recommendations for its future. Policy Reviews in Higher Education, 5(1), 28 to 46. https://doi.org/10.1080/23322969.2020.1820898 DiMaggio, P. J., and Powell, W. W. (1983). The iron cage revisited: institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147 to 160. https://doi.org/10.2307/2095101 Guil Gorostidi, S. del C., and Rubio Arostegui, J. A. (2026). Quality management in higher education from the perspective of institutional isomorphism: a scoping review. Frontiers in Education, 10, 1720224. https://doi.org/10.3389/feduc.2025.1720224 Hou, A. Y. C., Lin, A. F. Y., Su, E. H. C., Chen, Y., and Hill, C. (2024a). Revisiting quality assurance and qualification recognition of cross border higher education during pandemic: does physical mobility matter from an Asian perspective. Asian Education and Development Studies, 13(1), 78 to 94. https://doi.org/10.1108/AEDS-10-2023-0147 Hou, A. Y. C., Tao, C. H., Zhou, K. Z., Lin, A. F. Y., Su, E. H. C., and Chen, Y. (2024b). Evolution of quality assurance in higher education from INQAAHE GGP to ISGs: are quality assurance agencies in Asia ready for the emerging modules. Journal of International Cooperation in Education, 26(1), 85 to 100. https://doi.org/10.1108/JICE-09-2023-0022 Lu, H. P., and Wang, J. C. (2023). Exploring the effects of sudden institutional coercive pressure on digital transformation in colleges from teachers perspective. Education and Information Technologies, 28, 15991 to 16015. https://doi.org/10.1007/s10639-023-11781-x Marginson, S. (2022). What is global higher education. Oxford Review of Education, 48(4), 492 to 517. https://doi.org/10.1080/03054985.2022.2061438 Marginson, S., and Xu, X. (2023). Hegemony and inequality in global science: problems of the center periphery model. Comparative Education Review, 67(1). https://doi.org/10.1086/722760 Meyer, J. W., and Rowan, B. (1977). Institutionalized organizations: formal structure as myth and ceremony. American Journal of Sociology, 83(2), 340 to 363. https://doi.org/10.1086/226550 Miani, M. (2026). Glocal tensions in transnational higher education in China: the case of a foundation year in a Sino foreign cooperative university. Higher Education Quarterly, 80(1). https://doi.org/10.1111/hequ.70085 Mok, K. H. (2021). Education market with the Chinese characteristics: the rise of minban and transnational higher education in China. Higher Education Quarterly, 75(3), 398 to 417. https://doi.org/10.1111/hequ.12323 Mok, K. H., and Marginson, S. (2021). Massification, diversification and internationalisation of higher education in China: critical reflections of developments in the last two decades. International Journal of Educational Development, 84. Nguyen, H. T. L., Tran, L. T., Luu, A. P. T., Luong, D. H., and Pham, H. H. (2025). Trends and patterns of research on transnational education: a bibliometric analysis with Scopus data, 1972 to 2023. Journal of Studies in International Education. https://doi.org/10.1177/10283153251375406 Nunez Valdes, K., Quiros Alpera, S., and Cerda Suarez, L. M. (2021). An institutional perspective for evaluating digital transformation in higher education: insights from the Chilean case. Sustainability, 13(17), 9850. https://doi.org/10.3390/su13179850 Thondhlana, J., Garwe, E. C., de Wit, H., Gacel Avila, J., Huang, F., and Tamrat, W. (Eds.). (2021). The Bloomsbury Handbook of the Internationalization of Higher Education in the Global South. Bloomsbury Academic. Wilkins, S., and Annabi, C. A. (2023). Academic careers in transnational higher education: the rewards and challenges of teaching at international branch campuses. Journal of Studies in International Education, 27(3). https://doi.org/10.1177/10283153211052782 Wilkins, S., and Huisman, J. (2024). Transnational education redefined. International Higher Education, 121. Wilkins, S., and Muratov, B. (2024). Uzbekistan ranks third globally as host of international branch campuses. International Higher Education, 119. Xu, M. (2025). How cultural discount and strategic adaptation sustain international branch campuses in Malaysia. SAGE Open. https://doi.org/10.1177/21582440251411277 Yang, H. (2023). China's regulatory framework for transnational higher education: development, impacts, and rationales. International Journal of Chinese Education, 12(1). https://doi.org/10.1177/2212585X231162208 Zhan, T., and Marginson, S. (2024). Institutional dual identity in research capacity building in international branch campuses: the case of NYU Shanghai. Higher Education, 87(2), 471 to 490. https://doi.org/10.1007/s10734-023-01017-9 #transnational_higher_education #cross_border_education #international_branch_campus #global_accreditation #quality_assurance_in_education #institutional_theory #neo_institutionalism #curriculum_localisation #education_policy #higher_education_governance #academic_standards #glocal_education #cultural_identity_in_education #internationalisation_of_higher_education #education_research

Generative language models have moved from novelty to routine study tool in less than four years. Survey evidence from the United Kingdom shows that almost all undergraduates now use these systems in some form, and that the great majority use them in work that is graded. Universities built their integrity systems for a different problem: copied text that could be matched to a visible source. That old model does not fit a situation in which a machine can produce fluent, unpublished, source-free prose on demand. This article reviews the research published between 2021 and 2026 on plagiarism, originality, text detection, assessment design, and institutional policy, and it uses that evidence to argue for a rebuilt #academic_integrity framework. The review shows three consistent findings. First, machine detection of #generative_AI writing is unreliable enough, and easy enough to defeat, that it cannot carry the weight of a misconduct decision on its own. Second, the harms of misclassification fall unevenly, with multilingual and neurodivergent writers most exposed. Third, the most defensible responses are structural rather than forensic: they change what is assessed, how evidence of learning is gathered, and how decisions about suspected misconduct are made. The article proposes a layered framework built on five components: clarity of permitted use, assessment redesign around validity, process evidence, proportionate and fair investigation, and shared capability building for staff and students. The framework is offered as a practical instrument for institutions and as a guide for students who want to work honestly in a world where the machine is always within reach. Keywords: academic integrity; generative artificial intelligence; #plagiarism; #originality; #AI_detection; assessment design; higher education policy; academic misconduct 1. Introduction For roughly twenty years, the working definition of #plagiarism in universities was stable and, in a narrow way, workable. A student submitted a text. Software compared that text against a large index of published and previously submitted work. Where the words matched, the software drew a line under them and produced a percentage. Staff read the report, judged whether the overlap was properly cited, and acted. The system had well known flaws, but it rested on one comfortable assumption: if a student had not written the words, someone else had, and that someone else had left a trace somewhere. That assumption no longer holds. A large language model produces text that has never existed before, that matches no source in any index, and that can be tuned in seconds to sound like a nervous first-year student or a confident final-year one. The trace is gone. What remains is a submission that looks like writing but may contain no writing at all, in the sense that #higher_education has traditionally meant the word. The scale of use is no longer a matter of speculation. The 2026 HEPI and Kortext survey of full-time undergraduates in the United Kingdom found that ninety-five per cent of students report using artificial intelligence in at least one way, and ninety-four per cent say they use generative tools to help with assessed work (Stephenson and Armstrong, 2026). The same survey reports that twelve per cent of students say they have put machine-generated text directly into assessed work, up from eight per cent a year earlier and three per cent the year before that. The previous edition of the survey recorded an equally sharp jump, from fifty-three per cent to eighty-eight per cent of students using generative tools for assessments in a single year (Freeman, 2025). Whatever one thinks about self-report data, the direction is unmistakable. The tools are not at the door. They are in the room, on the desk, inside the word processor, and increasingly inside the library platform that the institution itself pays for. Institutions have responded in two broad ways. The first response is forensic. Buy a detector, scan the submissions, flag the suspicious ones, and run the misconduct process as before. The second response is structural. Change the assessment, state clearly what use of the tools is permitted, gather evidence of the process and not only the product, and build the judgement of staff and students rather than outsourcing that judgement to a classifier. This article argues, on the basis of the published evidence, that the first response is failing and that the second is where the sustainable work lies. That is not a claim that detection has no place at all. It is a claim about weight: about what a probability score can and cannot justify when a student's degree, visa, or professional registration is at stake. The argument matters for students in a direct and personal way. A false accusation of #academic_misconduct is not a technicality. It is months of stress, a mark withheld, a record that may follow a person into professional life, and in some jurisdictions a threat to the right to remain in the country. The 2026 HEPI survey found that students articulate real anxiety about being wrongly accused (Stephenson and Armstrong, 2026). At the same time, a system that quietly tolerates the wholesale outsourcing of thinking betrays the students who do the work and devalues the qualification that everyone is paying for. A defensible framework has to hold both concerns at once. This article proceeds as follows. Section 2 traces how the integrity apparatus of the university was built and why its foundations have shifted. Section 3 examines the concept of #originality itself, which turns out to be far less settled than the policy documents assume. Section 4 sets out the method of the review. Section 5 presents the findings across six themes: detection accuracy, adversarial evasion, equity effects, provenance and watermarking, institutional policy, and the views of students and staff. Section 6 proposes a layered framework. Sections 7 and 8 turn to implementation and to what all of this means in practice for a student sitting down to write. Sections 9 and 10 discuss limitations and conclude. 2. Background: How the Integrity Apparatus Was Built, and Why Its Foundations Moved 2.1 The similarity paradigm The dominant technology of academic integrity for two decades was #text_matching. Its logic was simple. Text that appears in a student submission and also appears in an indexed source is, potentially, copied. The output was a #similarity_score, a number that staff learned to read with practised suspicion. The number was never meant to be a verdict. Vendors said so, policies said so, and training sessions said so. In practice, the number often functioned as one, because a number is easy to act on and a judgement is hard. The similarity paradigm shaped how universities defined the offence. Policies were written around the reproduction of another person's words or ideas without acknowledgement. The word "another" was doing quiet but heavy work. It presupposed a human author somewhere upstream who owned the words. It also presupposed that the wrong being done was a wrong of attribution: a failure to point back to a source that existed. 2.2 The outsourcing turn The similarity paradigm was already under strain before 2022. #Contract_cheating, in which a student commissions bespoke work from a third party, produces text that matches nothing, because it is written to order. Text-matching software is close to blind to it. Institutions responded with a mixture of authorship analysis, stylistic comparison, and, more usefully, changes to assessment design that made outsourcing harder or less attractive (Dawson, 2021). The literature on cheating in this period is large, and Tight (2024) offers a useful synthesis of what that research has and has not established. The outsourcing turn taught the sector an important lesson that it has been slow to apply: the offence is not really about matching words. It is about a false claim of authorship. A student who submits commissioned work has not copied anything. They have lied about who did the thinking. The wrong is a wrong of #authorship and of misrepresentation, not primarily of #source_attribution. Policies that were written around the older wrong were poorly equipped for the newer one, and that mismatch is exactly the gap into which generative models arrived. 2.3 Generative models and the break in the chain A generative model breaks the chain between text and source in a way that essay mills never quite did. The essay mill still involved a person who wrote something. The model involves a statistical process that produces plausible sequences. There is no upstream author to name, no document to cite, no index entry to match. If a student passes off model output as their own composition, the traditional plagiarism definition strains to describe what has gone wrong, and the traditional #detection_tools cannot find it by matching. The early literature registered this shock quickly. Perkins (2023) set out the challenge for #academic_integrity policy in the immediate aftermath of the public release of conversational models, arguing that existing definitions of misconduct would need rewriting rather than patching. Cotton, Cotton and Shipway (2024) reached similar conclusions from the perspective of assessment practice, noting that many standard tasks could be completed to a passing standard by a model in seconds. Sullivan, Kelly and McLaughlan (2023) and Rudolph, Tan and Tan (2023) documented the same disruption across disciplines and institutions. Eaton (2023) went further, proposing that the sector had entered a condition she named postplagiarism, in which hybrid human and machine writing becomes normal and the old binary between original and copied ceases to describe reality. She later extended this argument, arguing that integrity frameworks must be built for a future in which human and machine contributions to text are routinely blended rather than separated (Eaton, 2025). The point is not that the concept of misconduct disappears. It is that its centre of gravity moves. What the institution needs to know is not "did these words appear somewhere else" but "did this student do the learning that this task was designed to require, and have they been honest about how the work was produced." That is a question about #assessment_security, about honesty, and above all about validity. It is not a question that a similarity index was ever built to answer. 3. What Do We Actually Mean by Originality? Before proposing a framework, it is worth pausing on the word that sits at the centre of every policy document and that almost no policy document defines: #originality. In the everyday language of universities, originality carries at least four distinct meanings, and they are frequently confused. Originality as novelty. The work says something that has not been said before. This is the sense used in doctoral examination. It is a demanding standard, and almost no undergraduate assessment actually requires it. A first-year essay on the causes of the French Revolution is not expected to advance the historiography. Originality as provenance. The words came from this person's hands. This is the sense that text-matching software operationalises. It is a claim about the causal history of a document, not about the quality or novelty of its content. Originality as effort. The student did the reading, the thinking, and the drafting. This is what most staff have in mind when they complain that machine-written work is not original: not that it lacks novelty, and not merely that the words came from elsewhere, but that the cognitive labour the task was meant to induce did not happen. Originality as authenticity of voice. The work sounds like this person, reflects their reasoning, and can be defended by them in conversation. This is the sense that oral examination tests. Most institutional policies slide between these meanings without noticing. A policy can demand originality in the sense of provenance while assessing tasks that require no novelty and while permitting extensive use of spelling checkers, grammar tools, reference managers, translation software, and peer feedback, all of which alter the causal history of the document. The line between acceptable and unacceptable assistance was never drawn on a principled basis. It was drawn by habit, and habit is now being asked to bear weight it cannot carry. Corbin, Dawson, Nicola-Richmond and Partridge (2025) found exactly this when they asked staff and students where the line sits: participants could not locate it consistently, and many recognised that the line they were defending was arbitrary. Dawson, Bearman, Dollinger and Boud (2024) offer the most useful reframing available in the recent literature. They argue that the sector's fixation on cheating as a moral category is a distraction, and that the real concern is #validity: whether the assessment still licenses the inference we want to draw from it, namely that the student has met the intended #learning_outcomes. On this account, cheating matters because it destroys validity, not because it is naughty. The reframing has a sharp consequence. An anti-cheating measure that reduces cheating but also damages validity, for example by pushing all assessment into artificial, time-limited, memory-based conditions that test nothing anyone cares about, has made things worse rather than better. This is a liberating move for policy design. It shifts the question from "how do we catch them" to "what would make this assessment worth trusting." It also connects with a second strand of recent work on #evaluative_judgement, the capacity to judge the quality of work, including one's own and that produced by a machine. Bearman, Tai, Dawson, Boud and Ajjawi (2024) argue that in a world of generative tools this capacity becomes a central learning outcome rather than an incidental one. A graduate who cannot tell whether a model's output is accurate, well-reasoned, or subtly wrong is not employable in any field where the model will be present, which is to say any field at all. Bearman and Ajjawi (2023) make a parallel argument about learning to work with systems whose internal workings are opaque. The implication for the concept of originality is significant. If the meaningful sense of originality is #process_over_product, then the assessment system needs to see the process. And if the graduate capability we care about includes critical use of machines, then a blanket demand for machine-free provenance is not only unenforceable but pedagogically incoherent. What replaces it must be #transparency about use, combined with assessment that can tell whether the learning happened. 4. Method This article is an integrative review and conceptual synthesis rather than an empirical study. The purpose of an #integrative_review is to bring together findings from studies with different designs, including experimental evaluations, surveys, policy analyses, and conceptual papers, in order to build a coherent account of a field that is moving too quickly for a conventional systematic review to capture. Scope. The review covers literature published between 2021 and 2026, with the great majority of sources published from 2023 onwards. This window was chosen deliberately. Work published before 2021 addresses a technological situation that no longer exists, and its empirical claims about detection performance are obsolete. Sources. Peer-reviewed journal articles, scholarly books, conference proceedings in machine learning, and reports from recognised higher education policy bodies were included. Vendor marketing material and news commentary were excluded from the evidence base, although the article notes where vendor claims and independent findings diverge, because that divergence is itself part of the problem. Selection. Studies were selected for their relevance to five questions: how reliable are AI text detectors under realistic conditions; how easily are they defeated; who is harmed when they fail; what alternatives to detection exist; and how are institutions actually responding. Priority was given to studies that report empirical results with a stated method and to conceptual papers that have shaped the policy conversation. Synthesis. Findings were grouped thematically. Where studies disagree, the disagreement is reported rather than smoothed over, because the disagreements are informative. In particular, the evidence on detector accuracy has shifted between 2023 and 2026, and any framework built on a snapshot of tool performance will be obsolete before it is implemented. That instability is itself one of the review's central findings and is built into the #conceptual_framework proposed in Section 6. Limitations of the method. An integrative review reflects the judgement of its author in selecting and weighting sources. The field is dominated by research from Anglophone, high-income systems, and the reader should treat generalisations to other contexts with caution. These limits are revisited in Section 9. 5. Findings 5.1 Detection accuracy under realistic conditions The first wave of independent evaluations, published through 2023, was uniformly discouraging for anyone hoping that software would solve the problem. Weber-Wulff and colleagues (2023) tested a set of detection tools across a range of text types and found accuracy well below what would be needed for high-stakes use, with inconsistent results and a troubling combination of #false_positives and #false_negatives. Their conclusion was blunt: the tools were not suitable as the sole basis for a misconduct finding. Elkhatat, Elsaid and Almeer (2023) reached a similar verdict, reporting that performance varied sharply depending on which model had produced the text and how long the sample was. Chaka (2023) tested several commercial detectors and found their outputs unstable and frequently contradictory. Dalalah and Dalalah (2023) documented both error types and warned specifically about the consequences of false accusations. Two structural features of the technology explain these results and are unlikely to disappear. The first is the base rate problem. Suppose a detector has a false positive rate of one per cent, a figure that would be considered excellent. In a cohort submitting one hundred thousand assignments a year, that is a thousand innocent students flagged, quite apart from any true positives. If even a fraction of those flags lead to formal proceedings, the institution has manufactured a serious injustice at industrial scale while congratulating itself on the tool's accuracy. Low error rates are not the same thing as low error counts, and #detector_reliability has to be judged against the volume of decisions being made, not against a benchmark figure in a vendor's brochure. The second is that detectors output probabilities and misconduct panels need facts. A score of seventy-eight per cent likelihood is #probabilistic_evidence about a text, not evidence about a person's conduct. It cannot be cross-examined. The student cannot inspect the reasoning. In most institutional procedures the standard of proof is the balance of probabilities, and a naive reading might suggest that a seventy-eight per cent score meets it. That reading is a statistical error. The score is a property of a classifier applied to a document, not a posterior probability that a particular student cheated, and treating it as the latter without any other evidence is precisely the kind of reasoning that appeals panels and courts have historically found unsafe. The picture is not entirely static, and honesty requires acknowledging that. More recent evaluations published in 2026 report that some tools have improved substantially. Van Vlasselaer, Van Droogenbroeck and Spruyt (2026) tested four widely used detectors against a controlled dataset containing fully human, fully machine, hybrid, and deliberately humanised texts. They found that one tool performed considerably better than the others across categories, and that false positives were rarer than earlier studies had reported. They also found that performance across tools remained uneven and highly context-dependent, and that findings of this kind date quickly as models change. Hadra, Cambridge and Mesbah (2026) similarly report inconsistent #classifier_accuracy across tools alongside continuing concerns about disclosure of methods by vendors. The correct reading of this trajectory is not "detection was bad and now it is fine." It is that detection performance is a moving target measured against a moving adversary, and that any policy which hard-codes reliance on a tool will be wrong within a year in one direction or the other. A framework must therefore be robust to the tool being better than expected and to its being worse. 5.2 Evasion is cheap, and that is the deeper problem Even if a detector performed perfectly on unaltered machine text, it would face a student population with access to trivial countermeasures. Perkins, Roe, Vu, Postma, Hickerson, McGaughran and Khuat (2024) generated samples with three leading models, applied six simple #adversarial_techniques of the kind a student could apply in a few minutes, and tested the results against seven detectors across more than eight hundred assessments. Detector accuracy fell by roughly seventeen per cent. The authors concluded that the tools could not be recommended as the basis for misconduct decisions, both because of accuracy limits and because of the risk of false accusation. Their finding is consistent with earlier work by Krishna and colleagues (2023), who showed in a machine learning context that paraphrasing pipelines substantially degrade the performance of detectors, and with the theoretical argument advanced by Sadasivan and colleagues (2023), who reasoned that as generated text approaches the distribution of human text, reliable detection becomes fundamentally difficult rather than merely technically hard. Meanwhile an entire market of #paraphrasing_tools and #humanising_tools has grown up for precisely this purpose. Roe and Perkins (2022) flagged automated paraphrasing as an integrity threat before the current generation of models arrived; the threat has since been absorbed into general-purpose systems, which will happily rewrite their own output in a different register on request. There is a further complication that receives too little attention: most real student work is now #hybrid_text. A student drafts a paragraph, asks a model to tighten it, rewrites the result, asks for a counter-argument, incorporates part of it, and runs the whole through a grammar tool. What is the ground truth for such a document? It has no clean label. Detectors are trained on a binary distinction that does not correspond to how writing now actually happens, and this mismatch is not a bug to be patched but a feature of the world. Van Vlasselaer and colleagues (2026) explicitly included hybrid documents in their design for this reason, and the differential performance of tools on those documents is among the most practically important findings in the recent literature. The strategic conclusion follows directly. Building an integrity system on detection means committing an institution to an unwinnable technical contest against a counterparty with more time, more motivation, and free access to better tools than the institution has. Any system whose integrity depends on staying ahead in that contest is a system that has already lost, because it has made its own legitimacy hostage to a race it does not control. 5.3 Who gets hurt: equity and the distribution of error If detection errors fell randomly, they would still be a problem. They do not fall randomly, and that is what turns a technical limitation into an ethical one. Liang, Yuksekgonul, Mao, Wu and Zou (2023) provided the pivotal finding: detectors systematically misclassified writing by non-native English speakers as machine-generated. The mechanism is not mysterious. Many detectors rely on statistical measures of the predictability of word choice. Writing that uses a narrower range of vocabulary and more conventional constructions scores as more predictable, and a fluent second-language writer often writes exactly that way. The detector is not detecting machine authorship. It is detecting a particular kind of linguistic profile, and then labelling it. When that profile correlates with being an international student, the tool has become a mechanism for distributing suspicion along lines of language background. The same logic exposes other groups. Students who write in short, plain, conventional sentences because of a disability, because they use assistive technology, because they have been taught to write formulaically, or because they are simply cautious writers, present the same statistical profile. Perkins and colleagues (2024) drew the inclusion implication explicitly in the title of their study: detection failure is not a neutral technical shortfall but a threat to #inclusive_assessment, because the burden of #false_accusation lands on those least able to absorb it and least likely to be believed. The consequences of a wrongful accusation deserve to be stated plainly rather than abstracted into a discussion of error rates. The student is presumed dishonest on the basis of a number they cannot interrogate. They are asked to prove a negative: to demonstrate that they wrote something. Students who kept no drafts, who compose directly into a browser, or who work in a second language and rely on tools for polish are structurally disadvantaged in that process. The reported #student_anxiety about false accusation in the HEPI survey is not irrational; it is a rational response to a system in which a machine's guess can trigger a process with life-altering stakes (Stephenson and Armstrong, 2026). Any framework that ignores this is not an integrity framework. It is a liability. Human judgement, incidentally, does not rescue the situation. Fleckenstein, Meyer, Jansen, Keller, Koller and Moller (2024) found that teachers were poor at identifying machine-generated essays, performing close to chance in some conditions and, importantly, being overconfident in their judgements. Casal and Kessler (2023) found that even trained linguists could not reliably distinguish machine from human academic writing. Perkins, Roe and colleagues (2023) found that combining academic judgement with software improved matters somewhat but did not produce anything approaching certainty. The implication is uncomfortable: neither the human nor the machine can reliably read authorship off the surface of a finished text. That is not a failure of effort. It is a property of the problem, and a framework has to be designed around it rather than in denial of it. 5.4 Watermarking and provenance: promising, partial, and not yet a foundation If reading authorship off the text is unreliable, an obvious alternative is to build the signal in at the point of generation. #Watermarking schemes embed a statistical signature into model output that a verifier holding the right key can detect while a reader cannot see it. Kirchenbauer and colleagues (2023) demonstrated a workable scheme, and the machine learning literature has developed the idea considerably since. Lancaster (2023) examined whether watermarking offers a solution to the academic integrity problem and concluded, sensibly, that it offers part of one at best. The reasons are practical rather than theoretical. Watermarking only works if the model provider implements it. Open-weight models can be run locally with the watermarking removed. Paraphrasing degrades the signal. And a student who retypes model output, or who uses one system to rewrite the output of another, may strip it entirely. A watermarking regime therefore catches the careless and misses the deliberate, which is close to the opposite of what an integrity system should do. The more durable idea in this space is not watermarking but #provenance: the systematic capture of how a document came to exist. Editor #version_history, timestamped drafts, revision logs, annotated reading, and recorded planning all constitute evidence of process that is far harder to fabricate convincingly than a finished text is. #Authorship_verification through process evidence has the crucial property of shifting the burden away from a probabilistic guess about the artefact and towards a record of the work, and it does so in a way the student controls and can present in their own defence. It has costs, discussed in Section 7, but it is the most promising technical direction currently available, and unlike detection it does not depend on winning an arms race. 5.5 What institutions have actually done Moorhouse, Yeo and Wan (2023) analysed the guidance issued by the world's highest-ranked universities in the first months after the arrival of conversational models. Their finding was that guidance existed but was thin, inconsistent, and heavily oriented towards prohibition and detection rather than towards assessment redesign. Chan (2023) proposed a more comprehensive #institutional_policy framework covering governance, pedagogy, and operational practice, arguing that piecemeal responses would fail. Foltynek and colleagues (2023), writing for the European Network for Academic Integrity, issued recommendations that emphasised transparency of use, the responsibility of institutions to educate rather than merely to police, and the unreliability of detection as an evidential basis. National regulators moved in the same direction: the Australian tertiary regulator's expert panel argued for assessment reform rather than surveillance as the primary response (Lodge, Howard, Bearman and Dawson, 2023), and UNESCO's guidance placed the emphasis on capability and governance rather than on prohibition (Miao and Holmes, 2023). Two practical models have gained traction and deserve description because they are usable tomorrow morning. The first is the AI Assessment Scale. Perkins, Furze, Roe and MacVaugh (2024) proposed a graduated scale on which each assessment task is explicitly located, running from no permitted use of generative tools, through use for idea generation and planning, through use for drafting and editing with declaration, to full collaborative use in which working with the model is itself the assessed skill. The scale's value is not theoretical elegance. It is that it replaces a vague institutional injunction with a task-level statement that a student can actually follow. Furze, Perkins, Roe and MacVaugh (2024) reported on a pilot implementation and found that clarity at the task level reduced the ambiguity that had been driving both anxiety and unintentional misconduct. The #AI_assessment_scale has since been adapted for particular contexts, including English for academic purposes teaching. The second is what is often called the #two_lane_model. In this approach, an institution designates a subset of assessment as #secured_assessment, conducted under conditions where the identity and unaided performance of the student can be assured, and treats the remainder as #open_assessment in which generative tools may be used with declaration. The secured lane carries the burden of assuring that the graduate can actually do the thing; the open lane carries the burden of developing the higher-order capabilities that the tools do not remove. The attraction of the model is that it stops pretending that every piece of coursework can be simultaneously authentic, open, and secure. Those three properties are in tension, and a framework that admits the tension is more honest than one that does not. Corbin, Dawson and Liu (2025) issue an important warning about both models. Their argument, captured in their title, is that talk is cheap: institutional statements about assessment reform are common, structural change is rare, and the gap between policy language and actual assessment practice is where integrity systems quietly fail. Corbin, Bearman, Boud and Dawson (2025) describe the whole situation as a wicked problem, in the technical sense: one with no stopping rule, no single correct formulation, and no solution that does not create new problems. Both papers are a useful antidote to framework enthusiasm, including the enthusiasm of this article. The most consistent gap across every policy analysis in this literature is #staff_development. Policies are written centrally and implemented by individual academics who have received no training, who are marking under time pressure, and whose default response to a suspicious submission is to run it through whatever tool is nearest. No framework survives contact with an untrained, overworked marking team. 5.6 What students and staff actually think The #student_voice literature complicates the picture in a way that policymakers should welcome. Chan and Hu (2023) surveyed students across disciplines and found broadly positive attitudes towards generative tools alongside genuine concern about accuracy, over-reliance, and the erosion of their own skills. Students were not, on the whole, looking for a way to avoid learning. They were looking for support, for feedback that arrives faster than a tutor's, and for help with the parts of academic writing that they had never been taught properly. Gruenhagen and colleagues (2024) found much the same in their study of student perceptions and use: students report using chatbots for assistance, often in ways that are pedagogically defensible, and are frequently unclear about where the institutional line sits. The HEPI data reinforce this. Students report using the tools mainly to explain concepts, summarise reading, and generate ideas, with the direct insertion of machine text into assessed work remaining a minority behaviour, albeit a growing one (Freeman, 2025; Stephenson and Armstrong, 2026). The 2026 survey also records that around two-thirds of students perceive that assessment has changed significantly in response to the technology, which suggests that structural reform is happening even if unevenly. What emerges is a picture of a student body that is ahead of its institutions in practice and behind them in guidance. Students are using the tools daily and are largely guessing at the rules. Where guidance is vague, they fill the gap with rumour, with what a friend on another course was told, and with the assumption that whatever is not detected is permitted. That is not a moral failing. It is what any rational person does in an under-specified system, and it points directly at the first component of the framework proposed below. The staff picture is one of pressure. Markers are asked to make judgements they are not equipped to make, using tools whose limitations they have not been trained to understand, within procedures designed for a different offence. The corrosive effect on #trust is significant. A relationship in which the tutor is primarily an investigator and the student is primarily a suspect is not one in which much learning happens, and the drift towards #surveillance carries pedagogical costs that rarely appear on the balance sheet. 6. A Layered Framework for Academic Integrity in the Age of Generative Models The evidence reviewed above supports a framework built on five interlocking components. They are presented as layers because no single one is sufficient, and because each compensates for the weaknesses of the others. The organising principle is #integrity_by_design: integrity should be built into the design of assessment and policy rather than bolted on afterwards through detection. Layer 1: Specify permitted use at the level of the task The single cheapest and most effective intervention available to any institution is to stop issuing vague, institution-wide statements about generative tools and to require every assessment brief to state precisely what use is permitted. The AI Assessment Scale (Perkins, Furze, Roe and MacVaugh, 2024) provides a ready-made vocabulary for this. Whether an institution adopts that specific scale or builds its own, the requirements are the same: the statement must be at the task level, not the policy level; it must be written in language a student can understand without a lawyer; it must be positive as well as prohibitive, saying what is allowed rather than only what is forbidden; and it must be consistent within a programme, because students taking five modules with five different unstated expectations will simply give up trying to comply. This layer alone resolves a large proportion of the integrity problem, because a large proportion of the integrity problem is not cheating. It is confusion. A student who uses a model to check their grammar and is then accused of misconduct has not cheated. They have been failed by a system that never told them where the line was. Layer 2: Redesign assessment around validity, not around detectability Following Dawson and colleagues (2024), the design question is not "can this be done by a machine" but "does this assessment still license the inference I want to draw." Some concrete moves follow. Assess the things that models are bad at, which currently include reasoning from context that only the student has, engaging with material generated in class, defending a position under questioning, connecting theory to a specific and personal placement or laboratory experience, and producing work that must be true of a particular local situation. Reduce the weight placed on the polished, decontextualised, take-home essay, which is now the single most vulnerable assessment format in the university and, not coincidentally, the format that has always been most vulnerable to contract cheating. Use #programmatic_assessment logic. Rather than trying to make every task individually secure, design a programme in which multiple pieces of evidence across time and modes together support the inference that the student has met the outcomes. A single compromised essay in such a system is a data point, not a catastrophe. Corbin, Dawson and Liu (2025) are right that this requires structural change at programme level and cannot be achieved by exhortation to individual academics. Add #oral_defence where the stakes justify it. A short viva on submitted work is the most robust available check on whether the student can actually stand behind what they submitted, and it tests exactly the capability we say we value. It is also expensive, so it should be deployed selectively: on high-stakes work, on a sample basis, or as a follow-up where concerns arise. Used as a routine, sampled component rather than as a punishment, it removes the accusatory framing entirely, because everyone is asked. Make the use of the tools itself an object of assessment. Ask students to submit a model transcript with a critical commentary on where the model was wrong. Ask them to fact-check a machine-generated literature summary against the actual sources. This develops #evaluative_judgement in the sense described by Bearman and colleagues (2024), and it has the pleasing property that a student who outsources this task to the model produces something visibly worse. Layer 3: Gather process evidence Where the product cannot be trusted, look at the process. #Process_evidence includes staged submissions with feedback at each stage, annotated bibliographies produced before the draft, planning documents, editor version histories, reflective commentaries, and in-class writing. The point is not to build a surveillance apparatus but to create a natural record of the work as it developed. This has three advantages. It gives the student a defence, which matters enormously given the equity findings in Section 5.3. It gives the institution something to look at that is not a probability score. And it improves learning independently of integrity, because staged work with feedback is simply better pedagogy than a single terminal submission. There are real costs. #Record_keeping requirements can burden students, particularly those working on shared or public computers, those with unstable housing, and those with disabilities. Version histories raise #data_protection questions that most institutions have not thought through. Process evidence must therefore be requested proportionately and with alternatives available, and it must never become a requirement that only the well-resourced student can satisfy. Layer 4: Investigate fairly, and never on a score alone This layer is where institutions currently create the most avoidable harm, and it can be stated as a set of rules. A detector output is a prompt to look, never a finding. It may justify a conversation. It may never, on its own, justify a penalty. Every serious independent evaluation reviewed in Section 5 reaches this conclusion, and institutions that have not written it into policy are carrying an unquantified legal and ethical exposure. The #burden_of_proof stays with the institution. Asking a student to prove they wrote something, on the basis of a score, inverts the burden and is indefensible. Where the institution alleges misconduct, the institution must produce evidence of misconduct. Evidence must be triangulated. #Triangulation means combining sources of a different kind: the student's account, process evidence, a conversation about the content, the student's demonstrated capability elsewhere, and, if used at all, tool output as one strand among several. #Academic_judgement is essential here, but Section 5.3 shows it is also unreliable on its own, so it must be exercised on evidence rather than on a hunch about writing style. #Due_process must be real. The student should be told what the concern is, shown the material on which it rests, given time to respond, allowed support, and given a route of appeal to someone who was not part of the original judgement. #Proportionality means that a first-year student who used a model to rephrase two sentences is not treated like a final-year student who submitted a wholly machine-written dissertation. Panels must be trained in what a probability score is and is not. A panel that believes a seventy-eight per cent score means a seventy-eight per cent chance the student cheated is a panel that will produce unsafe findings, and no amount of procedural correctness elsewhere will fix that. Layer 5: Build capability, in students and in staff The final layer is the one most likely to be cut when budgets are tight, and it is the one on which the others depend. For students, #AI_literacy means understanding what these systems are: pattern-completion machines that produce plausible text with no mechanism guaranteeing that the text is true. It means knowing that they invent citations, that they are confident when wrong, that they carry the biases of their training data, and that they cannot be trusted with facts the student cannot check. It also means being taught, explicitly, how to use them well: as a tutor to explain a difficult concept, as a critic of one's own argument, as a generator of counter-arguments, as a starting point that must then be verified. Very few institutions teach any of this, and then express surprise when students use the tools badly. For staff, capability means understanding detection well enough to stop over-trusting it, understanding assessment design well enough to redesign a module rather than merely complain about it, and having the time to do so. This is a resourcing question dressed up as a training question, and it will not be solved by another policy document. Finally, #declaration_statements should be normalised. A short statement at the end of a submission describing what tools were used and how does more for integrity culture than any detector, precisely because it treats the student as a person capable of honesty rather than as a suspect awaiting classification. Declarations are not self-enforcing, and a dishonest student will file a dishonest declaration. That objection misses the point. The purpose of a declaration is not to catch the dishonest; it is to make honesty easy, explicit, and normal for the large majority who want to comply, and to make the dishonest act a deliberate lie rather than a grey-area drift. 7. Implementation: What This Requires of an Institution A framework that cannot be implemented is a literature review with ambitions. Four practical conditions determine whether the layers above become real. Curriculum-level governance. #Curriculum_design decisions must be made at programme level, because the assessment vulnerabilities of a degree are a property of the whole degree and not of any one module. #Governance structures should require programme teams to map where their secured evidence of learning actually sits, and to notice when the honest answer is nowhere. Investment in staff time. #Staff_training is necessary but insufficient. Redesigning a module's assessment takes days, not hours, and a system that expects it to be done in the gaps between marking will get compliance theatre instead of reform. Corbin, Dawson and Liu (2025) are precise about this: without structural investment, institutional statements about assessment change remain talk. Procurement discipline. If detection tools are purchased at all, they should be procured with an explicit statement of what they may be used for and, crucially, what they may not. Vendor accuracy claims should be treated as marketing until independently verified, and the independent evidence should be read with attention to the year in which the testing was done. Contracts should be reviewed against #data_protection obligations, since these tools process student work, and in some cases the terms under which they do so have not been adequately scrutinised. Consistency of message. Students learn the rules from what happens, not from what is written. If one lecturer bans the tools, another requires them, and a third quietly runs everything through a detector, the effective institutional policy is chaos, and students will act accordingly. 8. What This Means for a Student Because this article is written for students as much as about them, it is worth translating the framework into direct terms. Assume the rules differ by task, and ask. The most common route into trouble is not deliberate cheating. It is a reasonable assumption that turned out to be wrong. If the brief does not say what use of generative tools is permitted, ask, and keep the answer in writing. Keep your process. Write in a tool that keeps a version history. Keep your notes, your outline, your reading. This costs nothing and it is the single most effective protection against a wrongful accusation, which, as Section 5.3 makes clear, is a risk that falls hardest on international students and on anyone whose writing is plain and conventional. #Reflective_practice about your own drafting is also, independently, a good way to write better. Use the tools for thinking, not instead of thinking. A model is an excellent explainer of concepts you do not understand, an excellent generator of objections to your argument, and an unreliable source of facts. It will invent references that look perfectly real. It will state falsehoods with total confidence. If you cannot check a claim, do not use it. #Self_regulation here is not a moral nicety; it is the difference between a graduate who can think and one who can only prompt. Declare what you did. #Ethical_use is not a matter of avoiding detection. It is a matter of being able to say honestly what your contribution was. If you would be uncomfortable describing your process to your tutor, that discomfort is the signal, and it arrives long before any detector does. Remember what the assessment is for. Every #study_skills conversation eventually reaches the same point. The degree is not the paper. The degree is the capability that the paper is meant to evidence. A student who outsources the work has bought the paper and skipped the capability, and the labour market will find that out faster than the misconduct panel ever could. 9. Limitations and Directions for Research This review has clear limits. It is not a systematic review, and its selection of sources reflects the judgement of the author. It draws overwhelmingly on research from English-speaking, high-income higher education systems, and its conclusions may not transfer cleanly to contexts with different assessment traditions, different resource levels, or different linguistic profiles. The equity findings in particular deserve replication in a wider range of language settings, since the evidence that detectors penalise second-language writers comes largely from studies of English. The empirical claims about detector performance carry an expiry date. The 2023 evaluations and the 2026 evaluations reach materially different conclusions about accuracy, and there is no reason to expect the picture in 2028 to resemble either. Any institution reading this article in a future year should treat Section 5.1 as a method for evaluating evidence rather than as a settled result. Several questions remain open. There is little robust evidence on the effect of the various proposed frameworks on actual student behaviour and learning outcomes, as opposed to on staff and student attitudes. There is almost no research on the experience of students who have been wrongly accused, which is a striking gap given how central that harm is to the ethical case. The learning effects of routine model use are poorly understood: whether the tools build or erode the capabilities they appear to support is the most important open question in the field and is not answerable with cross-sectional surveys. #Future_research should prioritise longitudinal designs, programme-level rather than task-level interventions, and studies conducted outside the Anglophone core. 10. Conclusion The generative model has not created a new kind of dishonesty. It has exposed how much of the university's integrity apparatus rested on a technical accident: the fact that copied text used to leave a trace. That accident is over, and no amount of investment in detection will bring it back. What is left is the thing that was always underneath. A university awards a qualification as a claim about a person's capability. Everything else, every policy, every detector, every misconduct panel, exists to protect the truth of that claim. When the claim is protected by a probability score that misfires on international students and can be defeated by a student with five minutes and a paraphrasing tool, the protection is an illusion, and an actively harmful one. The alternative is not comfortable, because it is expensive and slow. It means designing assessment that is worth trusting rather than assessment that is merely hard to fake. It means telling students clearly what they may do. It means looking at process rather than guessing at product. It means investigating fairly and never on a number alone. And it means building the capability, in students and staff alike, to work critically alongside systems that are not going away. None of this is achieved by a tool. It is achieved by a #culture_of_integrity in which honesty is easy, expectations are clear, and #human_judgement is exercised on evidence rather than on suspicion. #Trustworthy_assessment in the age of the machine will be built by people, or it will not be built at all. Hashtags #academic_integrity_frameworks #generative_AI_in_education #AI_text_detection #plagiarism_policy #originality_in_writing #assessment_redesign #academic_honesty #AI_literacy_for_students #higher_education_research #false_accusation_risk #student_assessment #educational_technology #university_policy #responsible_AI_use #future_of_assessment References Bearman, M. and Ajjawi, R. (2023) Learning to work with the black box: Pedagogy for a world with artificial intelligence. British Journal of Educational Technology, 54(5), 1160-1173. https://doi.org/10.1111/bjet.13337 Bearman, M., Tai, J., Dawson, P., Boud, D. and Ajjawi, R. (2024) Developing evaluative judgement for a time of generative artificial intelligence. Assessment and Evaluation in Higher Education, 49(6), 893-905. https://doi.org/10.1080/02602938.2024.2335321 Casal, J. E. and Kessler, M. (2023) Can linguists distinguish between ChatGPT and human writing? A study of research ethics and academic publishing. Research Methods in Applied Linguistics, 2(3), 100068. https://doi.org/10.1016/j.rmal.2023.100068 Chaka, C. (2023) Detecting AI content in responses generated by ChatGPT, YouChat, and Chatsonic: The case of five AI content detection tools. Journal of Applied Learning and Teaching, 6(2). https://doi.org/10.37074/jalt.2023.6.2.12 Chan, C. K. Y. (2023) A comprehensive AI policy education framework for university teaching and learning. International Journal of Educational Technology in Higher Education, 20, 38. https://doi.org/10.1186/s41239-023-00408-3 Chan, C. K. Y. and Hu, W. (2023) Students' voices on generative AI: Perceptions, benefits, and challenges in higher education. International Journal of Educational Technology in Higher Education, 20, 43. https://doi.org/10.1186/s41239-023-00411-8 Corbin, T., Bearman, M., Boud, D. and Dawson, P. (2025) The wicked problem of AI and assessment. Assessment and Evaluation in Higher Education, 1-17. https://doi.org/10.1080/02602938.2025.2553340 Corbin, T., Dawson, P. and Liu, D. (2025) Talk is cheap: Why structural assessment changes are needed for a time of GenAI. Assessment and Evaluation in Higher Education, 50(7), 1087-1097. https://doi.org/10.1080/02602938.2025.2503964 Corbin, T., Dawson, P., Nicola-Richmond, K. and Partridge, H. (2025) Where is the line? Towards a framework for acceptable uses of AI in assessment. Assessment and Evaluation in Higher Education. https://doi.org/10.1080/02602938.2025.2456207 Cotton, D. R. E., Cotton, P. A. and Shipway, J. R. (2024) Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61(2), 228-239. https://doi.org/10.1080/14703297.2023.2190148 Dalalah, D. and Dalalah, O. M. A. (2023) The false positives and false negatives of generative AI detection tools in education and academic research: The case of ChatGPT. The International Journal of Management Education, 21(2), 100822. https://doi.org/10.1016/j.ijme.2023.100822 Dawson, P. (2021) Defending Assessment Security in a Digital World: Preventing E-Cheating and Supporting Academic Integrity in Higher Education. Abingdon: Routledge. Dawson, P., Bearman, M., Dollinger, M. and Boud, D. (2024) Validity matters more than cheating. Assessment and Evaluation in Higher Education, 49(7), 1005-1016. https://doi.org/10.1080/02602938.2024.2386662 Eaton, S. E. (2021) Plagiarism in Higher Education: Tackling Tomorrow's Challenges. Santa Barbara: Libraries Unlimited. Eaton, S. E. (2023) Postplagiarism: Transdisciplinary ethics and integrity in the age of artificial intelligence and neurotechnology. International Journal for Educational Integrity, 19, 23. https://doi.org/10.1007/s40979-023-00144-1 Eaton, S. E. (2025) Global trends in education: Artificial intelligence, postplagiarism, and future-focused learning for 2025 and beyond. International Journal for Educational Integrity, 21, 12. https://doi.org/10.1007/s40979-025-00185-8 Elkhatat, A. M., Elsaid, K. and Almeer, S. (2023) Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19, 17. https://doi.org/10.1007/s40979-023-00140-5 Fleckenstein, J., Meyer, J., Jansen, T., Keller, S. D., Koller, O. and Moller, J. (2024) Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays. Computers and Education: Artificial Intelligence, 6, 100209. https://doi.org/10.1016/j.caeai.2024.100209 Foltynek, T., Bjelobaba, S., Glendinning, I., Khan, Z. R., Santos, R., Pavletic, P. and Kravjar, J. (2023) ENAI recommendations on the ethical use of artificial intelligence in education. International Journal for Educational Integrity, 19, 12. https://doi.org/10.1007/s40979-023-00133-4 Freeman, J. (2025) Student Generative AI Survey 2025. HEPI Policy Note 61. Oxford: Higher Education Policy Institute. Furze, L., Perkins, M., Roe, J. and MacVaugh, J. (2024) The AI Assessment Scale in action: A pilot implementation of GenAI-supported assessment. Australasian Journal of Educational Technology, 40(4), 38-55. https://doi.org/10.14742/ajet.9434 Gruenhagen, J. H., Sinclair, P. M., Carroll, J. A., Baker, P. R. A., Wilson, A. and Demant, D. (2024) The rapid rise of generative AI and its implications for academic integrity: Students' perceptions and use of chatbots for assistance with assessments. Computers and Education: Artificial Intelligence, 7, 100273. https://doi.org/10.1016/j.caeai.2024.100273 Hadra, M., Cambridge, K. and Mesbah, M. (2026) Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity, 22. https://doi.org/10.1007/s40979-026-00213-1 Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I. and Goldstein, T. (2023) A watermark for large language models. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 17061-17084. Krishna, K., Song, Y., Karpinska, M., Wieting, J. and Iyyer, M. (2023) Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36. Lancaster, T. (2023) Artificial intelligence, text generation tools and ChatGPT: Does digital watermarking offer a solution? International Journal for Educational Integrity, 19, 10. https://doi.org/10.1007/s40979-023-00131-6 Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023) GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 Lodge, J. M., Howard, S., Bearman, M. and Dawson, P. (2023) Assessment Reform for the Age of Artificial Intelligence. Melbourne: Tertiary Education Quality and Standards Agency. Miao, F. and Holmes, W. (2023) Guidance for Generative AI in Education and Research. Paris: UNESCO. Moorhouse, B. L., Yeo, M. A. and Wan, Y. (2023) Generative AI tools and assessment: Guidelines of the world's top-ranking universities. Computers and Education Open, 5, 100151. https://doi.org/10.1016/j.caeo.2023.100151 Perkins, M. (2023) Academic integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond. Journal of University Teaching and Learning Practice, 20(2). https://doi.org/10.53761/1.20.02.07 Perkins, M., Furze, L., Roe, J. and MacVaugh, J. (2024) The AI Assessment Scale: A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching and Learning Practice, 21(6). https://doi.org/10.53761/q3azde36 Perkins, M., Roe, J., Postma, D., McGaughran, J. and Hickerson, D. (2024) Detection of GPT-4 generated text in higher education: Combining academic judgement and software to identify generative AI tool misuse. Journal of Academic Ethics, 22, 89-113. https://doi.org/10.1007/s10805-023-09492-6 Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J. and Khuat, H. Q. (2024) Simple techniques to bypass GenAI text detectors: Implications for inclusive education. International Journal of Educational Technology in Higher Education, 21, 53. https://doi.org/10.1186/s41239-024-00487-w Roe, J. and Perkins, M. (2022) What are automated paraphrasing tools and how do we address them? A review of a growing threat to academic integrity. International Journal for Educational Integrity, 18, 15. https://doi.org/10.1007/s40979-022-00109-w Rudolph, J., Tan, S. and Tan, S. (2023) ChatGPT: Bane or boon for higher education? Journal of Applied Learning and Teaching, 6(1), 342-363. https://doi.org/10.37074/jalt.2023.6.1.9 Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W. and Feizi, S. (2023) Can AI-generated text be reliably detected? arXiv preprint. https://doi.org/10.48550/arXiv.2303.11156 Stephenson, R. and Armstrong, C. (2026) Student Generative Artificial Intelligence Survey 2026. HEPI Report 199. Oxford: Higher Education Policy Institute. Sullivan, M., Kelly, A. and McLaughlan, P. (2023) ChatGPT in higher education: Considerations for academic integrity and student learning. Journal of Applied Learning and Teaching, 6(1), 31-40. https://doi.org/10.37074/jalt.2023.6.1.17 Tight, M. (2024) Challenging cheating in higher education: A review of research and practice. Assessment and Evaluation in Higher Education, 49(7), 911-923. https://doi.org/10.1080/02602938.2023.2300104 Van Vlasselaer, M., Van Droogenbroeck, F. and Spruyt, B. (2026) Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, 22, 16. https://doi.org/10.1007/s40979-026-00226-w Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltynek, T., Guerrero-Dib, J., Popoola, O., Sigut, P. and Waddington, L. (2023) Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z

Universities that operate across national borders now carry a compliance load that would have looked impossible to a campus administrator thirty years ago. A single institution may run degree programmes in five countries, answer to four national regulators, hold two or three programme level accreditations, and still be expected to prove that a student in one branch receives an education of the same standard as a student at the home campus. This article examines whether agentic workflows, meaning software systems built on large language models that can plan, use tools, retrieve evidence, and act across several steps with limited human prompting, can carry a meaningful part of that load. The paper is conceptual and design oriented rather than empirical. It reviews recent literature on quality assurance, transnational higher education, and language model based agents, then proposes a layered architecture called the Agentic Quality Assurance Architecture. The architecture separates evidence, standards, reasoning, action, and oversight into distinct layers so that machine work can be inspected and reversed. The paper maps specific agent roles onto the accreditation lifecycle, describes how a standards crosswalk can allow one body of evidence to serve several regulators at once, and sets out a five level maturity model for adoption. It also examines the risks in detail, including fabricated citations, prompt injection through uploaded documents, automation bias among reviewers, the temptation to optimise for the appearance of quality rather than quality itself, and the accountability gap that opens when no named person can explain how a claim entered a self study report. The article concludes that agentic workflows are best understood as instruments of evidence logistics and not as substitutes for professional judgement, and that any institution adopting them should treat human sign off, provenance tracking, and independent audit as non negotiable design requirements rather than optional extras. Keywords: agentic artificial intelligence, quality assurance, accreditation, transnational education, international branch campuses, compliance automation, governance, large language models 1. Introduction 1.1 The problem in plain terms Quality assurance in higher education used to be a local activity. A university wrote a self study, a panel of peers visited, the panel wrote a report, and a decision followed. The whole cycle took place inside one legal system, one language, and one set of expectations about what counted as good teaching. That description no longer fits a large and growing part of the sector. Universities now open branch campuses abroad, franchise programmes to partner colleges, deliver online degrees to students in dozens of jurisdictions, and enter joint and dual award arrangements that split responsibility for a single qualification between two institutions in two countries. Each of those arrangements brings its own regulator, its own reporting calendar, and its own definition of evidence. The result is that compliance work has multiplied faster than the staff who do it. The practical consequence is familiar to anyone who has worked in a quality office. Staff spend an enormous share of their time on tasks that are cognitively simple but administratively heavy. They search for a document that they know exists somewhere. They copy a table from one report into another with slightly different column headings. They check that a policy referenced in a self study is the current version and not the one superseded eighteen months ago. They rewrite the same paragraph about assessment moderation for four different agencies because each agency wants it phrased in its own vocabulary. None of this work is trivial, because an error in any of it can cost an institution its licence to operate. But almost none of it requires the professional expertise that quality staff actually possess. This article asks a direct question. Can agentic workflows take on the heavy but low judgement portion of that labour, and if so, under what conditions can they do it without corrupting the thing they are supposed to protect? 1.2 What has changed recently Two developments make the question worth asking now rather than five years ago. The first is technical. Language models can now be wrapped in scaffolding that lets them plan a sequence of actions, call external tools, read the results, and revise their approach. This is the shift from a system that answers a question to a system that pursues a goal. Research on reasoning and acting loops, on tool calling, on self correction through reflection, and on retrieval from external knowledge stores has produced a set of engineering patterns that are now widely reused (Yao et al., 2023; Schick et al., 2023; Shinn et al., 2023; Lewis et al., 2020). Surveys of language model based autonomous agents describe a common structure of profile, memory, planning, and action that can be adapted to almost any document heavy domain (Wang et al., 2024; Xi et al., 2025). The second is regulatory. The European Union has adopted a horizontal law on artificial intelligence that places obligations on deployers according to risk, and national and sectoral regulators are following with their own expectations (European Union, 2024). Frameworks for managing artificial intelligence risk are now available in a form that an institutional #governance committee can actually work with (National Institute of Standards and Technology, 2023). International guidance aimed specifically at education has been published (Miao and Holmes, 2023). This matters because #quality_assurance is a regulated activity. A university cannot deploy an opaque system into its accreditation process and hope that nobody notices. It has to be able to explain what the system does, who checked it, and what happens when it is wrong. 1.3 Scope and contribution This paper does not report an experiment or a survey. It is a conceptual and design study, and it says so plainly, because one of the most common failures in this area is the presentation of speculation as if it were finding. The contribution is threefold. First, the paper synthesises three literatures that rarely speak to each other, namely research on internal and external #quality_assurance, research on transnational and cross border higher education, and technical research on agentic systems. Second, it proposes an architecture that is specific enough to be implemented and criticised, rather than a diagram of boxes with arrows. The architecture separates the layers where machine autonomy is acceptable from the layers where it is not. Third, it takes the risks seriously and at length. A great deal of writing about artificial intelligence in higher education either promises transformation or predicts catastrophe. Neither posture helps a pro vice chancellor who has to decide whether to sign a purchase order. This paper tries to give that person something useful. 1.4 Structure Section 2 sets out the context of borderless provision and the compliance burden it creates. Section 3 reviews the relevant literature. Section 4 defines agentic workflows precisely and distinguishes them from earlier forms of automation. Section 5 presents the proposed architecture. Section 6 maps agent roles onto the accreditation lifecycle. Section 7 addresses the specific problem of cross border standards and the crosswalk method. Section 8 presents an illustrative and explicitly hypothetical scenario. Section 9 examines risks. Section 10 proposes a governance framework. Section 11 offers a maturity model and implementation roadmap. Section 12 sets out evaluation metrics and a research agenda. Section 13 states limitations. Section 14 concludes. 2. The Borderless University and the Compliance Burden 2.1 The shape of cross border provision Cross border higher education takes several distinct forms, and the differences matter for compliance because each form allocates responsibility differently. An international branch campus is a physical site in a host country, owned or co owned by a foreign institution, awarding the foreign institution qualification. The home institution retains academic responsibility, but the site sits inside the host country legal system and is usually subject to host country licensing. Reviews of two decades of branch campus development describe a sector that has grown, consolidated, and occasionally retreated, with closures as well as openings, and with survival often depending on the regulatory and financial conditions in the host state (Wilkins, 2021). A franchise or validation arrangement places delivery in the hands of a partner institution while the awarding institution retains responsibility for the standard of the award. Here the awarding body must assure itself of teaching it does not directly control, which is a harder evidential problem than assuring its own classrooms. Joint and dual awards split responsibility between two institutions, each answerable to its own regulator, with the added difficulty that the two regulators may hold incompatible views about credit, contact hours, or assessment. Fully online cross border delivery removes the physical site altogether. The student may be in a country where the provider has no legal presence at all, which raises questions about consumer protection, data transfer, and recognition of the qualification. Comparative work on branch campuses in Malaysia, Singapore, China, and South Korea shows how differently host states approach regulation, ranging from light touch registration to detailed control over curriculum, staffing, and quality processes (Hou et al., 2021). An institution operating in several of these states simultaneously is therefore not managing one compliance regime with local variations. It is managing several genuinely different regimes at once. 2.2 Why the burden compounds The burden created by this pattern is not simply additive. It compounds, for four reasons. Different vocabularies for similar concepts. Almost every #quality_assurance system in the world cares about whether students achieve the intended learning outcomes, whether staff are qualified, whether assessment is fair, and whether the institution improves in response to what it learns. But the terms differ. What one agency calls a programme learning outcome another calls a graduate attribute, a competency, a course objective, or an exit profile. A human expert reads across these terms without difficulty. A filing system does not, which is why the same evidence gets re filed and re described repeatedly. Different evidential expectations. One agency wants a narrative account with illustrative examples. Another wants a matrix that maps every outcome to every assessment task. A third wants numerical thresholds and will not accept prose in their place. The underlying facts are identical. The presentation is not, and presentation takes time. Different cycles. Institutional review runs on one clock, programme #accreditation on another, professional body recognition on a third, and host country licence renewal on a fourth. Staff therefore live in a state of permanent submission preparation, which is exhausting and which crowds out the reflective work that #quality_assurance is supposed to enable. Different equivalence claims. A branch campus must usually demonstrate that its provision is equivalent to the home campus, but equivalence is not sameness. A class size that is normal in one country may be considered excessive in another. Library provision that is adequate where physical collections dominate may be inadequate where they do not. Demonstrating equivalence therefore requires careful comparative argument built on comparable data, and comparable data is precisely what a federated multi campus institution usually lacks. 2.3 The cost of the burden Research on internal quality assurance has warned for some time that quality systems can drift into compliance ritual, generating documents that satisfy external requirements while doing little to improve teaching (Cardoso et al., 2019). Studies of how agencies and review panels actually use European standards show variation in interpretation even within a common framework, which further increases the effort required of institutions that must satisfy multiple readers (Manatos and Huisman, 2020). The cost is not only financial. It is also a cost in trust and morale. Academic staff who experience quality assurance mainly as a demand for paperwork tend to disengage from it. Once they disengage, the documents produced become less truthful, because they are assembled by administrators who were not in the room when the teaching decisions were made. The system then measures its own paperwork rather than the education it was built to protect. Any technology introduced into this setting can either deepen that pathology or relieve it, and which of those happens depends less on the technology than on the design choices made around it. 2.4 A note on internationalisation values It is worth stating plainly that cross border provision is not merely a business model. Scholarship on internationalisation stresses that it should serve academic and social purposes, and warns against the drift toward commercial logic that treats students as revenue and campuses as assets (de Wit and Altbach, 2021). Analyses of global higher education similarly emphasise that the global and the national coexist rather than replace one another, and that institutions operate simultaneously in several normative worlds (Marginson, 2022). A model for aligning educational quality with the student experience in international provision argues that quality cannot be assessed apart from what students actually experience in the host setting (Tsiligiris and Hill, 2021). These arguments are directly relevant to automation. If #quality_assurance exists to protect the interests of students, then any system that makes the reporting easier while making the student experience less visible has failed, no matter how efficient it is. 3. Literature Review 3.1 Artificial intelligence in higher education The volume of research on artificial intelligence in higher education has grown quickly, and the quality of that research is uneven. A meta systematic review covering the field found recurring weaknesses, including thin ethical analysis, limited collaboration between technologists and educators, and methodological looseness that makes many claims difficult to trust (Bond et al., 2024). A state of the field review reached broadly similar conclusions, noting concentration of research in a small number of application areas and a shortage of work on institutional and administrative uses (Crompton and Burke, 2023). That last point is important for this paper. Most published work concerns teaching, learning, assessment, and student support. Administrative and #governance applications, including #quality_assurance, are comparatively under examined, even though they may be where language models are most immediately useful and least pedagogically risky. Critical scholarship supplies a necessary corrective to enthusiasm. Analyses of the discourse around artificial intelligence in higher education show how promotional language can narrow the range of futures that seem possible, and how claims of inevitability serve particular interests (Bearman et al., 2023). Historical work reminds us that educational technology has repeatedly promised to solve problems that turned out to be institutional rather than technical (Williamson and Eynon, 2020; Selwyn, 2022). Critical narratives of technology in higher education warn against confusing the automation of a process with its improvement (Popenici, 2023), and speculative approaches to digital futures argue for treating technological futures as things to be argued about rather than predicted (Ross, 2023). Work on #ethics in educational artificial intelligence has begun to converge on a set of principles, including transparency, fairness, accountability, privacy, and human oversight (Nguyen et al., 2023; Holmes and Tuomi, 2022). Frameworks aimed at university #policy makers propose structured approaches to governing use across teaching, research, and administration (Chan, 2023). Reviews of assessment policy in the generative era show institutions struggling to write rules that are both enforceable and educationally coherent (Luo, 2024). Work on assessment security offers a useful transferable insight, namely that the integrity of a process depends on the design of the process and not only on the detection of misconduct within it (Dawson, 2021). 3.2 Quality assurance and accreditation research The research base for #quality_assurance is mature but has not yet absorbed agentic technologies. The European Standards and Guidelines remain the most influential shared reference in the region and are built around the idea that the institution owns responsibility for quality and that external review verifies rather than replaces internal work (European Association for Quality Assurance in Higher Education, 2015). Studies of internal #quality_assurance repeatedly find tension between a #quality_culture, in which staff genuinely reflect and improve, and a #compliance culture, in which staff produce documents to satisfy inspectors (Cardoso et al., 2019). This tension defines the central design risk of the present paper. A system that produces excellent documents with less human effort could, if badly designed, remove the last remaining reason for academics to reflect on their practice. The reflection was often a by product of having to write the report. Take away the writing and you may take away the thinking. Any serious proposal for automation in this domain must confront that possibility rather than wave it away. 3.3 Language model based agents The technical literature provides the building blocks. Chain of thought prompting showed that intermediate reasoning steps improve performance on multi step problems (Wei et al., 2022). Reasoning and acting frameworks interleave thought with tool calls so that a model can gather information rather than guess (Yao et al., 2023). Tool learning demonstrated that models can be taught to call external functions when they need them (Schick et al., 2023). Reflection mechanisms let an agent criticise its own output and try again (Shinn et al., 2023). #retrieval_augmented_generation grounds outputs in retrieved documents rather than parametric memory, which is the single most important technique for a domain where every claim must be traceable to a source (Lewis et al., 2020). Surveys of autonomous agents describe architectures built from profile, memory, planning, and action modules, and catalogue applications across software engineering, science, and social simulation (Wang et al., 2024; Xi et al., 2025). Work on generative agents demonstrated multi agent environments with persistent memory and emergent coordination (Park et al., 2023). Foundation model analyses set out both the capability and the systemic risk of building many downstream applications on a small number of upstream models (Bommasani et al., 2021). The technical literature is also candid about limits. Hallucination remains an unsolved problem and is intrinsic to the way these systems generate text (Ji et al., 2023). Assurance of alignment and safety in large models faces foundational obstacles that are not close to resolution (Anwar et al., 2024). Indirect #prompt_injection allows an attacker to hide instructions inside a document that the model will later read, which turns any pipeline that ingests untrusted files into an attack surface (Greshake et al., 2023). Evaluation of agents is frequently unrigorous, with benchmarks that reward cost blind performance and results that do not survive replication (Kapoor et al., 2024). Taxonomies of language model risk cover misinformation, malicious use, and harms of representation, all of which have plausible analogues in a #quality_assurance setting (Weidinger et al., 2022). 3.4 Accountability, audit, and explanation A fourth literature supplies the #governance vocabulary. Internal algorithmic auditing frameworks describe how an organisation can check its own systems before deployment rather than after harm (Raji et al., 2020). Work on algorithmic supply chains shows that responsibility becomes diffuse when a system is assembled from components supplied by many parties, which is exactly the situation of a university buying a product built on a third party model (Cobbe et al., 2023). Documentation practices such as datasheets create a record of what data a system was built on and what it was intended for (Gebru et al., 2021). Human centred design principles argue for high automation and high human control together rather than a trade off between them (Shneiderman, 2022). Work on explanation surveys what current techniques can and cannot deliver (Ali et al., 2023; Longo et al., 2024). Legal analyses of the European framework explain the obligations that fall on deployers of higher risk systems (Veale and Zuiderveen Borgesius, 2021; Novelli et al., 2024). Broader philosophical treatments situate these questions within an account of what responsible use looks like (Floridi, 2023). Learning analytics research offers hard won lessons about student data, consent, and the limits of inference, and those lessons transfer directly to any #quality_assurance system that touches individual student records (Tzimas and Demetriadis, 2021). 3.5 The gap Putting these literatures together reveals a clear gap. #quality_assurance scholarship has not engaged with agentic systems. Agent research has not engaged with the peculiar evidential demands of #accreditation, where a claim must be attributable, dated, versioned, and defensible in front of a panel. Critical scholarship warns about hype but rarely offers an alternative design. This paper attempts to occupy that space. 4. What Agentic Workflows Are, and What They Are Not 4.1 A working definition For the purposes of this paper, an agentic workflow is a computational process in which one or more language model based components pursue a specified goal by planning a sequence of steps, invoking external tools, reading the results, updating their approach, and producing an output, with the whole process subject to defined limits and to human review at specified points. Four properties distinguish this from earlier automation. Goal direction rather than step specification. A traditional script is told what to do. An agent is told what to achieve and works out the steps. This is powerful and also dangerous, because a system that chooses its own steps can choose badly. Tool use. The agent can query a database, read a file, call an application programming interface, or run a search. Its knowledge is therefore not limited to what was in its training data, and its claims can be grounded in current institutional records (Schick et al., 2023; Lewis et al., 2020). Memory. The agent can retain state across steps and across sessions, which allows it to work on a task that takes weeks rather than seconds. Multi step correction. The agent can inspect its own draft, find a flaw, and revise (Shinn et al., 2023). 4.2 What agentic workflows are not They are not oracles. They do not know whether a programme is good. They cannot form the professional judgement that a peer reviewer forms after sitting in a classroom and talking to students. They are not deterministic. The same input may produce different outputs on different runs, which is uncomfortable in a domain that prizes consistency and which has direct implications for auditability. They are not neutral. The #standards they apply, the #evidence they privilege, and the language they produce all reflect choices made by the people who built and configured them. They are not accountable. A regulator cannot sanction a model. Responsibility must therefore rest, at every point, with a named human being (Cobbe et al., 2023). 4.3 The right framing: evidence logistics The most useful way to think about #agentic_workflows in this domain is as a system of evidence logistics. Logistics moves goods to where they are needed, in the right condition, at the right time, with a record of where they came from. It does not decide what the goods are worth. Applied to #quality_assurance, the agentic system finds relevant #evidence, checks that it is current, describes it accurately, arranges it against the relevant standard, flags what is missing, and hands it to a human who decides what it means. The judgement stays with the human. The fetching, checking, formatting, and cross referencing move to the machine. This framing is deliberately modest, and the modesty is the point. Most of the failure modes discussed later in this paper arise when institutions ask the system to make judgements it is not capable of making and cannot be held responsible for. 5. The Agentic Quality Assurance Architecture 5.1 Overview The proposed architecture has six layers. Each layer has a defined function, a defined level of permitted machine autonomy, and a defined human role. Separating the layers matters because it allows an institution to adopt part of the system without adopting all of it, and because it makes the boundary between machine work and human judgement explicit rather than implicit. Layer 1: Evidence. The institutional record. Layer 2: Standards. The machine readable representation of what regulators require. Layer 3: Retrieval and grounding. The mechanism that connects claims to sources. Layer 4: Reasoning and drafting. The agents that analyse, compare, and write. Layer 5: Action and workflow. The mechanism that moves work between people and systems. Layer 6: Oversight. The human and organisational controls that sit above everything else. 5.2 Layer 1: The evidence layer Nothing in this architecture works without a disciplined evidence layer, and this is where most institutions will find their real problem. The obstacle to automating #accreditation is rarely the reasoning. It is that the underlying records are scattered, inconsistently named, undated, duplicated, and held in systems that do not talk to each other. A branch campus keeps its own spreadsheets. A partner college holds the moderation records. The student information system in one country uses a different grading scale from the system in another. The evidence layer therefore requires several things. A canonical repository, or at least a federated index that can reach every source system without requiring documents to be copied. Mandatory #metadata for every artefact, including owner, campus, date of creation, date of approval, version, superseded version, applicable programmes, and retention class. Documentation practices from machine learning offer a good model here, since a datasheet is essentially a structured statement of what an artefact is and what it may legitimately be used for (Gebru et al., 2021). #provenance that survives transformation. If a number appears in a self study, it must be possible to trace it back through every intermediate calculation to the source record. This is the single most important technical requirement in the whole architecture, because a claim without #provenance is worthless in an #accreditation setting no matter how eloquently it is written. A clear data classification scheme distinguishing public documents, internal documents, commercially sensitive documents, and personal data, because the rules governing what an agent may read differ sharply across those categories. 5.3 Layer 2: The standards layer The second layer turns regulatory requirements into structures a machine can work with. Each standard is decomposed into atomic requirements. Each requirement carries an identifier, the issuing body, the version, the effective date, the type of #evidence normally accepted, and the level at which it applies, whether institutional, programme, or course. Once standards are represented this way, they can be related to each other. Requirement A from Agency One can be linked to Requirement B from Agency Two with a stated relationship, such as equivalent, broader, narrower, overlapping, or unrelated. This is the #crosswalk, and it is discussed at length in Section 7. It is the mechanism that allows one body of #evidence to serve several regulators. The standards layer must be maintained by human experts. A model may propose a mapping. Only a person who understands both regulatory systems may approve it. A wrong mapping is worse than no mapping, because it produces confident submissions that miss the requirement entirely. 5.4 Layer 3: Retrieval and grounding The retrieval layer is what prevents invention. Rather than asking a model what the institution does about assessment moderation, the system retrieves the moderation #policy, the minutes of the board that approved it, and the sample moderation records, and asks the model to describe what those documents say. Every sentence in a draft must carry a pointer to the source or sources that support it (Lewis et al., 2020). Three engineering commitments follow. Retrieval must be scoped by permission, so that an agent working on one campus cannot silently read another campus confidential material. Retrieval must be recency aware, so that a superseded #policy is never presented as current. This requires the #metadata discipline described above. Any sentence in a generated draft that cannot be traced to a retrieved source must be automatically flagged and shown to the human reviewer as unsupported. Unsupported does not mean false. It means unverified, and the distinction should be visible on the face of the document. 5.5 Layer 4: Reasoning and drafting agents This layer contains the specialised agents described in Section 6. The important architectural point is that they are separate. A single general purpose agent asked to produce a self study will produce something plausible and unverifiable. A set of narrow agents, each with a defined input, a defined output, and a defined check, produces work that can be inspected step by step. Separation also enables adversarial checking. The agent that drafts a claim should not be the agent that verifies it. A verification agent given only the draft and the source documents, and asked whether each sentence is supported, will catch errors that the drafting agent, which is motivated to produce fluent text, will not. 5.6 Layer 5: Action and workflow The action layer determines what the system may actually do in the world. This is where most institutions should be conservative. Reading is low risk. Drafting is low risk provided the draft is clearly marked as a draft. Notifying a person that a deadline approaches is low risk. Writing to the student record system is high risk. Sending a document to a regulator is high risk. Changing a #policy document is high risk. These actions should require explicit human authorisation, and the authorisation should be recorded with the identity of the person, the time, and the version of the artefact approved. The general principle is that the system should be permitted to prepare anything and to commit almost nothing. #human_in_the_loop is not a slogan here. It is a specific claim about which functions in the action layer carry an approval gate. 5.7 Layer 6: Oversight The oversight layer is described in Section 10. It comprises the #audit_trail, the model documentation, the assignment of accountable owners, the escalation path when the system errs, the periodic evaluation, and the independent review. It is listed last but should be built first, because retrofitting oversight onto a deployed system almost never succeeds (Raji et al., 2020). 6. Mapping Agents to the Accreditation Lifecycle 6.1 The lifecycle A typical external review cycle has five phases: continuous internal monitoring, self study preparation, submission, the review visit, and follow up on conditions and recommendations. Different agent roles suit different phases. 6.2 The evidence harvesting agent Function: locate and assemble the #evidence relevant to a given requirement. Given the requirement that an institution must demonstrate that academic staff are appropriately qualified for the level of the programme, the harvesting agent queries the human resources system for staff records, the programme database for teaching allocations, and the document repository for the recruitment and probation policies. It returns a set of artefacts with #metadata and a note of what it could not find. The most valuable output of this agent is not what it finds. It is what it cannot find. A reliable list of gaps, produced twelve months before a submission deadline rather than three weeks before, changes the character of the preparation process entirely. 6.3 The standards mapping agent Function: propose links between requirements from different bodies, and between requirements and existing #evidence. The mapping agent is a proposal engine, never an authority. It suggests that a given host country requirement about industry engagement is partially satisfied by evidence already collected for a professional body requirement about work integrated learning, and it explains the reasoning. A human confirms, rejects, or refines. The confirmed mapping enters the standards layer and is reused thereafter. 6.4 The gap analysis agent Function: compare current #evidence against requirements and report shortfalls with severity ratings. The severity rating must be conservative. It is far better to over report possible gaps than to under report them, because a false alarm costs staff time while a missed gap costs the institution its #accreditation. The agent should also distinguish three kinds of gap, namely evidence that does not exist, evidence that exists but has not been found, and evidence that exists and has been found but does not actually support the claim. These are different problems with different remedies. 6.5 The consistency and equivalence agent Function: compare data across campuses and flag divergences that would trouble a reviewer. If the home campus reports a first attempt pass rate of a certain level and a branch reports a substantially different one for the same module, the agent flags it. It does not explain it. Explaining it requires knowing that one campus admitted a cohort with different entry qualifications, or that a new lecturer was appointed mid semester, or that a national holiday disrupted the teaching schedule. Only a human knows these things. This agent is nonetheless one of the most valuable in the whole architecture, because inconsistency across campuses is precisely what external reviewers look for, and institutions routinely discover such inconsistencies at the worst possible moment. 6.6 The drafting agent Function: produce a first draft of a section of the self study, grounded in retrieved #evidence, in the register and structure expected by the specific agency. Two constraints are essential. Every sentence carries citations to source artefacts. The draft is watermarked and versioned as machine generated until a named human accepts it, at which point the record shows who accepted it and when. The drafting agent is genuinely useful because agencies want the same substance in different shapes. Producing four differently shaped versions of the same truthful account is exactly the kind of task where machine assistance saves real time without touching professional judgement. 6.7 The verification agent Function: check the draft against the sources. The verification agent receives the draft and the cited artefacts, and for each sentence returns a judgement of supported, partially supported, unsupported, or contradicted. It should be built to be sceptical, and it should be evaluated on its ability to catch deliberately introduced errors. An institution that does not periodically test its verification agent with planted mistakes has no basis for trusting it. 6.8 The monitoring agent Function: watch continuously rather than episodically. This is where #agentic_workflows could change #quality_assurance most fundamentally. Today, most institutions look closely at their quality data when a review approaches. A monitoring agent that runs weekly across all campuses, comparing indicators against thresholds and against each other, converts #quality_assurance from an episodic performance into ongoing management information. #continuous_monitoring of this kind is closer to the original intention of internal quality systems than the submission driven cycle most institutions actually run. The risk is equally clear. Continuous monitoring can become continuous surveillance of academic staff. The design must therefore aggregate at the level of programme and cohort rather than individual teacher wherever the purpose can be served by aggregate data, and any move to individual level analysis should require explicit #governance approval and staff consultation. 6.9 The submission and correspondence agent Function: assemble the final package in the required format, check completeness against the agency checklist, track deadlines, and prepare responses to agency queries. This agent should have no authority to send anything. It prepares. A human sends. 6.10 The orchestrator Above the specialised agents sits an #orchestration component that decomposes a request into tasks, assigns them, sequences them, handles failure, and escalates to a human when a task cannot be completed or when confidence falls below a threshold. The orchestrator is the part of the system most likely to fail in surprising ways, because it is the part with the most freedom. Its permitted action space should be tightly bounded, its decisions logged, and its escalation thresholds set conservatively at first and relaxed only in the light of evidence. Research on #multi_agent_systems shows both the coordination benefits and the difficulty of predicting emergent behaviour (Park et al., 2023; Xi et al., 2025). 7. Cross Border Compliance and the Crosswalk Method 7.1 The core insight The central insight of this paper is that the evidence a university holds is largely common across regulators, while the descriptions demanded by regulators are largely different. If the institution maintains one authoritative body of evidence and a set of mappings from that evidence to each regulator requirement, then producing a submission becomes a matter of projection rather than reconstruction. This is what a #crosswalk does. It is a many to many mapping between the requirements of different bodies, mediated by a shared internal representation. 7.2 Building the crosswalk The recommended sequence is as follows. Decompose each applicable standard into atomic requirements with stable identifiers. Define an internal evidence taxonomy that is independent of any single agency, organised around the things the institution actually has, such as approved policies, committee minutes, staff records, curriculum documents, assessment specifications, moderation records, external examiner reports, student feedback data, progression and completion statistics, employment outcomes, learning resources inventories, and complaints records. Map each atomic requirement to the evidence types that can satisfy it, with a note on the form of presentation each agency expects. Have the mapping proposed by an agent and approved by a human expert who knows both regulatory systems. Version the mapping. When an agency revises its standards, the mapping must be revisited, and every submission built on the old mapping must be identifiable so that its status can be reassessed. 7.3 What the crosswalk cannot do The #crosswalk handles the mechanical part of multi jurisdiction #compliance. It does not handle the substantive part, and it is important to be honest about this. Some requirements are genuinely incompatible. A host country rule requiring a fixed number of contact hours may sit awkwardly with a home country framework built on learning outcomes and notional study hours. No mapping resolves that. It requires negotiation, or a design decision, or in some cases a decision not to operate in that jurisdiction. Some requirements are culturally embedded in ways that resist translation. Expectations about the role of external examiners, the meaning of academic freedom, the acceptable degree of student influence on curriculum, and the relationship between the institution and the state differ across systems in ways that are not merely terminological. And some equivalence claims require argument rather than mapping. Demonstrating that a branch campus library provides equivalent access to scholarship, when its physical collection is small but its digital access is extensive, is an argument about what provision means. A machine can assemble the figures. Only a human can make the case. 7.4 Data protection across borders Cross border #quality_assurance necessarily involves cross border data. Student records, staff records, and feedback data sit in different jurisdictions with different rules about processing, transfer, and retention. Several design rules follow. Prefer aggregation. Most #accreditation requirements can be satisfied by aggregate statistics rather than individual records. Where aggregate data suffices, individual data should never be exposed to the agentic layer. Keep processing local where the law requires it. If a jurisdiction restricts transfer of personal data, the retrieval and analysis for that campus should occur within that jurisdiction, and only aggregate results should cross the border. Log every access to personal data, with purpose, actor, and time. This is required in several regulatory regimes and is good practice everywhere. Treat the model provider as part of the processing chain. If prompts containing institutional data are sent to an external model provider, that provider is processing the data, and the contractual and legal consequences must be addressed before deployment, not after (Cobbe et al., 2023). The lessons from learning analytics apply directly here. Institutions have repeatedly discovered that what is technically possible with student data is not always ethically or legally acceptable, and that consent obtained for one purpose does not extend indefinitely to others (Tzimas and Demetriadis, 2021). 8. An Illustrative Scenario The following scenario is hypothetical. It is offered to make the architecture concrete and contains no empirical claims. No institution, data, or result described here is real. Consider a university headquartered in one country with branch campuses in three others and franchise arrangements with two partner colleges. It holds institutional #accreditation at home, licences in each host country, and programme #accreditation from two professional bodies for its engineering and business provision. Eighteen months before the next institutional review. The monitoring agent has been running for a year. It has produced a standing report of divergences across campuses. Three items stand out. Assessment moderation records for one branch are incomplete for two semesters. Staff qualification records for the franchise partners are held by the partners and have not been verified for a year. Student feedback response rates at one campus have fallen to a level that undermines any claim based on that feedback. None of these are catastrophic. All of them would be damaging if discovered by a review panel. Discovered eighteen months in advance, they are simply work. Twelve months out. The evidence harvesting agent runs the full requirement set for the institutional review, the two host country licence renewals falling in the same period, and one professional body reaccreditation. The gap analysis agent reports the results by severity. The #crosswalk shows that a substantial proportion of the required evidence is shared across the four processes, which allows the quality office to plan one evidence gathering exercise rather than four. Six months out. The drafting agent produces first drafts of each self study section, grounded in retrieved evidence with citations. Roughly speaking, some sentences are fine, some need reworking, and some are flagged as unsupported. The unsupported flags are the most useful output, because they identify where the institution is asserting something it cannot demonstrate. Human authors, who are academic and professional staff with real knowledge of the provision, rewrite, correct, and add the judgement that the draft lacks. Three months out. The verification agent runs against the human edited drafts. It finds some claims that no longer match their sources after editing, which is a common and easily missed error. The consistency agent runs a final comparison across campuses. The submission agent assembles the packages and checks each against the agency checklist. The visit. During the review, the panel asks for evidence that was not in the submission. The harvesting agent retrieves candidates within minutes. A human checks them and provides them. The panel receives a faster and more complete response than would otherwise be possible. Afterwards. Conditions and recommendations are entered into the system as new requirements with deadlines. The monitoring agent tracks them. What has changed in this scenario is not the standard of judgement. Humans still wrote the arguments, made the equivalence claims, decided what the data meant, and signed everything. What changed is that they spent their time on those things instead of on searching, formatting, cross referencing, and chasing. 9. Risks and Failure Modes This section is deliberately the longest analytical section in the paper, because the risks are where institutional decisions are actually made. 9.1 Fabrication Language models generate plausible text, and plausibility is not truth. Hallucination is a well documented and structurally persistent property of these systems (Ji et al., 2023). In an #accreditation context, a fabricated citation, a policy that does not exist, a committee that never met, or a statistic that was never calculated is not an embarrassment. It is potentially a false statement to a regulator. Mitigations: strict grounding through retrieval; mandatory citation of every claim; automatic flagging of unsupported sentences; an independent verification agent; human sign off at named points; periodic adversarial testing with planted errors. None of these mitigations eliminates the risk. They reduce it and, more importantly, they make it detectable. An institution should assume that fabrication will occur and design so that it is caught before it leaves the building. 9.2 Prompt injection and evidence poisoning Any system that reads documents can be attacked through documents. Instructions hidden in a file, an email, or a form submission can redirect an agent that reads them (Greshake et al., 2023). In a #quality_assurance pipeline, the ingested documents include material from partner colleges, from student submissions, and from external parties. This is a real and underappreciated exposure. Mitigations: treat all ingested content as untrusted data rather than instruction; sanitise and isolate; give agents least privilege so that a compromised agent can read little and write nothing; log all tool calls; require human authorisation for any external action; monitor for anomalous agent behaviour. There is also a subtler variant, which is evidence poisoning by insiders. If staff learn that a monitoring agent flags certain patterns, they may learn to produce records that avoid the flags. This is not a security problem in the technical sense. It is a #governance problem, and it is discussed below. 9.3 Automation bias When a system produces confident, fluent, well formatted output, humans tend to accept it. This is #automation_bias, and it is the most likely way that a well designed system fails in practice, because it defeats the human oversight on which every other control depends. A reviewer who is asked to check two hundred pages of machine drafted self study, under time pressure, will not check it properly. The oversight becomes a formality, the sign off becomes a ritual, and the institution ends up with the appearance of human control and the reality of machine authorship. Mitigations: limit the volume of machine output that any one person is asked to check; require reviewers to record the specific checks they performed rather than simply confirming approval; display confidence and support status prominently on the face of the document; deliberately surface disagreement between agents rather than resolving it silently; rotate reviewers; sample audit human approvals to see whether errors are being caught. Human centred design work argues that high automation and high human control can coexist, but only when the interface is designed to make control real rather than nominal (Shneiderman, 2022). 9.4 Goodhart effects and the corruption of quality If an institution optimises for the metrics that #accreditation bodies measure, and if a machine can optimise those metrics faster and more thoroughly than a human, the institution may become extremely good at appearing to be good. This is the deepest risk in the paper and it has no technical fix. It is a matter of institutional intent. A university that uses #agentic_workflows to find and fix real weaknesses will improve. A university that uses them to find and conceal real weaknesses will decline while its reports improve. The technology serves both purposes equally well. Partial mitigations exist. Keep #quality_assurance connected to actual student experience through direct evidence, including student voice, that is harder to manufacture (Tsiligiris and Hill, 2021). Ensure that the people who use the system are the people who teach, not only the people who report. Have external reviewers examine the process by which submissions are produced, not only the submissions themselves. Retain unannounced and qualitative elements in review so that the paper record cannot be the only thing examined. 9.5 Erosion of quality culture Writing a self study is painful, and some of the pain is productive. Staff are forced to articulate what they do, notice that they cannot justify parts of it, and change them. If a machine writes the first draft, that reflection may not happen. This is a serious concern given existing evidence that internal #quality_assurance already tends toward ritual (Cardoso et al., 2019). Automation could accelerate that drift. The design response is to automate the searching and formatting while deliberately preserving the reflection. The machine should present staff with the questions it cannot answer, the inconsistencies it cannot explain, and the claims it cannot support, and require them to respond. Used this way, the system creates reflective work rather than removing it. Whether institutions will use it that way, when the alternative is faster and cheaper, is an open question and one that should be studied empirically. 9.6 The accountability gap If a self study contains an error, who is responsible? The staff member who approved it may say the system produced it. The system vendor may point to terms of service disclaiming responsibility for outputs. The model provider may point to its own disclaimers. Responsibility diffuses across the supply chain until nobody holds it (Cobbe et al., 2023). Regulators will not accept this, and neither should institutions. Every artefact should have a named accountable human owner. The #audit_trail should record who approved what, when, and on what basis. Contracts with vendors should address responsibility explicitly. The institution should assume that it, and not its supplier, will be answerable to the regulator. 9.7 Inequity between campuses The home campus usually has better systems, better records, and better staff capacity than a young branch. If the agentic system depends on high quality structured records, it will work well at the home campus and badly at the branch. The branch will then appear worse in comparative reporting, not because its education is worse but because its record keeping is. This is an important and easily overlooked injustice, and it is a reason to invest in the evidence layer at the periphery before deploying clever reasoning at the centre. #equity in this sense is a design requirement, not an afterthought. 9.8 Vendor lock in and model dependency Building institutional #quality_assurance on a single external model creates dependency. Models are deprecated, prices change, terms change, and behaviour changes between versions in ways that can silently alter outputs. The foundation model literature warns precisely about this concentration of dependency (Bommasani et al., 2021). Mitigations: keep the evidence layer, standards layer, and #crosswalk in open, portable formats owned by the institution; treat the model as a replaceable component; test outputs when a model version changes; avoid designs in which institutional knowledge exists only inside a vendor system. 9.9 Evaluation weakness Institutions will be tempted to judge these systems by whether staff like them and whether submissions go through. Neither is a valid measure of accuracy. Agent evaluation is known to be weak across the field, with benchmarks that reward the wrong things and results that do not replicate (Kapoor et al., 2024). An institution deploying this technology in a regulated process must build its own evaluation, and Section 12 proposes what it should measure. 10. A Governance Framework 10.1 Principles Six principles should govern deployment. Human accountability is absolute. Every output that leaves the institution has a named human owner who is answerable for it. Grounding is mandatory. No claim without a source. No source without #provenance. Autonomy is graduated. Machine autonomy is permitted in proportion to the reversibility and the stakes of the action. Everything is logged. Prompts, retrievals, tool calls, outputs, approvals, and overrides. Transparency to those affected. Staff and students should know that these systems are in use, what they do, and what they do not do. Independent review. The system is audited by people who did not build it (Raji et al., 2020). 10.2 Structures A university deploying #agentic_workflows in #quality_assurance should establish, at minimum, a #governance committee with academic, professional, legal, data protection, and student membership; a documented register of every agentic use case with its risk classification; a designated owner for each use case; an incident reporting route for errors; and a scheduled review cycle. This maps reasonably well onto established artificial intelligence risk management practice, which organises activity around governing, mapping, measuring, and managing risk (National Institute of Standards and Technology, 2023). 10.3 Regulatory positioning Institutions operating in or serving the European Union must consider where their systems fall under the European framework, which imposes obligations that scale with risk and which places specific duties on deployers as well as providers (European Union, 2024; Veale and Zuiderveen Borgesius, 2021; Novelli et al., 2024). A system that drafts a self study is not obviously in the highest risk category, but a system that touches decisions about individual students or staff may be, and the boundary requires legal advice rather than assumption. Beyond formal law, sector guidance on generative artificial intelligence in education sets expectations about human agency, transparency, and equity that regulators are likely to invoke even where they are not binding (Miao and Holmes, 2023). Institutions should also expect #accreditation bodies themselves to begin asking how submissions were prepared. Publishers have already moved to require disclosure of artificial intelligence use in manuscripts, and an analogous expectation in #accreditation is a reasonable prediction (Perkins and Roe, 2024). 10.4 Disclosure Institutions should decide, in advance and as a matter of #policy, what they will tell #accreditation bodies about their use of these systems. The safest position, and in the view of this paper the correct one, is full disclosure. A submission that states clearly that drafting was machine assisted, that every claim is grounded in cited evidence, that verification was performed, and that named humans approved the content is more credible than one that conceals its method and is later found out. Concealment carries an obvious and severe downside risk. Disclosure, accompanied by a description of controls, may even become a marker of institutional maturity. 11. A Maturity Model for Adoption Institutions should not attempt to build the full architecture at once. The following five levels describe a plausible progression. Level 0: Fragmented. Evidence is scattered across systems and people. No shared taxonomy. Submissions are assembled by heroic individual effort under deadline pressure. Most institutions are here or close to it. No agentic technology should be deployed at this level, because there is nothing reliable for it to stand on. Level 1: Organised. A single evidence index exists. Artefacts carry #metadata. Versions are controlled. Standards are documented. This level requires no artificial intelligence at all and delivers substantial benefit on its own. It is also the precondition for everything that follows, and institutions that skip it will fail at Level 2. Level 2: Assisted retrieval. Agents help staff find evidence and answer questions about the institutional record, with citations. Low risk, immediately useful, and an excellent way for staff to build calibrated trust in the technology by seeing where it succeeds and where it fails. Level 3: Assisted analysis and drafting. Gap analysis, consistency checking across campuses, and grounded first drafts, all with mandatory human review. This is where meaningful workload reduction begins. It is also where #automation_bias becomes a live risk and where oversight design must be strong. Level 4: Orchestrated workflows. Multi agent #orchestration across the full lifecycle, with #continuous_monitoring, human approval gates at defined points, and a full #audit_trail. Substantial benefit and substantial governance requirement. Level 5: Continuous assurance. Quality assurance ceases to be episodic. The institution knows its own position against every applicable standard at any moment. Submissions become projections of a continuously maintained state rather than construction projects. This is the theoretical destination. No institution is there, and reaching it would require solving organisational problems that are considerably harder than the technical ones. The most important message of this model is that the early levels deliver most of the value and carry almost none of the risk. An institution that reaches Level 2 with disciplined evidence management has already improved its position enormously. The temptation to jump to Level 4 with poor underlying records should be resisted. 12. Evaluation and a Research Agenda 12.1 What to measure An institution deploying these systems should measure, at minimum, the following. Grounding accuracy. The proportion of generated claims that are fully supported by their cited sources, assessed by independent human review of a random sample. Fabrication rate. The proportion of outputs containing at least one unsupported factual assertion presented as fact. Gap detection sensitivity and specificity. How many real gaps the system finds, and how many false alarms it raises, measured against expert review of the same material. Human catch rate. Of errors deliberately planted in machine output, what proportion do human reviewers detect? This is the single most informative measure of whether oversight is real, and it should be run regularly and unannounced. Time and cost. Staff hours per submission, before and after, honestly accounted, including the time spent checking machine output and maintaining the system. Cost blind reporting of efficiency gains is a known weakness in agent evaluation (Kapoor et al., 2024). Equity across campuses. Whether the system performs comparably at branches and partners as at the home campus. Effect on quality culture. Whether staff engagement with reflection and improvement rises or falls. This requires qualitative work and cannot be inferred from usage statistics. 12.2 A research agenda The following questions are open and would repay serious empirical study. Does agentic assistance improve the accuracy of self studies, or only their fluency? Do review panels evaluate machine assisted submissions differently when they know how they were produced? Does #continuous_monitoring change academic behaviour, and in which direction? Can a #crosswalk between major #accreditation frameworks be constructed with sufficient reliability to be reused across institutions, and who should own and maintain it? What happens to the reflective function of the self study when the drafting burden is removed? How do staff at branch campuses experience a system built and configured at the home campus, and does it centralise power in ways that undermine local academic autonomy? Under what conditions does #automation_bias defeat human oversight in this specific domain, and what interface designs reduce it? These questions require collaboration between quality professionals, education researchers, and computer scientists. The meta reviews cited earlier suggest such collaboration is exactly what the field has been missing (Bond et al., 2024). 13. Limitations This paper is conceptual. The architecture has not been implemented, and the illustrative scenario is hypothetical. Claims about workload reduction are plausible but untested, and the reader should treat them as hypotheses rather than findings. The paper reflects assumptions drawn mainly from European, North American, and Asia Pacific #quality_assurance systems, and those assumptions may not transfer to systems with different regulatory philosophies. The technical capabilities described are moving quickly, and specific engineering claims will date. The architectural principles, particularly the insistence on grounding, #provenance, graduated autonomy, and human accountability, are likely to prove more durable than any particular implementation. Finally, the paper takes an essentially reformist position. It assumes that #accreditation is worth doing and that making it less wasteful is worthwhile. Readers who believe that the entire apparatus of external #quality_assurance is a bureaucratic imposition that damages universities will find the paper answers the wrong question, and that critique deserves a serious response that this paper does not attempt to give. 14. Conclusion Universities that operate across borders face a #compliance burden that is growing faster than their capacity to absorb it, and much of that burden consists of work that is administratively heavy but intellectually light. #agentic_workflows are well suited to precisely that kind of work. They can find evidence, check that it is current, describe it accurately, arrange it against the requirements of several different regulators, and tell a human what is missing. What they cannot do is form professional judgement, and they cannot be held responsible. Those two facts should determine the entire design. The machine handles logistics. The human holds judgement and accountability. The architecture proposed here, with its separation of evidence, standards, retrieval, reasoning, action, and oversight, is an attempt to make that division of labour structural rather than aspirational. The risks are real and they are not primarily technical. Fabrication can be caught with grounding and verification. Injection can be contained with least privilege. The harder risks are institutional. #automation_bias hollows out oversight until human approval means nothing. The pursuit of measurable quality can displace actual quality. The removal of the drafting burden can remove the reflection that made the drafting worthwhile. None of these are solved by better models. They are solved, if at all, by better institutions. The most useful advice this paper can offer is therefore also the least exciting. Fix the evidence layer first. Build the #audit_trail before the agents. Keep humans accountable by name. Disclose what you are doing. Measure whether your oversight actually catches errors, and assume it does not until you have shown that it does. Automate the searching, not the thinking. If universities do that, #agentic_workflows could give quality offices back the thing they have least of, which is time to think about whether the education they are assuring is any good. If they do not, these systems will simply help them produce more convincing accounts of provision nobody has genuinely examined. The technology is indifferent between those outcomes. The institutions are not, or at least they should not be. References Ali, S., Abuhmed, T., El-Sappagh, S., Muhammad, K., Alonso-Moral, J. M., Confalonieri, R., Guidotti, R., Del Ser, J., Diaz-Rodriguez, N., and Herrera, F. (2023). Explainable artificial intelligence: what we know and what is left to attain trustworthy artificial intelligence. Information Fusion, 99, 101805. Anwar, U., Saparov, A., Rando, J., Paleka, D., Turpin, M., Hase, P., Lubana, E. S., Jenner, E., Casper, S., Sourbut, O., et al. (2024). Foundational challenges in assuring alignment and safety of large language models. Transactions on Machine Learning Research. Bearman, M., Ryan, J., and Ajjawi, R. (2023). Discourses of artificial intelligence in higher education: a critical literature review. Higher Education, 86(2), 369-385. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021). On the opportunities and risks of foundation models. Center for Research on Foundation Models, Stanford Institute for Human-Centered Artificial Intelligence, Stanford University. Bond, M., Khosravi, H., De Laat, M., Bergdahl, N., Negrea, V., Oxley, E., Pham, P., Chong, S. W., and Siemens, G. (2024). A meta systematic review of artificial intelligence in higher education: a call for increased ethics, collaboration, and rigour. International Journal of Educational Technology in Higher Education, 21(1), 4. Cardoso, S., Rosa, M. J., Videira, P., and Amaral, A. (2019). Internal quality assurance: a new culture or added bureaucracy? Assessment and Evaluation in Higher Education, 44(2), 249-262. Chan, C. K. Y. (2023). A comprehensive AI policy education framework for university teaching and learning. International Journal of Educational Technology in Higher Education, 20(1), 38. Cobbe, J., Veale, M., and Singh, J. (2023). Understanding accountability in algorithmic supply chains. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1186-1197. Crompton, H., and Burke, D. (2023). Artificial intelligence in higher education: the state of the field. International Journal of Educational Technology in Higher Education, 20(1), 22. Dawson, P. (2021). Defending Assessment Security in a Digital World: Preventing E-Cheating and Supporting Academic Integrity in Higher Education. Routledge, London. de Wit, H., and Altbach, P. G. (2021). Internationalization in higher education: global trends and recommendations for its future. Policy Reviews in Higher Education, 5(1), 28-46. European Association for Quality Assurance in Higher Education (2015). Standards and Guidelines for Quality Assurance in the European Higher Education Area. ENQA, Brussels. European Union (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence. Official Journal of the European Union. Floridi, L. (2023). The Ethics of Artificial Intelligence: Principles, Challenges, and Opportunities. Oxford University Press, Oxford. Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daume III, H., and Crawford, K. (2021). Datasheets for datasets. Communications of the ACM, 64(12), 86-92. Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., and Fritz, M. (2023). Not what you have signed up for: compromising real world LLM integrated applications with indirect prompt injection. Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, 79-90. Holmes, W., and Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, 57(4), 542-570. Hou, A. Y. C., Hill, C., Chen, K. H. J., and Tsai, S. (2021). A comparative study of international branch campuses in Malaysia, Singapore, China and South Korea: regulation, governance and quality assurance. Asia Pacific Education Review, 22(2), 181-195. Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., and Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., and Narayanan, A. (2024). AI agents that matter. arXiv preprint, Princeton University. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Kuttler, H., Lewis, M., Yih, W., Rocktaschel, T., Riedel, S., and Kiela, D. (2020). Retrieval augmented generation for knowledge intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474. Longo, L., Brcic, M., Cabitza, F., Choi, J., Confalonieri, R., Del Ser, J., Guidotti, R., Hayashi, Y., Herrera, F., Holzinger, A., et al. (2024). Explainable artificial intelligence (XAI) 2.0: a manifesto of open challenges and interdisciplinary research directions. Information Fusion, 106, 102301. Luo, J. (2024). A critical review of GenAI policies in higher education assessment: a call to reconsider the originality of students work. Assessment and Evaluation in Higher Education, 49(5), 651-664. Manatos, M. J., and Huisman, J. (2020). The use of the European Standards and Guidelines by national accreditation agencies and local review panels. Quality in Higher Education, 26(1), 48-65. Marginson, S. (2022). What is global higher education? Oxford Review of Education, 48(4), 492-517. Miao, F., and Holmes, W. (2023). Guidance for Generative AI in Education and Research. UNESCO, Paris. National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST, United States Department of Commerce, Gaithersburg. Nguyen, A., Ngo, H. N., Hong, Y., Dang, B., and Nguyen, B. T. (2023). Ethical principles for artificial intelligence in education. Education and Information Technologies, 28(4), 4221-4241. Novelli, C., Casolari, F., Rotolo, A., Taddeo, M., and Floridi, L. (2024). Taking AI risks seriously: a new assessment model for the AI Act. AI and Society, 39(5), 2493-2497. Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. (2023). Generative agents: interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 1-22. Perkins, M., and Roe, J. (2024). Academic publisher guidelines on AI usage: a ChatGPT supported thematic analysis. F1000Research, 12, 1398. Popenici, S. (2023). Artificial Intelligence and Learning Futures: Critical Narratives of Technology and Imagination in Higher Education. Routledge, London. Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D., and Barnes, P. (2020). Closing the AI accountability gap: defining an end to end framework for internal algorithmic auditing. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 33-44. Ross, J. (2023). Digital Futures for Learning: Speculative Methods and Pedagogies. Routledge, London. Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., and Scialom, T. (2023). Toolformer: language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36. Selwyn, N. (2022). Education and Technology: Key Issues and Debates, 3rd edition. Bloomsbury Academic, London. Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36. Shneiderman, B. (2022). Human Centered AI. Oxford University Press, Oxford. Tsiligiris, V., and Hill, C. (2021). A prospective model for aligning educational quality and student experience in international higher education. Studies in Higher Education, 46(2), 228-244. Tzimas, D., and Demetriadis, S. (2021). Ethical issues in learning analytics: a review of the field. Educational Technology Research and Development, 69(2), 1101-1133. Veale, M., and Zuiderveen Borgesius, F. (2021). Demystifying the draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), 97-112. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6), 186345. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., and Zhou, D. (2022). Chain of thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35, 24824-24837. Weidinger, L., Uesato, J., Rauh, M., Griffin, C., Huang, P. S., Mellor, J., Glaese, A., Cheng, M., Balle, B., Kasirzadeh, A., et al. (2022). Taxonomy of risks posed by language models. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 214-229. Wilkins, S. (2021). Two decades of international branch campus development, 2000-2020: a review. International Journal of Educational Management, 35(1), 311-326. Williamson, B., and Eynon, R. (2020). Historical threads, missing links, and future directions in AI in education. Learning, Media and Technology, 45(3), 223-235. Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2025). The rise and potential of large language model based agents: a survey. Science China Information Sciences, 68(2), 121101. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: synergizing reasoning and acting in language models. Proceedings of the Eleventh International Conference on Learning Representations.

Universities across very different countries have started to look alike in the way they govern artificial intelligence. They publish similar #policy_documents, buy similar software, write similar rules about student use of generative tools, and describe their work using a shared vocabulary of risk, transparency, and responsible use. This article asks why that convergence is happening and what it costs. It uses #neo_institutional_theory, and in particular the idea of #institutional_isomorphism, to explain how universities become structurally similar not because similarity works better, but because it produces #legitimacy. The analysis is a conceptual and integrative review of scholarship published mainly between 2021 and 2026 on AI policy in higher education, quality assurance systems, accreditation, and transnational technology regulation. It identifies four #isomorphic_pressures that push institutions toward the same model: coercive pressure from binding law and funding conditions, mimetic pressure from imitation under uncertainty, normative pressure from professional bodies and quality agencies, and a fourth pressure that classical theory did not anticipate, which this article calls #infrastructural_isomorphism, produced by dependence on a small number of global technology vendors. The article argues that the practical result is a gap between what universities say and what they do, a phenomenon that institutional theory calls #decoupling. It also argues that conformity is unevenly distributed: institutions in well resourced systems shape the standards, while institutions in less resourced systems import them. The article closes with propositions for future research and with recommendations for regulators, quality agencies, university leaders, and students. Keywords: institutional isomorphism, artificial intelligence regulation, quality assurance, accreditation, higher education governance, policy diffusion, academic integrity, decoupling 1. Introduction In the space of about three years, artificial intelligence moved from a specialist research topic to a governance problem that every university had to answer. The public release of large language models in late 2022 created an immediate practical question for teachers and administrators: what should a student be allowed to do with a machine that writes fluent prose on demand? Institutions answered quickly. They wrote guidance notes, updated academic integrity codes, bought detection software, formed committees, and published statements of principle. What is striking is not that they answered. It is that they answered in the same way. A reader who compares the AI policy of a large public university in the United States with that of a private university in Southeast Asia, a technical institute in the Gulf, and a research university in Northern Europe will find remarkable overlap. The documents use the same categories of permitted, restricted, and prohibited use. They invoke the same values of transparency, accountability, human oversight, and fairness. They point to the same risks of bias, hallucination, and data protection failure. They often name the same commercial tools. Document analyses of institutional guidelines confirm this pattern across national systems, showing that universities converge on a narrow set of policy templates even when their legal environments, student populations, and academic traditions differ sharply (An et al., 2025; Dabis and Csaki, 2024; Humble, 2025; Wang et al., 2024). This article treats that #similarity as a puzzle rather than as evidence of good practice. If a solution spreads because it works, we would expect adoption to follow evidence. But the evidence base for many of the most widely adopted AI measures in higher education is weak. Detection tools, for example, are now standard equipment in thousands of institutions, and yet a substantial body of testing shows that they are unreliable and that their errors fall hardest on students who write in a second language (Elkhatat et al., 2023; Liang et al., 2023; Perkins et al., 2024a). Adoption in this case cannot be explained by performance. It has to be explained by something else. The something else is #legitimacy. Organizations that operate in heavily regulated, publicly funded, and reputationally sensitive fields do not only need to work. They need to be seen to be behaving as a proper organization of their type is expected to behave. In the classic formulation of neo-institutional sociology, formal structures are often adopted as ceremony rather than as technique, because they signal to external audiences that the organization is modern, responsible, and safe to fund, accredit, and trust (DiMaggio and Powell, 1983; Meyer and Rowan, 1977). Universities are among the most legitimacy dependent organizations in existence. Their core product, the #degree, is worthless if external audiences stop believing in it. Two features of the current moment intensify this dependence. The first is the arrival of binding transnational law. The European Union's Artificial Intelligence Act, adopted in 2024, is the first comprehensive statutory regime for AI in a major jurisdiction, and it explicitly classifies several educational uses of AI, including admissions decisions and the evaluation of learning outcomes, as high risk. Because compliance is cheaper to standardize than to localize, and because global technology vendors prefer to build one compliant product rather than many, the effects of this law travel far beyond Europe, a phenomenon widely discussed under the label of the #brussels_effect (Almada and Radu, 2024; Bradford, 2023; Crum, 2025). The second feature is the maturity of the global #quality_assurance apparatus. Long before AI, higher education had already built a dense infrastructure of accreditation agencies, external review cycles, standards frameworks, benchmarking exercises, and international rankings. A recent scoping review of 113 studies concluded that this apparatus is itself a powerful engine of homogenisation, producing convergence in university structures, management practices, and even institutional culture, while at the same time encouraging symbolic compliance rather than substantive change (Guil Gorostidi and Rubio-Arostegui, 2026). AI governance did not arrive in an empty room. It arrived in a room already furnished with standards, indicators, audits, and the professional habits they created. This article brings these two literatures together. Research on AI policy in universities has documented convergence but has rarely explained it with organizational theory. Research on institutional isomorphism in higher education has explained convergence but has rarely examined AI. The contribution here is to connect them, and to argue that AI governance is not a new kind of policy problem. It is the newest and most technologically dense case of an old one. 1.1 Research questions The article addresses four questions. Through which mechanisms do international AI policy regimes and global quality assurance frameworks produce #technological_conformity in universities? What is distinctive about AI as a driver of isomorphism, compared with earlier waves of educational reform? What are the consequences of this conformity for academic autonomy, epistemic diversity, and student fairness? What alternatives exist, and under what conditions can institutions retain #institutional_agency? 1.2 Contribution and structure The article makes three contributions. Conceptually, it extends the classical triad of coercive, mimetic, and normative isomorphism by adding a fourth mechanism specific to the AI era, in which conformity is enforced not by law or imitation but by the technical architecture of the tools themselves. Empirically, it synthesizes recent studies of institutional AI policy across several regions and shows a consistent pattern of convergence coupled with #symbolic_compliance. Practically, it offers concrete recommendations for actors who want accountability without uniformity. Section 2 sets out the theoretical framework. Section 3 reviews the current governance environment. Section 4 explains the method. Sections 5 and 6 present the analysis. Section 7 discusses consequences. Section 8 offers implications for practice. Section 9 states limitations. Section 10 concludes. 2. Theoretical framework 2.1 The core claim of neo-institutional theory The starting point is a simple and counterintuitive claim. Organizations do not become similar mainly because similarity is efficient. They become similar because they operate in a shared environment that rewards recognizable forms. Meyer and Rowan (1977) argued that many formal structures in modern organizations function as myth and ceremony. A university adopts a quality office, an ethics committee, a strategic plan, and a set of key performance indicators partly because these structures do useful work, but also because their absence would be read by funders, regulators, employers, and students as a sign that the institution is not serious. The structures are, in a precise sense, a public performance of #competence. DiMaggio and Powell (1983) then asked what makes organizations in a given field converge, and answered with three mechanisms. #Coercive_isomorphism comes from formal and informal pressure exerted by organizations on which an institution depends. Law, regulation, funding conditions, and contractual requirements are the obvious cases. A ministry that requires an AI policy as a condition of a grant is exercising coercive pressure. So is a regulator that treats automated grading as a high risk activity requiring documentation. #Mimetic_isomorphism comes from uncertainty. When an organization does not know what to do, and when the technology is poorly understood and the consequences of error are severe, it copies organizations it regards as successful. #Copying is cheap, it is defensible, and it distributes blame. If a peer institution of high prestige has adopted a particular AI policy, adopting the same policy is a reasonable bet even in the absence of evidence. #Normative_isomorphism comes from professionalization. Academic developers, quality managers, librarians, registrars, and integrity officers are trained in similar programmes, attend the same conferences, read the same guidance, and belong to the same associations. They carry shared assumptions about what good practice looks like from one institution to the next. Convergence here is not imposed. It is #internalized. 2.2 Organizational fields and legitimacy These mechanisms operate within what the theory calls an #organizational_field, meaning the total set of actors that constitute a recognized area of institutional life. For higher education, the field includes not only universities but ministries, accreditation agencies, ranking publishers, funding councils, publishers, professional bodies, consultancies, and, increasingly, technology vendors. The field defines what counts as a legitimate university and what counts as a legitimate way of doing university work. The critical insight is that the field, not the individual institution, is the unit in which standards are set. A single university cannot unilaterally decide that AI detection is unreliable and abandon it if its accreditor, its national regulator, its insurers, and its peer institutions all expect detection to be in place. To deviate is to invite the question of whether the institution is taking integrity seriously. #Conformity is therefore rational at the level of the individual institution even when it is collectively harmful. 2.3 Decoupling The theory also predicts a specific failure mode. Because the point of adoption is often legitimacy rather than function, organizations frequently adopt a structure formally while leaving daily practice untouched. Meyer and Rowan called this #decoupling. The AI policy exists, is published on the website, and is cited in the accreditation self study, but teaching staff have not read it, students do not understand it, and assessment design has not changed. Decoupling is not simple hypocrisy. It is often the only way an organization can satisfy contradictory demands. A university may be told simultaneously to protect academic integrity, to prepare students for an AI saturated labour market, to respect academic freedom, to reduce staff workload, and to avoid legal risk. No single coherent practice satisfies all of these. A formal policy that is loosely enforced satisfies all of them on paper. 2.4 Extending the theory: infrastructural isomorphism Classical isomorphism theory was developed before the platform economy. It assumed that the pressures on organizations were social, legal, and professional. In the AI era, a fourth mechanism operates, and it is materially different in kind. Universities do not build their own foundation models. They rent access to a small number of models owned by a small number of firms, and they access those models through learning management systems, plagiarism services, student information systems, and productivity suites supplied by an equally small number of firms. The choices those firms make about what the software can and cannot do become, in effect, institutional policy. If the learning platform offers a specific AI feature, it will be used. If the assessment system offers an originality score, that score will enter integrity procedures. If the vendor's model refuses certain tasks, those tasks disappear from the curriculum. This article calls this #infrastructural_isomorphism, and defines it as convergence produced by shared dependence on standardized technical architectures, interoperability requirements, and vendor defaults, rather than by legal command, imitation, or professional norms. The recent literature on post pandemic digitalization has begun to describe exactly this dynamic, noting that university governance is increasingly mediated by algorithmic logic, shared data frameworks, and platform dependence on global providers (Guil Gorostidi and Rubio-Arostegui, 2026; Nunez Valdes et al., 2021; Sangwa et al., 2025). Infrastructural isomorphism differs from the other three in an important way. Coercive rules can be contested in court. Mimetic choices can be revisited. Normative beliefs can be argued with. But a default setting in a procurement contract is invisible, non negotiable for most institutions, and rarely discussed as policy at all. It is #governance_by_default, and it is the least examined form of #technology_policy in universities today. 2.5 A working model The framework used in the remainder of this article can be stated compactly. International regimes and quality agencies generate legitimacy criteria. Universities, being legitimacy dependent, adopt structures that satisfy those criteria. Because the criteria are generated centrally and the adopting institutions are numerous and unequal, adoption produces convergence. Because the criteria are often symbolic and the underlying technology is poorly understood, adoption is frequently decoupled from practice. And because the technology is supplied by a concentrated market, convergence is reinforced by the tools themselves. 3. The governance environment 3.1 Binding law and the reach of the AI Act The European Union's AI Act is the anchor of the current regulatory environment. Its structure is risk based: obligations scale with the potential of a system to harm health, safety, or fundamental rights. Educational uses are treated seriously. Systems used to decide access to education, to evaluate learning outcomes, to assign students to programmes, or to monitor and detect prohibited behaviour during examinations fall within the high risk category, which triggers requirements on documentation, data governance, transparency, human oversight, and post market monitoring. Two consequences follow for universities everywhere, not only in Europe. The first is direct. A university outside the EU that enrols European students, that runs a branch campus in Europe, that participates in European research consortia, or that seeks recognition of its qualifications in Europe has a practical interest in demonstrating alignment. The second is indirect and more powerful. Vendors serving a global market rarely maintain two versions of a product. The cost of maintaining a compliant and a non compliant variant usually exceeds the cost of making the compliant variant the default everywhere. Scholarship on regulatory diffusion has examined this pathway closely. Some accounts stress that the EU's market size makes unilateral standard setting likely, producing a de facto #regulatory_diffusion in which non European users receive European rules without ever having voted on them (Bradford, 2023). Others are more cautious, arguing that the AI Act's dependence on technical standards and its heavy reliance on implementation detail may weaken its external pull, or transform it into a form of experimentalist governance rather than straightforward export (Almada and Radu, 2024; Crum, 2025). The debate is unresolved. What is not in doubt is that a rulebook written in one region now shapes procurement conversations in universities that had no part in writing it. 3.2 Soft law, guidance, and the ethics industry Alongside binding law sits a much larger body of non binding instruments: intergovernmental recommendations, national AI strategies, ministerial guidance, sectoral codes, and institutional principles. These instruments have no legal force. Their influence is nonetheless considerable, because in a legitimacy driven field, the ability to cite an internationally recognized principle is itself valuable. A recurring finding in policy analysis is that these instruments converge on a small and repetitive set of values. Transparency, fairness, accountability, privacy, human oversight, and beneficence appear again and again. The convergence has been read optimistically as an emerging global consensus. It can equally be read as evidence of #normative_isomorphism at the level of states and agencies, which then transmit the same vocabulary downward to institutions. Studies of university AI policies find that institutions reproduce this vocabulary closely, often without operational definitions, without allocation of responsibility, and without any mechanism for detecting failure (Azevedo, 2025; Dabis and Csaki, 2024; Humble, 2025). The result is a documentary genre. #Ethics_language becomes the accepted way of demonstrating seriousness, and the demonstration substitutes for the substance. 3.3 Quality assurance and accreditation as transmission belts The most important channel through which global expectations reach the individual academic is #accreditation. Quality assurance in higher education is a mature, professionalized, and internationally coordinated activity. In Europe, the standards and guidelines associated with the European Higher Education Area define what an internal quality system should contain. In other regions, national agencies perform equivalent functions, and many explicitly benchmark themselves against international reference points. Programme level accreditation by professional bodies adds a further layer, particularly in engineering, business, medicine, and teacher education. The scoping review by Guil Gorostidi and Rubio-Arostegui (2026) is direct about the consequences. Quality systems, it concludes, act as vectors of isomorphism in all three classical modes, and their effects are strongest in young institutions and in peripheral systems, which replicate dominant models in order to obtain legitimacy. It also documents the predictable side effects: bureaucratisation, the dominance of what can be counted, administrative overload, and a persistent gap between declared standards and actual practice. AI has now entered this apparatus. Accrediting bodies have begun issuing guidance on the use of AI in accreditation processes and on the governance of AI within accredited institutions. National quality agencies have started to build AI expectations into their frameworks, and at least one national system has embedded AI directly into its quality assurance and accreditation architecture as a model intended to be adaptable elsewhere (Ramezani and Mostafavi, 2025). The mechanism is exactly the one the theory predicts. Once AI governance appears in an accreditation standard, it becomes a compliance object. Once it is a compliance object, every institution must produce evidence of it. Once every institution must produce evidence, evidence production converges on whatever format the agency finds legible. 3.4 Rankings, benchmarking, and reputational pressure Rankings deserve separate mention because they operate on a different logic. They are not regulators and they have no authority, yet they change behaviour, because they are read by students, employers, ministries, and donors. #Rankings reward what is measurable. If AI readiness becomes a ranked or benchmarked category, institutions will optimize for the indicator rather than for the capability. This is the classic reactivity problem in performance measurement, and it has been documented repeatedly in higher education, where metrics do not merely describe institutional reality but actively reshape it. The risk in the AI case is specific: an institution can score highly on AI readiness by purchasing licences, publishing a policy, and running a workshop, while changing nothing about how it teaches or assesses. 3.5 Four regional patterns Convergence does not mean identity. It is useful to look at how the same pressures land differently in four settings, because the differences reveal which mechanism is doing the work. Europe. Here #coercive_pressure dominates. A binding statute, a mature quality assurance architecture built around the European Higher Education Area, and a strong data protection tradition combine to produce compliance oriented AI governance. European institutions are documenting, classifying, and recording. The risk in this setting is not lawlessness. It is bureaucratic drift, in which the volume of documentation grows while teaching stays the same. North America. Here #mimetic_pressure dominates, because binding federal AI law is absent and institutions are left to look at each other. Comparative mapping of institutional guidance at large public universities shows both wide adoption and wide variation in depth, with many documents offering principles without procedures. Regional accreditors and professional bodies are now filling the vacuum, which means that a system that began as imitative is being converted into a normative one. The Gulf and parts of Asia. Here national strategy is the engine. Governments have made AI a stated pillar of economic policy, and universities are expected to align. Quality agencies in the region have begun to embed AI expectations directly into accreditation frameworks, which makes the coercive channel unusually short and unusually fast. Where accreditation is linked to AI benchmarks, institutions respond quickly, and neighbouring systems then imitate the leader. This is coercive and mimetic pressure operating in sequence, and it produces convergence at a speed that Europe, with its slower legislative machinery, cannot match. Africa, South Asia, and Latin America. Here the dominant experience is #importation. National AI strategies exist in many countries, but enforcement capacity is thin, budgets are constrained, and connectivity is uneven. What arrives is the tool and the template, usually together, usually from elsewhere. Research on AI and knowledge systems in the Global South describes the pattern as one in which ethical, legal, and policy frameworks are adopted wholesale without contextual adaptation, while the underlying systems encode linguistic and cultural assumptions drawn from Western training data. The formal similarity between an AI policy in Nairobi and one in Toronto can be very high. The practical meaning of the two documents can be completely different. The comparison supports a claim that runs through the rest of this article. The same global standard, applied across unequal systems, does not produce equal outcomes. It produces #asymmetric_conformity: substantive compliance where capacity exists, and symbolic compliance where it does not. 4. Method 4.1 Design This article is a conceptual and integrative literature review with a theory building purpose. It is not an empirical study, it does not report new primary data, and it does not claim statistical generalizability. Its aim is to synthesize an emerging body of work and to propose an explanatory framework that can be tested empirically by others. 4.2 Source selection Sources were selected according to four criteria. First, relevance to at least one of the three core domains: institutional theory in higher education, AI policy in higher education, and transnational regulation of AI. Second, recency, with a strong preference for work published from 2021 onward, reflecting the fact that the object of study barely existed before that date. Third, quality, with priority given to peer reviewed journal articles, scholarly books, and systematic or scoping reviews. Fourth, geographical breadth, with deliberate inclusion of work addressing contexts outside Europe and North America, since a literature that describes conformity while itself being confined to the systems that set the standards would reproduce the very bias it studies. A small number of foundational theoretical texts predating this window are retained because the framework is unusable without them. They are cited as theory, not as evidence about current conditions. 4.3 Analytical procedure The analysis proceeded in three stages. In the first stage, the classical isomorphism triad was used as an initial coding frame, and claims from the reviewed literature were sorted into coercive, mimetic, and normative categories. In the second stage, material that fitted none of the three categories was examined separately. This residual material clustered strongly around platform dependence and technical defaults, and generated the fourth mechanism proposed in Section 2.4. In the third stage, the literature was re read for evidence of decoupling, resistance, and adaptation, in order to avoid a deterministic account in which institutions merely absorb whatever the field demands. 4.4 Limitations of the method The limitations are real and should be stated at the outset rather than buried. The evidence base is young, dominated by document analyses and by studies of English language institutions, and heavily weighted toward well resourced systems. Document analysis reveals what institutions say, not what they do, which means the literature is structurally better at detecting formal convergence than at detecting decoupling. The account of infrastructural isomorphism offered here is theoretically motivated and consistent with recent scholarship on digital transformation, but it has not yet been tested with procurement data or with observational study of how vendor defaults shape teaching. These gaps are converted into research propositions in Section 9. 5. Analysis: the four mechanisms at work 5.1 Coercive pressure Coercive pressure in AI governance operates through four channels. The first is statutory obligation. Where a university deploys a system that a law classifies as high risk, obligations attach directly. Documentation must exist, human oversight must be demonstrable, and records must be retained. These are not aspirational values. They are auditable artefacts. The second is #funding_conditionality. Research councils, ministries, and international donors increasingly attach AI related conditions to funding: disclosure of AI use in grant applications, data management requirements, ethical review of AI enabled research. Because grant income is existential for research intensive institutions, these conditions are obeyed quickly and without much argument. The third is contractual and liability pressure. Insurers, data protection authorities, and legal counsel push institutions toward defensible positions. A university that can show it followed a recognized framework is in a stronger position after an incident than one that exercised independent judgment and got it wrong. This asymmetry systematically favours conformity. Doing what everyone else does is the cheapest available form of #risk_management. The fourth is #accreditation_conditionality, discussed above. It is the most powerful of the four in ordinary institutional life, because accreditation cycles reach every programme and every department, whereas statutes and grants reach only some. An instructive case is documented in a study of Omani higher education, which found coercive pressure operating through top down national mandates while, in the absence of enforcement structures, institutions also looked sideways to regional exemplars in neighbouring states and imitated them informally. The same study noted that accreditation linked AI benchmarks in a neighbouring system created exactly the compliance incentive that was missing at home. This is a clean illustration of the mechanisms operating simultaneously and of how the presence or absence of an enforcement structure changes which mechanism dominates. 5.2 Mimetic pressure Mimetic pressure is the mechanism most visible in the first generation of institutional AI policies, and the reason is straightforward: in 2023, nobody knew what to do. Under acute uncertainty, an institution has three options. It can research the problem, which is slow and expensive. It can improvise, which is risky and exposes leadership to blame. Or it can copy a respected peer, which is fast, cheap, and defensible. Most copied. Document analyses show the fingerprints of copying clearly. Policies share structure, sequence, phrasing, and categories. Many institutions adopted a three tier permission scheme distinguishing prohibited, permitted with disclosure, and freely permitted uses. Many adopted the same assessment taxonomy. Many adopted the same disclosure statement template. The convergence is too tight to be explained by independent invention (An et al., 2025; Humble, 2025; Wang et al., 2024). Two features of academic mimicry deserve attention. First, prestige asymmetry governs the direction of copying. Institutions copy upward. The policies of a small number of highly ranked, mostly Anglophone universities propagate through the system, which means the assumptions embedded in those policies (about class size, staff to student ratios, assessment format, and technological access) propagate as well, into institutions where those assumptions do not hold. Second, frameworks act as accelerants. When a well designed, freely available framework appears, it spreads rapidly because it solves the copying problem elegantly. The Artificial Intelligence Assessment Scale, for example, offers a graduated scale of permitted AI involvement in assessment, from no AI to full AI integration, and was explicitly designed to give institutions a defensible, teachable structure (Perkins et al., 2024b). Frameworks of this kind are genuinely useful, and the point here is not to criticize them. The point is structural: the very quality that makes a framework useful, namely that it can be adopted without local research, is also what makes it an engine of #policy_convergence. 5.3 Normative pressure Normative pressure is quieter and more durable than the other two, because it works through people rather than documents. The professionals who write and implement AI policy in universities form a transnational community. Academic integrity officers, learning technologists, quality managers, and educational developers attend the same conferences, follow the same practitioner literature, hold the same credentials, and increasingly hold the same certifications. They arrive at their institutions already carrying a template of what a responsible AI policy looks like. When they move between institutions, the template moves with them. Professional bodies reinforce this. Disciplinary accreditors in engineering, business, and health professions issue AI guidance that becomes, in practice, mandatory for the programmes they accredit. Publishers issue authorship rules that become de facto research policy. Vendors of integrity software run training that becomes de facto staff development. The effect is that institutional AI policy is written by a professional field with a shared worldview, and that worldview has a definite content. It treats AI primarily as a #compliance_risk to be managed rather than as an epistemological development to be argued about. It privileges procedure over pedagogy. It is comfortable with documentation and uncomfortable with disagreement. 5.4 Infrastructural pressure The fourth mechanism is the one that distinguishes AI from earlier reform waves. Consider the practical sequence by which AI actually enters a classroom. A university licenses a learning management system. The vendor adds AI features to that system. Those features are enabled by default. Teaching staff, who did not choose them and were not consulted, discover them in use. The institution then writes a policy governing what has already happened. Consider the same sequence for assessment. A university licenses similarity checking software. The vendor adds an AI writing indicator. That indicator produces a number. The number appears in the same interface where academic misconduct decisions are already made. Staff, reasonably, treat it as evidence. In both cases the substantive policy decision was made by a vendor, in a product meeting, in another country, and was transmitted to the institution as a software update. This is what governance by default means. It is isomorphism without a decision. Three features make infrastructural pressure especially strong. #Market_concentration. The number of organizations capable of training frontier models is very small, and the number of platforms through which universities access educational software is only slightly larger. Concentration at the supply end guarantees similarity at the adoption end. #Interoperability_requirements. Systems must talk to each other. Data must move between the student record, the learning platform, the library, and the analytics dashboard. Interoperability standards are, by construction, homogenising: they define the shape that data must take, and therefore the shape that institutional processes must take (Sangwa et al., 2025). #Invisibility. Coercive rules are published. Mimetic choices are discussed in committees. Normative beliefs can be challenged in a seminar. A default setting is none of these. It does not appear on any agenda. It cannot be appealed. Most academic staff do not know it exists. The consequence is that the deepest layer of AI conformity in universities is also the least governed. #Procurement, not policy, is where the decisive choices are being made, and procurement is treated as an administrative function rather than an academic one. 6. Case domains: where conformity becomes visible Abstract mechanisms are best examined where they touch practice. Four domains show the pattern with unusual clarity. 6.1 Academic integrity and the detection industry The #detection_story is the strongest available evidence that conformity in this field is driven by legitimacy rather than by evidence. The empirical position is not seriously contested. Testing shows that AI text detectors perform inconsistently, that their accuracy degrades sharply when the text is lightly edited, that they generate false positives on human writing, and that they are biased against writers whose first language is not English (Elkhatat et al., 2023; Liang et al., 2023; Perkins et al., 2024a). The bias has a technical explanation. Many detectors rely, directly or indirectly, on statistical measures of predictability, and second language writers tend to use a more limited and more conventional vocabulary. The detector reads restraint as machine generation. The consequences fall on identifiable people. A false accusation of misconduct is a serious event in a student's life, and the accusation is difficult to rebut because there is no source document to point to, unlike in conventional plagiarism (Giray, 2024). Where the accused student is an international student on a visa, the stakes rise further. The evidence has been public for some time. Detection tools nevertheless remain widely deployed. Why? Because deployment is legitimacy producing. An institution that has purchased detection software can tell its accreditor, its board, its national regulator, and the press that it takes integrity seriously and has invested in safeguards. An institution that has examined the evidence and declined to purchase detection software must explain, repeatedly, why it is not doing what everyone else is doing. The first position is easy to defend and the second is not, regardless of which is better supported. #Detection_software is therefore best understood not primarily as a technology but as a legitimacy artefact. The pattern is a textbook case of decoupling. The purpose declared is integrity. The function performed is reassurance. And a subset of students pays the cost. 6.2 Assessment design #Assessment_design is where AI has produced genuine intellectual work and also its most striking uniformity. The dominant institutional response has been to classify. Assessments are sorted into categories according to how much AI involvement is permitted, and students are required to disclose. Scales of this kind are pedagogically defensible and represent a real improvement on blanket prohibition (Perkins et al., 2024b). But two problems follow from the speed of adoption. The first is that scales are being applied without the assessment redesign that gives them meaning. Declaring that an essay is a no AI task does not make it one. In the absence of supervised conditions, the declaration is unenforceable, and everybody involved knows it. The scale then becomes a ritual of honesty rather than a design decision, which is decoupling in its purest form. The second is that the categories themselves import assumptions. A permission scale assumes that AI use is a discrete, disclosable event, like citing a source. But for many students AI is now ambient: it is in the word processor, the search engine, the browser, the note taking application, and the phone. A framework built around discrete disclosure will misdescribe the reality it governs, and it will do so identically in every institution that adopts it. 6.3 Admissions, analytics, and automated decision making The high risk classifications in the AI Act point at exactly the places where universities use AI to make consequential decisions about people: admissions, progression, and prediction of student outcomes. Here, conformity has an obvious #upside. Requirements for documentation, human oversight, and bias testing are genuinely protective, and their spread through the field is a public good. If the Brussels effect delivers due process protections to students in jurisdictions whose own legislatures have not acted, that is a benefit, not a cost. But the same requirements produce a subtler distortion. Compliance is expensive. Documenting a model, testing it for bias across protected groups, and maintaining a monitoring regime requires expertise that most institutions do not have and cannot afford. The predictable response is to outsource: to buy a compliant system from a vendor rather than to build or adapt one locally. Compliance thus strengthens the vendors, deepens infrastructural dependence, and narrows the range of systems in use. A regulation designed to constrain technology providers ends up #consolidating them. 6.4 Research and publication The research domain shows normative isomorphism operating almost without institutional mediation. Publishers, not universities, set the rules on AI in scholarly writing, and they set them almost identically: AI cannot be an author, AI use must be disclosed, and authors remain responsible for the content. Universities have largely adopted these rules by reference. The rules are sensible. The point of interest is jurisdictional: the substantive research policy of a university is now, in this area, written by a handful of commercial and scholarly publishers and transmitted to institutions as a condition of publication. That is a significant transfer of #governance_authority, and it took place without any institution deciding to make it. 7. Discussion: what conformity costs 7.1 The uneven geography of conformity The most serious consequence of isomorphism in AI governance is that it is not #symmetrical. Some institutions write the standards and others receive them. The standards being globalized were produced in a small number of wealthy systems, and they encode the conditions of those systems. They assume reliable connectivity, institutional legal capacity, staff development budgets, and the ability to negotiate with vendors. They assume that the languages of instruction are well represented in model training data. They assume that students have private access to devices. Where these assumptions fail, the standard does not adapt. It simply becomes harder to meet, and meeting it consumes resources that were needed elsewhere. Scholarship on AI and knowledge systems in the Global South describes this precisely: frameworks are imported wholesale without contextual adaptation, imported tools embed the linguistic and cultural assumptions of their training data, and automated systems undervalue rhetorical traditions that differ from the Anglophone norm. The result is a form of #governance_colonialism in which the appearance of universal standards conceals a highly particular origin. There is a bitter irony in the detection case. Systems that are biased against second language writers are exported, through the same isomorphic channels, to institutions where most students are second language writers. Conformity does not merely fail to help those institutions. It actively harms their students. The scoping review of quality management reaches a similar conclusion by a different route, noting that similar standards produce divergent trajectories: strategic adaptation in consolidated systems, and hybrid, symbolic, and compliance oriented responses in systems with fewer resources and less regulatory autonomy (Guil Gorostidi and Rubio-Arostegui, 2026). #Epistemic_diversity is lost not through argument but through procurement and accreditation. 7.2 The narrowing of institutional judgment A second cost is the erosion of the university's capacity to reach its own conclusions. The intellectual question raised by generative AI is genuinely difficult and genuinely disputed. What is the purpose of an essay? What skills does a graduate need when machine writing is free? Is the ability to produce prose still a proxy for the ability to think? Different disciplines and different educational philosophies will answer differently, and they should be allowed to. Isomorphic pressure short circuits this. It converts a contested educational question into a settled compliance question, and it does so before the argument has been had. A department that concludes, after careful thought, that its students should learn to write without machine assistance because writing is how they learn to think, and a department that concludes the opposite, both end up filling in the same disclosure form. The form is the same in both cases because the form was never about pedagogy. It was about legitimacy. This is what the literature on quality assurance has long warned about: standardization reduces institutional diversity, discourages pedagogical innovation, and shifts attention from substantive improvement to formal compliance. #Academic_autonomy does not usually end by decree. It ends by template. 7.3 Compliance without competence A third cost is that conformity crowds out #capability_building. The resources available for responding to AI are finite. Every hour spent writing a policy, populating a compliance matrix, or preparing an accreditation narrative is an hour not spent redesigning an assessment, training a lecturer, or talking to students about how they actually use these tools. Because compliance is legible to external audiences and capability is not, the incentive is to invest in compliance. A university can be fully compliant and completely unprepared. Indeed, survey evidence on the gap between how widely AI is used by students and staff and how thinly it is governed suggests that many institutions are exactly this: documented and unready. 7.4 The students Students appear in most institutional AI policies as objects of #regulation rather than as participants in it. They are told what is prohibited, warned about misconduct, and required to disclose. They are rarely consulted about the design of the rules, rarely told what the detection software does or how reliable it is, and rarely given a route to contest an algorithmic accusation. This matters for a reason that isomorphism theory itself supplies. If policies are adopted for legitimacy in the eyes of external audiences, then the audiences that matter are regulators, accreditors, and peers. Students are not an audience whose approval is needed. They are the population on which the policy operates. A governance regime built to satisfy external observers will predictably underweight the interests of the people inside it. The practical implications are concrete. Students should know whether their work will be scanned, what the tool's error rate is, what happens if they are flagged, and how they can appeal. They should be told that a detection score is not proof. In a great many institutions, none of this is currently true. 7.5 Resistance, translation, and agency It would be a mistake to end with a picture of universities as passive. Institutional theory has always insisted that organizations negotiate, reinterpret, and sometimes refuse the pressures they face, and the higher education literature documents exactly this. Universities translate global standards into local practice, and translation is a creative act. Some institutions have publicly disabled detection tools after examining the #evidence. Some have rejected surveillance based proctoring on ethical grounds. Some have chosen to redesign assessment rather than to police it. Some have built AI literacy into the curriculum instead of treating AI as a threat to be contained. These decisions share a feature: they were made by institutions that had the capacity to examine evidence independently and the standing to defend a deviation. That is exactly the capacity that is unequally distributed. #Institutional_resistance is available to the strong. This is why the recommendations in the next section are aimed disproportionately at the actors who set the standards, rather than at the institutions that receive them. 8. Implications for practice 8.1 For regulators and international bodies Write standards that #scale_down, not only up. A requirement that presumes a legal team and a data science unit is a requirement that only wealthy institutions can meet. Proportionality should be built into the instrument, not left to the discretion of overstretched national agencies. Regulate the supply side more, and the adoption side less. Universities are the weakest party in the chain. Obligations placed on model providers and platform vendors, including obligations of transparency about error rates and about default settings, will do more for students than obligations placed on institutions that cannot inspect the systems they license. Take an explicit position on unreliable tools. Where a class of system is known to produce discriminatory error, saying so plainly would relieve thousands of institutions of the pressure to adopt it. Silence from regulators is read by institutions as permission, and, worse, as expectation. 8.2 For quality assurance agencies and accreditors Do not ask for AI policies. Ask for evidence of #AI_practice. The question is not whether an institution has a document. It is whether teaching staff have changed anything, whether students understand the rules, whether accusations are handled fairly, and whether anyone has checked. A standard that can be satisfied by a document will be satisfied by a document, and nothing else will happen. Accept #plural_answers. An accreditor that requires a specific approach guarantees convergence. An accreditor that requires a defensible, evidenced, and internally coherent approach permits diversity while still demanding rigour. #Audit_the_tools, not only the institutions. Agencies are far better placed than individual universities to evaluate whether a widely used commercial system performs as claimed. Doing this centrally, once, would save thousands of institutions from doing it badly or not at all. Ask about #procurement. If the decisive AI choices in a university are made in contracts, then quality review that never examines contracts is reviewing the wrong thing. 8.3 For university leaders Treat #procurement_as_policy. Which platform is licensed, which features are enabled, and which defaults are accepted are curriculum decisions. They should be made with academic input, not delegated to a purchasing function. Distinguish sharply between what is legally required, what is professionally expected, and what is merely being copied. The three demand different responses, and conflating them causes institutions to obey pressures that do not actually exist. Build #capability before writing policy. A policy written by people who do not understand the technology will be copied from a policy written by people who did not understand it either. Give students #standing. A student facing an algorithmic accusation should have the right to see the evidence, to know the tool's limitations, and to appeal to a human being who is empowered to disregard the machine. 8.4 For academic staff The most important observation for individual teachers is that assessment design remains a local decision, and it is the decision that matters most. #Detection is a policing response to a design problem. Where assessment rewards process, iteration, defence of one's reasoning, and supervised demonstration of understanding, the incentive to outsource collapses without any surveillance at all. 8.5 For students Students reading this in a library, which is the likely audience, should take three things from it. Understand that the rules governing your AI use were probably not written by anyone who taught you, and were probably copied from an institution unlike yours. That does not make them illegitimate, but it does mean they can be questioned. Understand that a detection score is a statistical guess and not evidence. If you are ever accused on the basis of one, you are entitled to ask what the tool's false positive rate is, and to point out that the published research on this question does not favour the accuser. Understand that #AI_literacy, meaning the ability to use these systems well, to know what they are bad at, and to explain your own reasoning without them, is the skill that no policy can give you and no tool can take away. 9. Limitations and a research agenda 9.1 Limitations This is a conceptual article. It proposes an explanation and does not test one. Its evidence base is drawn largely from document analyses, which capture stated policy rather than enacted practice, and which are therefore better suited to demonstrating formal convergence than to measuring decoupling. The literature it reviews is skewed toward Anglophone and European institutions, which is itself a symptom of the inequality the article describes. The fourth mechanism it proposes, infrastructural isomorphism, is argued from theory and from adjacent empirical work on digital transformation rather than from direct evidence about procurement decisions. 9.2 Propositions for empirical testing The framework generates testable claims. Six are offered here. Proposition 1. The similarity between two institutions' AI policies will be better predicted by their shared accreditor, shared vendor, and shared professional networks than by similarity in their student populations, disciplines, or national legal environment. Proposition 2. Institutions with lower resource capacity will show higher rates of textual similarity to prestige exemplars, and higher rates of decoupling between policy and classroom practice. Proposition 3. The adoption of AI detection tools will correlate more strongly with peer adoption and accreditation expectations than with any institutional assessment of tool accuracy. Proposition 4. Changes to vendor default settings will precede corresponding changes in institutional policy, rather than following them. Proposition 5. Institutions that examined the evidence base independently before adopting AI tools will show greater variation in policy, and greater alignment between policy and practice, than institutions that did not. Proposition 6. Compliance investment and pedagogical investment will be substitutes rather than complements under resource constraint, so that measured compliance maturity will be negatively associated with measured assessment redesign. 9.3 Methods worth using Testing these claims requires methods the field is not yet using much. Computational text similarity analysis across large corpora of institutional AI policies would establish diffusion pathways directly. Procurement contract analysis would open the black box of infrastructural isomorphism. Ethnographic study of committees would show how legitimacy arguments actually win. Longitudinal work would distinguish genuine institutionalization from a policy fashion that fades. Comparative case studies between systems at the centre and the periphery of standard setting would test the asymmetry claim. And, notably absent so far, systematic study of student experience of algorithmic accusation would put the people most affected into the evidence base. 10. Conclusion Universities are converging on a shared model of artificial intelligence governance, and the convergence is happening faster than the evidence that would justify it. This article has argued that the explanation lies in organizational sociology rather than in technology. Universities are legitimacy dependent organizations operating in a densely institutionalized field. When a novel and poorly understood technology arrives, they adopt the structures that their environment recognizes as responsible, because doing so is the cheapest available way to remain credible. Binding law supplies coercive pressure. Uncertainty supplies mimetic pressure. Professional communities and quality agencies supply normative pressure. And, in a development that classical theory did not foresee, a concentrated technology market supplies infrastructural pressure through defaults that no institution voted on and few can even see. The result is a field in which institutions look increasingly alike, in which formal policy is frequently disconnected from what happens in classrooms, in which the costs of unreliable tools fall on the students least able to contest them, and in which the standards being globalized carry, invisibly, the assumptions of the wealthy systems that produced them. None of this is an argument against regulation. AI in education needs rules, and some of the rules now spreading, particularly those requiring documentation, human oversight, and the right to challenge automated decisions, are genuinely protective. The argument is against a specific failure: the confusion of #conformity_with_quality. A university is not primarily a compliance entity. It is an institution whose purpose is to make and examine claims about the world. The question of what a machine that writes should mean for the education of people who think is exactly the sort of question that universities exist to argue about. It would be a strange failure if that argument were settled, quietly and everywhere at once, by a template, a standard, and a default setting. Hashtags #Institutional_Isomorphism #AI_Regulation #Higher_Education_Governance #Quality_Assurance_Frameworks #Technological_Conformity #Academic_Integrity #Neo_Institutional_Theory #Policy_Diffusion #Accreditation_Standards #EU_AI_Act #Educational_Technology_Policy #Global_Higher_Education #AI_in_Academia #Coercive_Mimetic_Normative #Decoupling_in_Universities References Almada, M., and Radu, A. (2024). The Brussels side-effect: How the AI Act can reduce the global reach of EU policy. German Law Journal, 25(4), 646-663. https://doi.org/10.1017/glj.2023.108 An, Y., Yu, J. H., and James, S. (2025). Investigating the higher education institutions' guidelines and policies regarding the use of generative AI in teaching, learning, research, and administration. International Journal of Educational Technology in Higher Education, 22, 10. https://doi.org/10.1186/s41239-025-00507-3 Azevedo, L. (2025). Institutional policies on artificial intelligence in higher education: Frameworks and best practices for faculty. New Directions for Adult and Continuing Education. https://doi.org/10.1002/ace.70013 Bradford, A. (2023). Digital empires: The global battle to regulate technology. Oxford University Press. Crum, B. (2025). Brussels effect or experimentalism? The EU AI Act and global standard-setting. Internet Policy Review, 14(3). https://doi.org/10.14763/2025.3.2032 Dabis, A., and Csaki, C. (2024). AI and ethics: Investigating the first policy responses of higher education institutions to the challenge of generative AI. Humanities and Social Sciences Communications, 11(1), 1-13. https://doi.org/10.1057/s41599-024-03526-z DiMaggio, P. J., and Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147-160. https://doi.org/10.2307/2095101 Elkhatat, A. M., Elsaid, K., and Almeer, S. (2023). Evaluating the efficacy of AI content detection tools in differentiating between human and AI-generated text. International Journal for Educational Integrity, 19, 17. https://doi.org/10.1007/s40979-023-00140-5 Giray, L. (2024). The problem with false positives: AI detection unfairly accuses scholars of AI plagiarism. The Serials Librarian, 85(5-6). https://doi.org/10.1080/0361526X.2024.2433256 Guil Gorostidi, S. C., and Rubio-Arostegui, J. A. (2026). Quality management in higher education from the perspective of institutional isomorphism: A scoping review. Frontiers in Education, 10, 1720224. https://doi.org/10.3389/feduc.2025.1720224 Humble, N. (2025). Higher education AI policies: A document analysis of university guidelines. European Journal of Education. https://doi.org/10.1111/ejed.70214 Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779 Lu, H. P., and Wang, J. C. (2023). Exploring the effects of sudden institutional coercive pressure on digital transformation in colleges from teachers' perspective. Education and Information Technologies, 28, 15991-16015. https://doi.org/10.1007/s10639-023-11781-x Meyer, J. W., and Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. American Journal of Sociology, 83(2), 340-363. https://doi.org/10.1086/226550 Nunez Valdes, K., Quiros-Alpera, S., and Cerda Suarez, L. M. (2021). An institutional perspective for evaluating digital transformation in higher education: Insights from the Chilean case. Sustainability, 13(17), 9850. https://doi.org/10.3390/su13179850 Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., and Khuat, H. Q. (2024a). GenAI detection tools, adversarial techniques and implications for inclusivity in higher education. https://doi.org/10.48550/arXiv.2403.19148 Perkins, M., Furze, L., Roe, J., and MacVaugh, J. (2024b). The Artificial Intelligence Assessment Scale (AIAS): A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching and Learning Practice, 21(6), 49-66. Ramezani, S. G., and Mostafavi, Z. S. (2025). Developing and validating a comprehensive scale for accreditation standards and quality assurance in e-learning institutions. Education and Information Technologies, 30(15), 21139-21187. https://doi.org/10.1007/s10639-025-13587-5 Sangwa, S., Butera, A., and Mutabazi, P. (2025). Digital transformation of higher education: A post-COVID review of adoption, quality assurance, and governance challenges. Current Research Bulletin, 2, 120-133. https://doi.org/10.55677/CRB/I07-07-CRB2025 Wang, H., Dang, A., Wu, Z., and Mac, S. (2024). Generative AI in higher education: Seeing ChatGPT through universities' policies, resources, and guidelines. Computers and Education: Artificial Intelligence, 7, 100326. https://doi.org/10.1016/j.caeai.2024.100326

Organizational behavior grew up as a science of people. Its core ideas, such as motivation, leadership, group norms, and culture, were built on the assumption that the only actors inside a firm who make decisions, hold status, and shape the behavior of others are human beings. That assumption no longer describes the modern firm. Scheduling systems assign shifts, ranking systems allocate work, scoring systems decide who is hired, pricing systems set the terms of exchange, and increasingly capable software agents now execute multi-step tasks with limited supervision. This article argues that the discipline needs a revised vocabulary and a set of sociological models that treat #algorithmic_actors not as tools that humans use, but as participants in the social order of the organization. Drawing on an integrative review of research published between 2020 and 2025 across organization studies, information systems, economic sociology, and the sociology of work, the article develops a framework called Post-Human Organizational Behavior (PHOB). The framework rests on four constructs: distributed agency, machine readable legitimacy, recursive control, and ecosystem drift. From these constructs the article derives four analytical models: the Layered Agency Model, the Model of Algorithmic Institutional Work, the Post-Human Status Hierarchy, and the Ecosystem Coupling Model. Nine propositions are offered for empirical testing. The article then applies the framework to five settings drawn from existing empirical work: platform logistics, algorithmic finance, clinical decision support, automated hiring, and agentic software development. The contribution is theoretical rather than empirical. The aim is to give students and researchers a workable language for studying firms in which authority is exercised partly by systems that no single person fully understands, and in which the #social_order of work is negotiated between people and machines rather than among people alone. Keywords: post-human organizational behavior, algorithmic management, distributed agency, autonomous agents, sociomateriality, corporate ecosystems, algorithmic control, organizational sociology 1. Introduction Walk into a large distribution center, a hospital radiology unit, a currency trading desk, or the recruiting office of a multinational, and you will find a familiar social world. There are managers and subordinates. There are informal leaders. There are stories people tell about who is competent and who is not. There are rules on paper and rules in practice. All of this is the traditional subject matter of organizational behavior. Look more carefully, though, and you will notice something that the classic textbooks do not prepare you for. In the distribution center, the pace of work is set by a routing system that no supervisor can override without filing an exception. In the radiology unit, the order in which images are read is influenced by a triage model that flags certain scans as urgent. On the trading desk, most orders are placed by execution systems that respond to signals faster than any person can read them. In the recruiting office, the first cut of applicants is made by a screening model whose reasons for rejection are not fully available even to the recruiter who signs the letter. In each of these settings, something is exercising authority that is not a person. It sets priorities, allocates scarce resources, evaluates performance, and shapes what people believe is possible. It has effects that look very much like the effects of a manager, and yet it has none of the properties that our theories of managers assume. It has no career ambitions, no need for esteem, no capacity for shame, and, importantly, no ability to explain itself in the way a supervisor can when challenged in a corridor. This article takes that situation seriously. It proposes that organizational behavior needs to move from a #human_centred model of the firm to what can be called a post-human model. The term post-human here is not science fiction, and it is not a claim that people no longer matter. It is a technical claim about the unit of analysis. A post-human model of the firm is one in which the basic building block of #social_order is not the individual person but the relationship between people and the computational systems with which they act. In such a model, agency is not the property of one entity. It is something produced jointly, in a chain, by people and machines together. The urgency of this shift comes from a change in the nature of the systems themselves. For most of the history of workplace computing, software was a passive instrument. It executed instructions that a person had written and that another person could inspect. The systems now spreading through large firms are different in three respects. First, they learn, which means their behavior is a function of data rather than of explicit rules, and it changes over time without anyone rewriting the code. Second, they are opaque, which means that even their designers often cannot give a full account of why a specific output appeared. Third, and most recently, they are agentic, which means they are able to plan and carry out sequences of actions across multiple tools and systems with only intermittent human checkpoints. Research over the last five years has documented what happens when such systems enter the workplace. Scholars have shown how #algorithmic_management restructures the terrain of control in organizations, how workers respond to being evaluated by systems they cannot see, how professionals defend their expertise when a model contradicts them, and how the boundary between augmentation and automation is unstable rather than fixed (Kellogg, Valentine, & Christin, 2020; Rahman, 2021; Anthony, 2021; Raisch & Krakowski, 2021). This is valuable work. What is still missing is an integrating framework. Studies of gig platforms, studies of clinical AI, and studies of automated hiring tend to speak past one another, because each borrows a different theoretical vocabulary. The result is a fragmented literature in which the same underlying phenomenon appears under many names. This article attempts a synthesis. It has four objectives. The first objective is to explain why the existing conceptual toolkit of organizational behavior does not travel well into settings governed by autonomous systems. The problem is not that our concepts are wrong. It is that they are silently indexed to human properties. Motivation theory assumes a subject who wants something. Leadership theory assumes an influence process between beings who can recognize each other. Culture theory assumes shared meaning. When a scheduling system reshapes the norms of an entire warehouse without ever meaning anything by it, our theories do not so much fail as fall silent. The second objective is to build a framework that lets us speak about such situations precisely. The framework, called Post-Human Organizational Behavior, is developed around four constructs. Distributed agency describes how the capacity to act is spread across a chain of human and non-human contributors. Machine readable #legitimacy describes the way organizations increasingly grant authority to claims that can be represented in data and withhold it from claims that cannot. Recursive control describes the loop in which a system shapes behavior, records that behavior, and then learns from the very behavior it produced. Ecosystem drift describes the slow movement of a corporate ecosystem toward a state that no participant chose, produced by the interaction of many optimizing systems. The third objective is to convert the framework into models that can guide research. Four are offered here. They are analytical devices rather than statistical models, in the tradition of conceptual work in organization theory. The fourth objective is to show that the framework does real work, by using it to reinterpret five empirical settings that have been studied in the recent literature. The article is aimed at graduate students and early career researchers. It is written in plain language on purpose. The subject matter is hard enough without ornamental prose. The remainder of the article proceeds as follows. Section 2 states the problem in the language of theory. Section 3 reviews the relevant literature across four traditions. Section 4 sets out the framework and its propositions. Section 5 describes the method used to build the argument. Section 6 presents the four models. Section 7 applies them to empirical settings. Section 8 discusses implications for the core topics of organizational behavior. Section 9 covers practical implications. Section 10 states limitations, and Section 11 sets out a research agenda before the conclusion. 2. The Problem: Why the Standard Toolkit Struggles 2.1 The hidden anthropocentrism of organizational behavior Every scientific field carries assumptions that are so basic that nobody writes them down. In organizational behavior, the deepest of these is that the entities whose behavior we study are human. This assumption is not stated because until recently it did not need to be. Consider the standard content of an introductory course. Individual differences. Perception and attribution. Motivation. Job design. Teams and group dynamics. Communication. Power and politics. Leadership. Decision making. Organizational culture. Change. Every one of these topics presupposes an actor with an inner life. Perception requires a perceiver. Attribution requires someone who wonders why. Motivation requires a want. Culture requires shared understanding. When we introduce a system that allocates work, evaluates output, and adapts its own behavior, we have introduced something that participates in all of these processes without possessing any of the properties on which the theories depend. A routing system does not perceive, yet it changes what workers perceive. It does not attribute, yet it produces the data from which humans make attributions about each other. It is not motivated, yet it shapes the incentive structure of an entire workforce. It does not share meaning, yet it becomes part of the meaning that people share. The standard move is to say that the system is a tool and that the real actor is the manager who deployed it. This move has become increasingly hard to defend. In many large firms, no single manager chose the system, no single manager understands it, and no single manager can predict its behavior on a given day. The system was procured by one unit, configured by another, trained on data produced by a third, and is now maintained by a vendor. Its outputs shape the working lives of thousands of people. To describe this as a manager using a tool is to describe a river as a person carrying water. 2.2 Three shifts that break the old model Three specific properties of contemporary systems are responsible for the breakdown. The first is learned behavior. A rules based system does what it was told. A learned system does what the data implies. This makes its behavior a moving target. Research on how organizations manage machine learning systems has emphasized that such systems are best understood as continuously evolving rather than as finished artifacts, which means the organization must manage a relationship rather than operate an instrument (Berente, Gu, Recker, & Santhanam, 2021). The second is opacity. Even where the mathematics is public, the reason for a particular output is often not reconstructible in terms a human would accept as an explanation. This is not simply a technical inconvenience. It changes the social structure of the organization, because #accountability depends on the ability to give reasons. Work on how professionals cope with opaque analytical technologies shows that people develop practices of black boxing, in which they stop asking why the system produced an output and instead treat the output as a fact to be worked around (Anthony, 2021). The consequences are organizational, not only technical. The third is agency in the operational sense. Newer systems do not merely output a score for a person to act on. They take actions. They send messages, place orders, open tickets, allocate budgets, escalate cases, and call other systems. The point at which a human being can intervene is no longer at every step but at a checkpoint, and often after the fact. Theoretical work on delegation to and from computational artifacts has argued that this reverses a long standing assumption, since artifacts may now delegate to humans rather than only receiving delegation from them (Baird & Maruping, 2021). Taken together, these three properties mean that the object we are studying is not a tool inside a social system. It is a participant in the social system. 2.3 What a post-human model must deliver A useful post-human model of organizational behavior must satisfy four requirements. It must explain how authority is produced when the entity exercising it cannot justify itself. It must explain how status is allocated in a workforce where proximity to a system, and the ability to interpret it, becomes a source of standing. It must explain how norms form when a large share of behavior is prompted by a system rather than by other people. And it must explain how organizational outcomes emerge when many optimizing systems interact within and across firm boundaries. The framework offered in Section 4 is designed against these four requirements. 3. Literature Review The relevant literature sits in four traditions that rarely cite one another. This section summarizes each, identifies what it contributes, and states what it leaves unresolved. 3.1 Algorithmic management and the sociology of control The most developed body of work concerns the use of computational systems to direct, evaluate, and discipline labor. The now standard synthesis argues that algorithmic control operates through a set of mechanisms that recall classic managerial control but operate at a different speed and scale, and that this creates a contested terrain in which workers develop new forms of response (Kellogg, Valentine, & Christin, 2020). Empirical work in this tradition is rich. Studies of ride hailing and delivery platforms have shown how #power_asymmetry is produced by information advantages rather than by formal hierarchy, and how workers nonetheless retain forms of agency by anticipating, gaming, and reinterpreting the system (Curchod, Patriotta, Cohen, & Neysen, 2020; Bucher, Schou, & Waldkirch, 2021; Cameron, 2022). Others have shown that #resistance is not a separate activity from control but is co-constituted with it, so that the same practices that allow workers to cope also help the system function (Cameron & Rahman, 2022). Research on opaque evaluation has demonstrated that when workers cannot see the criteria by which they are judged, they do not become free of control. They become more controlled, because they over-comply with an imagined standard, a condition described as an invisible cage (Rahman, 2021). Parallel work has mapped algorithmic management as a general phenomenon rather than a platform specific one, and has traced its spread into standard employment relationships, where it interacts with human resource practices in ways that existing employment relations theory did not anticipate (Duggan, Sherman, Carbery, & McDonnell, 2020; Jarrahi et al., 2021; Möhlmann, Zalmanson, Henfridsson, & Gregory, 2021). Studies of #platform_work have also stressed that platforms are not simply intermediaries but institutions that restructure labor markets (Vallas & Schor, 2020; Wood & Lehdonvirta, 2021), and that #surveillance in these settings is spatial and organizational rather than merely informational (Newlands, 2021). Legitimacy has emerged as a central concern, since the acceptance of algorithmic control by those subject to it cannot be assumed and must be produced (Wiener, Cram, & Benlian, 2023). Contribution. This tradition established that computational systems exercise something that functions like managerial authority, and it gave us a mechanism level vocabulary for describing it. Limitation. It remains largely a literature about control over workers. It treats the system as an instrument of management, which is reasonable in platform settings where an identifiable firm designed the system to direct labor. It does not easily extend to settings where the system directs managers, where it acts on other systems, or where no one is in a position to use it as an instrument at all. 3.2 Agency, sociomateriality, and the human-machine relation A second tradition asks a more basic question: what does it mean to say that a technology acts? The strongest recent statement holds that agency in organizations is best understood as conjoined, meaning that neither the human nor the machine acts alone and that different forms of conjoining produce different organizational consequences (Murray, Rhymer, & Sirmon, 2021). A related literature has framed the central management problem as a paradox, since attempts to augment human work with machines tend to slide toward automation, and attempts to automate tend to create new needs for augmentation (Raisch & Krakowski, 2021). Work on #delegation has argued for a symmetrical view in which agentic artifacts and humans delegate to each other (Baird & Maruping, 2021), while conceptual work in information systems has proposed treating artificial intelligence as a form of digital agency that must be theorized in its own terms (Ågerfalk, 2020; Schuetz & Venkatesh, 2020). Field studies give this abstraction flesh. Research on robotic surgery showed that introducing a machine between the surgeon and the patient did not simply change the tool but reorganized coordination, altered who could see what, and dissolved forms of embodied learning that the profession had depended on (Sergeeva, Faraj, & Huysman, 2020). Studies of human in the loop configurations have shown that the loop is not a fixed design but an emergent arrangement that workers construct and reconstruct (Grønsund & Aanestad, 2020). Research on machine learning in policing found that the value of the system depended less on its predictions than on the emergence of a new occupational role, the person who could translate between the model and the officers, a form of #knowledge_brokerage that no one had planned (Waardenburg, Huysman, & Sergeeva, 2022). Scholars have also argued that all organizational actors are now, in effect, theorists of technology, since people continuously develop working accounts of what a system is and what it can do, and those accounts shape outcomes as much as the system itself does (Bailey, Faraj, Hinds, Leonardi, & von Krogh, 2022; Bailey & Barley, 2020). Contribution. This tradition dissolved the idea that agency belongs to one party. It gives us the concept of distributed or conjoined action. Limitation. It is strong on the micro relation between a person and a system and weak on aggregation. It does not tell us what happens when thousands of such relations are stacked into a hierarchy, or when systems interact with other systems across firms. 3.3 Knowledge, expertise, and organizational learning A third tradition asks what happens to knowledge when machines produce it. Studies of clinical settings found that practitioners could not verify whether the training labels on which a diagnostic model was built were correct, which meant that the apparent objectivity of the system rested on a foundation that nobody could inspect (Lebovitz, Levina, & Lifshitz-Assaf, 2021). Follow up work showed that professionals responded by selectively engaging with the system, choosing when to consult it in ways that preserved their own judgment (Lebovitz, Lifshitz-Assaf, & Levina, 2022). Research on organizational change in a firm adopting predictive analytics found that occupational groups engaged in symbolic actions to defend or dismantle existing regimes of knowing, so that the arrival of the algorithm became a political contest over what counts as knowledge (Pachidi, Berends, Faraj, & Huysman, 2021). Theoretical work has warned that substituting machine judgment for human judgment changes the learning properties of the organization itself, because the organization stops generating the kind of experience from which humans learn (Balasubramanian, Ye, & Xu, 2022; Sturm et al., 2021). Related work in strategy has examined complementarity between #machine_learning and human capital rather than pure substitution (Choudhury, Starr, & Agarwal, 2020), and work on human and machine collaboration has framed the issue as one of organization design rather than individual usage (Puranam, 2021). A parallel argument holds that the framing of technology as a substitute for people, rather than a complement to them, is a choice with consequences for productivity and for the distribution of gains (Brynjolfsson, 2022; Acemoglu & Johnson, 2023). Field work on skill acquisition has shown that the pathways through which novices become experts are being quietly cut off, since the routine tasks that once served as apprenticeship are the first to be automated (Beane, 2024). Contribution. This tradition shows that #organizational_learning is not neutral with respect to who does the thinking. When machines take over judgment, the organization's future capacity changes. Limitation. It has focused on professional and expert settings. Its findings have not been generalized into a model of the organization as a whole. 3.4 The society of algorithms and the political economy of ratings A fourth tradition, based largely in sociology, treats classification and scoring as a social force in their own right. A recent review argues that algorithms should be studied as social institutions that classify, rank, and thereby constitute the social world rather than merely describe it (Burrell & Fourcade, 2021). This is extended into an argument that contemporary societies increasingly order people through ranking and scoring, producing a form of stratification based on the ordinal position that data assigns to a person (Fourcade & Healy, 2024). Related work argues that the political economy of computational systems must be understood in terms of the extraction of resources, labor, and data on which they depend (Crawford, 2021), and that the design of machine systems should be constrained by principles that keep human judgment central in professional domains (Pasquale, 2020). Ethnographic work in warehousing and in startups has shown that automation frequently redistributes work rather than eliminating it, and that human labor is often hidden inside systems that present themselves as autonomous (Delfanti, 2021; Shestakofsky, 2024). Within organization studies, a compatible line of work argues that #datafication produces a new condition of behavioral visibility, in which the behavior of employees becomes persistently observable and therefore governable in ways that were not previously possible (Leonardi & Treem, 2020; Hafermalz, 2021). Work on people analytics has warned that the pursuit of measurable performance can crowd out the practical wisdom on which good judgment depends (Gal, Jensen, & Stein, 2020), and work on the lifecycle of algorithmic decision systems has stressed that the ethical stakes arise not at a single moment of design but across the whole life of the system (Marabelli, Newell, & Handunge, 2021). Research on metrics in professional work has documented how quantification reshapes the very activities it claims to measure (Christin, 2020). Contribution. This tradition supplies the macro dimension. It explains why classification systems have effects that outrun anyone's intentions. Limitation. It says relatively little about the internal life of firms. It tends to speak about society, markets, and states rather than about departments, teams, and careers. 3.5 The gap Put together, these four traditions establish four separate facts. Systems exercise control. Agency is distributed. Machine knowledge reshapes expertise. Classification stratifies. What is missing is a model that connects them. We do not currently have a way of saying, in one theoretical language, that a warehouse routing system controls workers, is conjoined with the supervisor who interprets it, degrades the tacit knowledge that once made supervision possible, and simultaneously locks the firm into a competitive position determined by its interaction with the routing systems of other firms. This article proposes such a language. 4. The Post-Human Organizational Behavior Framework 4.1 Definition and scope Post-Human Organizational Behavior (PHOB) is defined here as the study of behavior, authority, status, and norm formation in organizations where a material share of consequential action is initiated, shaped, or executed by computational systems that operate with limited and intermittent human supervision. Three boundary conditions apply. First, PHOB concerns organizations in which computational systems act, not merely inform. A dashboard that reports last quarter's sales is outside scope. A system that reallocates the sales territory on the basis of that data is inside scope. Second, PHOB concerns systems whose behavior is not fully specified by any person. A payroll system that applies a published formula is outside scope, however important it is. A model that scores candidates on learned patterns is inside scope. Third, PHOB is agnostic about consciousness. It makes no claim that systems have experience, intention, or interests. It claims only that they occupy positions in the social structure of the organization that were previously occupied only by people, and that this has consequences that can be studied. This is a functional claim, not a metaphysical one. The distinction matters, because critics of post-human language often assume that it involves attributing an inner life to machines. It does not. 4.2 Construct one: distributed agency The first construct holds that in post-human organizations the capacity to act is not located in a single entity but is distributed along a chain. Consider a rejected job application. Who rejected the candidate? The recruiter signed the letter, but the recruiter never saw the file, because the model ranked it below a threshold. The threshold was set by an analytics team responding to a request from human resources to reduce time to hire. The model learned its patterns from a decade of past hiring decisions made by people who no longer work at the firm. The vendor supplied the base architecture. No one in this chain performed the act of rejection in the way the concept of an act normally implies, yet the rejection unambiguously occurred. This is what is meant by #distributed_agency. It builds directly on the argument that agency in contemporary organizations is conjoined rather than singular (Murray, Rhymer, & Sirmon, 2021), and on the symmetrical view of delegation between humans and artifacts (Baird & Maruping, 2021). PHOB extends these arguments in one respect: it insists that the chain has a temporal dimension. Some of the contributing actors acted years ago and are no longer present. Their decisions persist inside the model as learned patterns. The organization is therefore acted upon by its own past in a way that is not reducible to routine or memory as those terms are normally used. The practical implication is that questions of the form who decided this are frequently unanswerable, and that organizations respond to this unanswerability with a set of social devices. They appoint nominal decision owners. They construct after the fact narratives. They create human sign off steps whose real function is to supply a signature rather than a judgment. These devices are themselves objects of study. Proposition 1. In organizations with high #distributed_agency, formal accountability structures will increasingly diverge from actual causal influence, and the gap will be managed through symbolic sign off practices rather than through redesign of the decision process. Proposition 2. The greater the temporal depth of the agency chain, meaning the extent to which current system behavior reflects decisions by people who have left the organization, the more resistant the system will be to reform through ordinary managerial intervention. 4.3 Construct two: machine readable legitimacy The second construct concerns how claims come to be accepted as valid. In classical accounts, authority in organizations rests on some combination of formal position, expertise, and personal standing. PHOB proposes an additional and increasingly dominant basis: a claim is legitimate to the extent that it can be represented in a form the organization's systems can process. The mechanism is simple and observable. Suppose an experienced nurse believes a patient is deteriorating. The belief is based on years of practice and on cues that she cannot fully articulate. Suppose the early warning system, drawing on vital signs, shows the patient as stable. In many organizations, the nurse's judgment can only be acted upon if she can convert it into the categories the system recognizes. If she cannot, her judgment does not simply lose an argument. It fails to enter the argument at all, because there is no field in which to record it. This is #machine_readable_legitimacy. It connects to the finding that quantification does not merely measure work but reshapes what work is understood to be (Christin, 2020), to the argument that datafication makes behavior persistently visible and therefore governable (Leonardi & Treem, 2020), and to the observation that ranking and scoring have become general mechanisms of social ordering (Fourcade & Healy, 2024; Burrell & Fourcade, 2021). The organizational consequences are significant. Knowledge that resists formalization loses standing, not because anyone decided it was less valuable, but because it cannot be entered. Over time, this creates a selection pressure inside the workforce. People who are good at rendering their work legible to systems rise. People whose contribution is real but unrecordable become invisible, a pattern consistent with evidence that the ability to interpret and translate for systems becomes a source of #status even when it was never a formal part of the job (Waardenburg, Huysman, & Sergeeva, 2022). Proposition 3. As the share of organizational decisions mediated by computational systems rises, the perceived legitimacy of unquantifiable expertise will decline, independent of its demonstrated accuracy. Proposition 4. Employees will engage in translation work, meaning effort spent converting their judgment into machine readable form, and this work will be systematically under-recognized in formal performance evaluation. 4.4 Construct three: recursive control The third construct concerns feedback. A learning system that governs behavior sits in a loop. It shapes what people do. What people do generates data. That data trains the next version of the system. The system is therefore not observing an independent world. It is observing a world it made. Call this #recursive_control. The concept differs from ordinary feedback in one crucial way. In ordinary feedback, the measuring instrument is separate from the thing measured. Here it is not. If a delivery platform rewards speed, drivers drive faster. The data then shows that fast driving is normal, and the model recalibrates its expectations upward. The standard rises without anyone raising it. Evidence for this dynamic appears in the finding that workers subject to opaque evaluation over-comply with an imagined standard, effectively producing the very behavior the system will later learn to expect (Rahman, 2021), and in accounts of anticipatory practices among platform workers who adjust their conduct to what they believe the system wants (Bucher, Schou, & Waldkirch, 2021). Recursive control has three properties worth naming. It is self-confirming. Because the system produces the data that validates it, apparent accuracy rises over time even where genuine predictive validity does not. A hiring model that screens out a category of applicant will never observe that category succeeding, and will therefore never learn that it was wrong. This is closely related to the problem that the ground truth on which such systems are trained may itself be an artifact rather than a fact (Lebovitz, Levina, & Lifshitz-Assaf, 2021). It is norm-generating. Because the loop runs continuously, the standard of acceptable behavior moves without any explicit decision to move it. Norms in post-human organizations therefore arise partly from a process that has no author. It is ratchet-like. The loop tends to move in one direction, usually toward greater intensity, because the behaviors that the system rewards are the behaviors that generate the data that reinforces the reward. Proposition 5. In organizations exhibiting recursive control, performance standards will exhibit unidirectional drift over time in the absence of deliberate counter-intervention, and this drift will not be attributable to any explicit managerial decision. Proposition 6. Systems operating under recursive control will display rising internal validity metrics alongside stable or declining external validity, producing overconfidence in system performance at the organizational level. 4.5 Construct four: ecosystem drift The fourth construct moves from the firm to the field. Firms do not deploy systems in isolation. A retailer's pricing model responds to a competitor's pricing model. A logistics system negotiates with a supplier's scheduling system. A recruitment model draws on a labor market shaped by other recruitment models. An investment system trades against other investment systems. The result is that the environment of any given firm increasingly consists of other firms' systems. This is what is meant by a #corporate_ecosystem driven by algorithmic actors. And because each system is optimizing locally, the ecosystem as a whole can move toward configurations that no participant selected and none prefers. Call this #ecosystem_drift. It is not a conspiracy and it is not a market failure in the standard sense. It is the aggregate consequence of many local optimizations interacting. Three examples illustrate. In pricing, mutually responsive systems can converge on higher prices without any communication between the firms, because each learns that undercutting is punished by rapid matching. Nobody agreed to anything. The outcome nonetheless resembles the outcome of an agreement. In hiring, if many firms screen on similar learned patterns drawn from similar historical data, then a candidate rejected by one is likely to be rejected by all. The labor market develops an effectively coordinated exclusion that no employer chose and no employer can see, because each observes only its own funnel. In content and communication, if firms optimize internal messaging for engagement metrics, attention within the ecosystem migrates toward whatever those metrics reward, and the substantive quality of communication becomes a residual. Drift is hard to detect from inside a single firm, because each firm sees only its own local optimization working as designed. This is precisely why it requires a field level construct. The argument connects to work on data network effects, which shows that the value of a system increases with the data it accumulates and thereby produces self-reinforcing positions (Gregory, Henfridsson, Kaganer, & Kyriakou, 2021), and to political economy arguments about the direction of technological development being a matter of choice rather than fate (Acemoglu & Johnson, 2023). Proposition 7. In fields with high algorithmic density, meaning a high proportion of inter-firm interactions mediated by learning systems on both sides, outcome variance across firms will decline over time as strategies converge, even in the absence of imitation. Proposition 8. Ecosystem drift will be systematically under-detected by individual firms, because the monitoring systems available to each firm are optimized for local performance rather than field level pattern. Proposition 9. Firms that retain a meaningful stock of non-formalized human judgment will show greater strategic variance, and therefore greater potential for both outperformance and failure, than firms whose decision processes are fully formalized. 5. Method This is a conceptual article. It contributes theory rather than data, and it should be evaluated by the standards appropriate to that genre: coherence, novelty, and usefulness for future empirical work. The argument was constructed through an integrative review. Integrative review is appropriate when the aim is to synthesize findings from separate literatures into a new conceptual structure rather than to summarize a mature field. The review proceeded in four stages. Stage one: scoping. Four literatures were identified as relevant on the grounds that each addresses one facet of the phenomenon: algorithmic management, human-machine agency, machine knowledge and expertise, and the sociology of classification. Sources were drawn primarily from work published between 2020 and 2025 in organization studies, information systems, sociology, and management theory outlets, together with a small number of scholarly books from the same period. Stage two: extraction. From each source, three things were recorded: the phenomenon observed, the mechanism proposed, and the assumption about agency that the mechanism implied. The third of these turned out to be the most useful, because it revealed that the literatures differ less in their findings than in what they take an actor to be. Stage three: construct building. Recurring mechanisms across literatures were grouped. Four groupings survived. Each was formalized into a construct with a definition, a stated mechanism, and observable indicators. Where a construct restated an existing idea, the existing term was retained and credited. Where the synthesis produced something that had no name in the literature, a new term was introduced. Stage four: model construction and propositions. Constructs were combined into four models. Nine propositions were derived, each framed so that it could in principle be falsified. Two limitations of the method should be stated at the outset. First, an integrative review is inevitably shaped by the reviewer's judgment about what counts as relevant, and a different reviewer might have drawn different boundaries. Second, because the phenomenon is developing quickly, some of the empirical settings on which the review draws may change faster than the theory can be tested. This is a general hazard of theorizing about emerging technology and is not unique to this article. 6. Four Models 6.1 Model one: the Layered Agency Model The Layered Agency Model (LAM) is a device for locating where action actually occurs in a post-human organization. It replaces the question who decided with the question at which layer was the decision effectively closed. The model proposes five layers. Layer 1: Constitutive. Decisions embedded in the architecture, the choice of objective function, the definition of the target variable, and the selection of the training population. These decisions are usually made once, often by people outside the operating unit, sometimes by a vendor, and they constrain everything above them. They are rarely revisited. Layer 2: Configurational. Decisions about thresholds, weights, exception rules, and the scope of deployment. These are made inside the firm, usually by a technical unit, and they are revisited occasionally. Layer 3: Operational. The moment to moment behavior of the system as it processes cases. No human is present. This is where the great majority of consequential acts occur, measured by volume. Layer 4: Interpretive. The work done by humans to make sense of system outputs, to decide when to trust them, and to convert them into action. This layer is where the literature on selective engagement and #knowledge_brokerage lives (Lebovitz, Lifshitz-Assaf, & Levina, 2022; Waardenburg, Huysman, & Sergeeva, 2022). Layer 5: Accountable. The formal act by which a human being takes responsibility, signs, or communicates the outcome. Legally decisive. Causally often trivial. The model's central claim is that in most post-human organizations the causal weight sits at Layers 1 and 3 while the accountability sits at Layer 5. This mismatch is the structural source of most #governance failures. It also explains a common frustration: reform efforts aimed at Layer 5, such as requiring a human to review every decision, change very little, because the reviewer at Layer 5 has neither the information nor the standing to reopen what was closed at Layer 1. The model yields a diagnostic question for researchers and practitioners: for a given decision type, at which layer was the outcome effectively determined, and at which layer is responsibility formally located? The distance between these two answers is a measure of the organization's exposure. Layer Who acts Frequency Visibility Typical accountability 1 Constitutive Designers, vendors Rare Very low None assigned 2 Configurational Technical units Periodic Low Diffuse 3 Operational System alone Continuous Low None assigned 4 Interpretive Frontline staff Continuous Moderate Informal 5 Accountable Managers, professionals Per case High Formal and full 6.2 Model two: Algorithmic Institutional Work The second model concerns how rules and norms get made. Institutional theory has long argued that organizations are governed not only by formal rules but by taken for granted understandings of what is appropriate, and that actors engage in institutional work to create, maintain, or disrupt these understandings. The classic assumption is that institutional work is done by people, purposefully. PHOB proposes that computational systems perform functionally equivalent work without purpose. Call this #algorithmic_institutional_work. It operates through three mechanisms. Sedimentation. A system that repeatedly produces the same category of outcome makes that outcome appear natural. If a scheduling system consistently assigns night shifts to a certain group, the arrangement acquires the character of a fact about the world rather than a decision that someone made. Over time, people stop asking why. Foreclosure. A system forecloses alternatives simply by not representing them. If a project approval tool has no field for long term or strategic in its list of justifications, then long term projects do not get rejected. They fail to be proposed. The absence of a field becomes the absence of an option. Amplification. A system amplifies whatever it measures. Because measured things become visible and visible things become manageable, the measured aspects of work expand to fill the available attention, and the unmeasured aspects contract. This is consistent with evidence that metrics reshape professional work rather than merely record it (Christin, 2020; Gal, Jensen, & Stein, 2020). The importance of this model is that it explains institutional change without an author. Existing theory struggles here, because it is built around the idea of the institutional entrepreneur, a purposive actor who works to change the rules. In post-human organizations, rules change because a threshold was retuned. There was no entrepreneur, only an engineer, and the engineer was not thinking about institutions. This connects directly to the finding that the arrival of predictive analytics in an organization triggers symbolic contests between occupational groups over the legitimate way of knowing (Pachidi, Berends, Faraj, & Huysman, 2021). The PHOB extension is that the system is not merely the object of that contest. It is a participant in it, because its outputs continuously reinforce one side. 6.3 Model three: the Post-Human Status Hierarchy The third model addresses stratification inside the workforce. Traditional models of workplace status rest on hierarchy, expertise, and social capital. PHOB proposes that a further dimension has emerged, which can be called algorithmic proximity. This is the degree to which a person's role gives them the ability to see, question, configure, or credibly interpret the systems that govern the work of others. The model proposes five positions, which cut across formal hierarchy. Architects. Those who set objective functions and design systems. Very few in number. Extremely high causal influence. Often organizationally junior and physically distant from the work their decisions govern. Frequently employed by a vendor rather than the firm. Configurers. Those who tune, retrain, and set thresholds. Moderate in number. High influence, low visibility. Their work is treated as technical maintenance and is rarely reviewed as policy, even though its effects are policy. Brokers. Those who translate between system and worker. They explain why the model said what it said, and they decide when it should be believed. Their standing derives entirely from their interpretive monopoly. This position is well documented empirically (Waardenburg, Huysman, & Sergeeva, 2022), and is characteristically informal, meaning the person occupying it holds real power without any formal recognition of it. Subjects. Those whose work is directed and evaluated by the system, and who have no channel to contest it other than through a broker. This is the position of most frontline workers in platform and warehouse settings (Curchod et al., 2020; Delfanti, 2021). Ghosts. Those whose contribution is real but not representable in the data. Mentors, people who calm conflicts, people who notice the thing that goes wrong before it goes wrong. In a system of #machine_readable_legitimacy these people are invisible by construction. They tend to be identified only when they leave and something quietly breaks. The presence of hidden human labor inside apparently autonomous systems is a recurring finding in ethnographic work (Shestakofsky, 2024; Delfanti, 2021). The key claim of this model is that the post-human status order is only loosely coupled to the formal org chart. A junior data scientist who owns a threshold may exercise more influence over the daily lives of a thousand people than the director who nominally manages them. Conventional organizational behavior, which reads power from position, will systematically misdescribe such an organization. Position Basis of standing Formal recognition Real influence Architects Design authority Low to moderate Very high Configurers Parameter control Low High Brokers Interpretive monopoly Very low High and local Subjects None Formal role only Low Ghosts Unrecorded contribution None Invisible until absent 6.4 Model four: the Ecosystem Coupling Model The fourth model operates at the level of the field. It proposes that the relationship between firms in an algorithmically dense field can be characterized along two dimensions. The first dimension is coupling tightness, meaning the speed and automaticity with which one firm's systems respond to another's. Where a competitor's price change triggers an automatic response within seconds, coupling is tight. Where it triggers a quarterly review by a pricing committee, coupling is loose. The second dimension is representational overlap, meaning the extent to which the firms' systems learn from the same or similar data and represent the world using the same categories. Two banks using similar risk models trained on similar data have high representational overlap, whether or not they realize it. Crossing these produces four field types. Loose coupling, low overlap. The classic competitive field of the twentieth century. Firms differ, respond slowly, and outcomes vary widely. Strategic variety is high. Tight coupling, low overlap. A volatile field. Firms respond instantly to each other but see the world differently, so responses are not aligned. Expect rapid oscillation and occasional cascading failures, as has been observed in automated trading environments. Loose coupling, high overlap. A slowly converging field. Firms think alike but act slowly. Outcomes converge gradually and the field homogenizes without anyone noticing. Tight coupling, high overlap. The condition of maximum drift. Firms think alike and respond instantly. The field can move rapidly and collectively to a position that no participant chose. Systemic risk is highest here, and so is the likelihood of parallel exclusion in labor and credit markets. This cell is where Proposition 7 should be tested. The model implies a strategic point that runs against current management practice. Firms currently compete to adopt the same systems from the same vendors trained on the same data. This maximizes representational overlap. If the model is right, the effect is to reduce the variance of outcomes across the field, which is to say it destroys the very differentiation that strategy is supposed to produce. The competitive advantage of retaining idiosyncratic human judgment may therefore be greater, not smaller, in an algorithmically dense field. This is the substance of Proposition 9. 7. Applying the Framework: Five Settings The value of a framework is what it lets you see. This section applies PHOB to five settings that are well documented in recent research. The purpose is illustration, not evidence. 7.1 Platform logistics and delivery work The empirical picture is established. Workers are directed by systems that assign tasks, monitor pace, and score performance, and they cannot see the criteria being used (Rahman, 2021; Möhlmann et al., 2021). They develop coping practices, including anticipation and gaming (Bucher, Schou, & Waldkirch, 2021; Cameron, 2022). Resistance and control turn out to be entangled rather than opposed (Cameron & Rahman, 2022). PHOB adds three readings. Under the Layered Agency Model, the worker's grievance is normally directed at Layer 5, the support agent who answers the appeal, or at Layer 4, the local coordinator. But the deactivation was determined at Layer 1, by the choice of an objective function that treats a completion rate below a threshold as a risk. The appeal process is therefore not merely unhelpful. It is structurally incapable of helping, because the person hearing it operates at a layer where the decision is already closed. This reframes a complaint about bad customer service as a claim about organizational architecture. Under recursive control, the acceleration of the work is explained without recourse to managerial intent. As workers speed up to satisfy an invisible standard, the data on which the system trains shifts. The expected pace rises. Nobody ordered it. This is #ratchet_effect behavior, and it predicts that pace will intensify even in firms whose managers sincerely oppose intensification. Under the Ecosystem Coupling Model, competing platforms operating in the same city with similar models and similar data occupy the tight coupling, high overlap cell. This predicts convergence in pay rates and working conditions across platforms without any communication between them, and it predicts that a worker excluded from one platform will find similar conditions at the next. Whether this holds is an empirical question, and it is a testable one. 7.2 Algorithmic finance Trading is the oldest domain of autonomous algorithmic action, and it is the clearest illustration of ecosystem drift. Here, the majority of consequential action occurs at Layer 3 in the Layered Agency Model. No human is present when the order is placed. Accountability, however, sits at Layer 5, with a desk head who may not have been able to observe the trade until it completed. The layers are separated not only by information but by time, since the system acts on a scale at which human intervention is physically impossible. The Ecosystem Coupling Model describes this field as tightly coupled with substantial representational overlap, since many participants use similar signals derived from similar data. The model therefore predicts occasional collective movements that no participant intended, which is a reasonable description of episodes in which markets have moved violently without any external news. The theoretical point for organizational behavior is that these are not failures of individual firms. They are properties of the field, and they cannot be prevented by any single firm behaving well. The status model also applies. The quantitative researcher who designs a signal exercises influence far beyond the formal seniority of the role, while the experienced trader whose feel for the market was once the firm's core asset drifts toward the Ghost position, valuable but increasingly unrecordable. 7.3 Clinical decision support Healthcare has produced the most careful studies of what happens when a model contradicts a professional. Practitioners cannot verify the labels on which diagnostic models were trained, so the apparent objectivity of the system rests on a foundation nobody can inspect (Lebovitz, Levina, & Lifshitz-Assaf, 2021). Professionals respond by engaging selectively, consulting the system in some circumstances and not others (Lebovitz, Lifshitz-Assaf, & Levina, 2022). The introduction of machines into surgical work reorganizes coordination and can erode the embodied learning through which the profession reproduced itself (Sergeeva, Faraj, & Huysman, 2020). PHOB reads this through machine readable legitimacy. The clinician's intuition, however well founded, has no field in the record. The model's score does. Over time, the score wins, not because it is more accurate but because it is more enterable. The consequence is a slow erosion of the standing of clinical judgment that is invisible in any single case, since in any single case the clinician retains the formal right to override. The Ghost position is also visible here. The nurse who notices that something is wrong before the vital signs move is performing work of enormous value that the system cannot see. She is not resisting the system. She is simply not represented in it. The learning implication is serious. If routine cases are handled by the model and only difficult cases reach the human, then junior clinicians lose the routine practice through which they once learned to recognize the difficult case. The pipeline of expertise is cut at the base, a dynamic identified across occupations (Beane, 2024; Balasubramanian, Ye, & Xu, 2022). 7.4 Automated hiring Hiring shows all four constructs at once. #Distributed_agency: the rejection has no author, as described in Section 4.2. Machine readable legitimacy: candidates whose merits are unrecordable are not rejected. They are unseen. Recursive control: the model never observes the counterfactual. It cannot learn that the people it screened out would have succeeded, because they were never hired. Its accuracy therefore appears to improve while its validity may not, exactly as Proposition 6 predicts. Ecosystem drift: because many firms screen on similar learned patterns drawn from similar historical data, a candidate rejected by one is likely to be rejected by many. This produces a coordinated exclusion with no coordinator. Individual firms cannot detect it because each sees only its own funnel. The point is not that any firm behaved badly. The point is that the field produced an outcome that no firm intended and none can observe, which is precisely the definition of drift given in Section 4.5. This reading has a policy implication. Auditing individual firms for bias, which is the standard remedy, addresses the firm level but not the field level. If the mechanism is representational overlap across firms, then firm by firm audit will find each firm within tolerance while the aggregate exclusion persists. 7.5 Agentic software development The newest setting is also the one where PHOB is most necessary, because the systems involved come closest to acting as organizational members. In software organizations, systems now write code, open and review pull requests, respond to incidents, and call other systems to complete multi-step tasks. The human role shifts from producing work to specifying it and to approving it after the fact. Under the Layered Agency Model, this pushes the volume of consequential action decisively into Layer 3 while formal responsibility remains at Layer 5. Under the status model, the Broker position becomes central, since the person who knows how to specify a task well, and who can tell when the output is subtly wrong, acquires enormous informal standing. Under recursive control, systems trained substantially on the output of previous systems risk a compounding of their own conventions, which raises questions about long term diversity in engineering practice that the field has barely begun to examine. The skill implication mirrors the clinical case. If the routine work through which junior engineers once learned is absorbed by systems, the profession loses its training ground. This is the automation and augmentation paradox in a particularly sharp form (Raisch & Krakowski, 2021; Beane, 2024). The organization gains short term productivity and loses the mechanism by which it renews its own capability. This trade is rarely made explicit, because the cost falls in a future budget period and on a future manager. 8. Discussion: Rewriting the Core Topics of Organizational Behavior If the framework holds, then the standard topics of the discipline require revision rather than replacement. This section takes them one at a time. 8.1 Motivation Motivation theory asks what makes people put in effort. Its dominant traditions locate the answer inside the person, in needs, expectations, or the meaning of the work. Post-human organizations complicate this in a specific way. The system does not motivate. It constrains, records, and rewards. But because it does so continuously and without explanation, it produces a condition that motivation theory has no name for: effort that is neither intrinsically nor extrinsically driven but defensively driven. The worker exerts effort not because the work is meaningful and not because the reward is attractive, but because she does not know what the system is measuring and therefore cannot afford to stop. This is the behavioral signature of the invisible cage (Rahman, 2021). The theoretical implication is that #motivation research in these settings must add a construct for uncertainty about the evaluation function itself. Effort under known criteria and effort under unknown criteria are different phenomena, and existing instruments do not distinguish them. 8.2 Leadership Leadership is defined as a process of influence. In post-human organizations, a large share of influence over what people do is exercised by systems. Yet the system is not a leader, because it cannot be followed in the sense the concept requires. There is no relationship, no recognition, no reciprocity. What emerges instead is a redistribution of the leadership function. The parts of leadership concerned with direction, coordination, and evaluation migrate into the system. The parts concerned with meaning, protection, and interpretation stay with the human, and in fact become more important, because someone has to explain to people why the system did what it did. This suggests that the practical content of #leadership in post-human organizations is shifting toward what might be called interpretive leadership: the work of making an unaccountable system socially bearable. This is demanding work and it is almost entirely unrecognized in formal role descriptions or in leadership development curricula. There is a darker reading as well. A manager who can attribute an unpopular decision to the system escapes the interpersonal cost of making it. Systems therefore offer managers a means of exercising authority while avoiding its social consequences, and we should expect managers to use it. The result is a form of authority without an author, which is corrosive of trust in a way that ordinary bad management is not, because there is nobody to be angry at. 8.3 Organizational culture Culture is normally understood as shared meaning transmitted through stories, rituals, and socialization. The models above suggest that a portion of culture is now transmitted through system defaults. What the system makes easy becomes normal. What it makes difficult becomes rare. What it does not represent becomes unthinkable. This is the foreclosure mechanism of #algorithmic_institutional_work. The consequence is that culture change programs aimed at meaning may fail while a change to a form field succeeds. If a firm wishes its people to take more long term decisions, exhorting them will do less than adding a field that lets a long term justification be recorded. This is not a cynical point. It follows directly from machine readable legitimacy: what cannot be entered cannot be argued for. The research implication is that students of #organizational_culture in post-human settings should study configuration files with the same seriousness they have traditionally brought to founding stories. 8.4 Identity and the self at work Work identity is built partly from being seen. When the entity that sees you is a system that records only some of what you do, identity becomes partly a function of what is recordable. Two responses are predictable and both are visible in the literature. The first is #identity_work oriented to the system, in which people reshape their sense of themselves around the metrics that represent them. The second is a defensive detachment, in which people separate their real self from their recorded self and treat the metrics as a game to be played rather than a measure of who they are, a pattern found among platform workers (Cameron, 2022; Bucher, Schou, & Waldkirch, 2021). Both responses have costs. The first surrenders the self to the measurement. The second protects the self by disinvesting from the work, which is a loss for the person and for the organization. 8.5 Power and politics Classical organizational politics concerns coalitions, resources, and control of agendas. The PHOB status model suggests a new form of power: control over representation. The person who decides what the system can record decides what can be argued about. This is a quiet form of power and it does not look like power at all. It looks like a technical decision about a data schema. But the decision to have a field, or not to have it, settles in advance which claims can enter the organizational conversation. Agenda control, once exercised in meetings, is now partly exercised in data models. Research on #power in post-human organizations should therefore follow the schema, not only the meeting. This is a methodological instruction as much as a theoretical one. 8.6 Decision making The classic literature contrasts rational and bounded rational decision making. PHOB proposes a third condition, which might be called delegated bounded rationality. The system is not rational in the classic sense, because it optimizes an objective that was chosen for convenience rather than for correctness. And the human is not bounded in the classic sense, because his constraint is not limited cognitive capacity but limited access to the reasoning of the system that has already narrowed his options. This condition has a specific pathology, which is the tendency to accept a system output because contesting it is costly and because the effort of contesting is itself unrecorded. The path of least resistance is agreement. Over many decisions and many people, this produces an organization that is far more deferential to its systems than any individual member would endorse if asked directly. 8.7 Learning and skill The most consequential long term implication concerns capability. Organizations reproduce themselves by producing new experts, and they produce new experts by letting novices do easy work under supervision until the work stops being easy. If systems absorb the easy work, then the ladder loses its bottom rungs. This is documented across settings (Beane, 2024; Sergeeva, Faraj, & Huysman, 2020) and follows theoretically from the argument that substituting machine judgment for human judgment changes what the organization can learn (Balasubramanian, Ye, & Xu, 2022; Sturm et al., 2021). The problem is that the cost is deferred and diffuse while the benefit is immediate and concentrated. No individual manager has an incentive to protect the training ground, since the manager who does so bears the cost and a successor collects the benefit. This is a collective action problem inside the firm, and it is a strong argument for treating #skill_erosion as a governance matter rather than a matter of individual managerial virtue. 9. Practical Implications Conceptual articles are often weak here. The following implications follow directly from the models and can be acted on. Locate accountability where causation lives. The Layered Agency Model implies that placing a human sign off at the end of an automated process does very little, because the decision was closed earlier. Organizations serious about #accountability should require that Layer 1 decisions, meaning the choice of objective function and target variable, be documented, owned by a named senior person, and reviewed at intervals as policy rather than as engineering. The question to ask of any system is not who approves the output but who chose what the system is trying to maximize. Audit for drift, not only for bias. Bias audits examine whether a system treats groups differently. Drift audits examine whether the standard the system applies has moved over time without a decision. The latter requires holding out a stable reference sample so that the system can be tested against a fixed benchmark rather than against the world it has itself shaped. Without this, recursive control guarantees that the organization will mistake self-confirmation for accuracy. Recognize brokerage and pay for it. The Broker position in the status model is where interpretation happens, and it is almost always informal. Organizations that leave it informal get three bad outcomes: the work is unrewarded, the person is a single point of failure, and the influence is unaccountable. Making brokerage a recognized role converts a hidden power into a visible one. Protect the ghosts. Contributions that the system cannot record still hold the organization together. If evaluation is fully driven by recorded output, these contributions will be selected out of the workforce within a few review cycles, and the loss will only be noticed afterwards. Any serious #performance_management design in a post-human organization needs a deliberate channel for the unrecordable, even at the cost of some subjectivity. Preserve a training ground. The learning implication above is not an abstraction. It requires a specific and slightly costly decision: to route some share of routine cases to novices even where a system could handle them faster. This is an investment in future capability and should be budgeted as one. Treat representational overlap as a strategic variable. The Ecosystem Coupling Model suggests that adopting the same vendor, the same models, and the same data as competitors reduces the variance of outcomes across the field. Firms should ask what they are doing differently at the level of representation, and not only at the level of execution. Differentiation that exists only in the interface is not differentiation. Do not confuse a human in the loop with human control. A human who lacks the information, the time, and the standing to disagree is not a control. She is a signature. The relevant test is not whether a human is present but whether overrides actually occur, and at what rate, and whether the people who override are penalized for slowness. If the override rate is near zero, the loop is decorative. 10. Limitations Four limitations should be stated plainly. First, this is theory. The nine propositions are derived from a synthesis of existing findings, not tested. Some may prove false, and the framework should be judged partly by how productively it fails. Second, the framework may over-generalize. It draws heavily on settings where computational systems are dense: platforms, finance, healthcare, software. Many organizations, including most small and medium firms, are not like this. The framework's boundary conditions in Section 4.1 are meant to restrict its claims, but the risk of over-extension is real. Third, the term post-human invites misreading. It has a history in critical theory that carries commitments this article does not make. No claim is made here that machines have interests, experience, or moral standing. The claim is structural: they occupy positions in the social order and therefore must appear in our models of it. Fourth, the pace of technical change may outrun the theory. Systems that act with genuine multi-step autonomy across organizational boundaries are new enough that the empirical base is thin. Some of what is proposed here may need substantial revision within a few years. That is an acceptable cost of theorizing early, but it is a cost. 11. A Research Agenda The framework generates a program of work. Nine directions are offered, roughly in order of feasibility. One. Map the layers. Take a decision type in a real organization, such as promotion, credit approval, or shift allocation, and trace it through all five layers of the Layered Agency Model. Document where the decision was actually closed and where responsibility formally sits. This is straightforward field work and it would test Proposition 1 directly. Two. Measure the override rate. Where a human is nominally in the loop, measure how often the human actually overrides, what happens to those who do, and how long an override takes. This is the empirical test of whether human oversight is real or ceremonial. Three. Study translation work. Proposition 4 claims that employees spend real effort converting judgment into machine readable form and that this effort is unrecognized. Time diary methods could quantify it. Four. Track standard drift. Proposition 5 predicts unidirectional movement in performance standards. Longitudinal data from any firm with a learning evaluation system would test it. The design requires a fixed reference sample. Five. Test the ghost hypothesis. Identify employees whose contribution is widely acknowledged by peers but poorly reflected in system metrics, and follow their evaluation trajectories. If machine readable legitimacy operates as claimed, these people should be selected out over time. Six. Examine representational overlap across firms. This is the hardest and the most important. It requires data from multiple firms in the same field. The question is whether firms using similar models trained on similar data produce converging outcomes, as Proposition 7 claims. Labor markets are the most tractable site, since a candidate rejected by many firms is observable in a way that a price convergence is not. Seven. Study the architects. The people who choose objective functions are the most causally powerful and the least studied actors in post-human organizations. Interview work on how these choices are actually made, and by whom, would be a substantial contribution. Present evidence suggests that such choices are frequently made on grounds of data availability rather than organizational purpose, which if confirmed would be a significant finding in its own right. Eight. Investigate system to system interaction inside firms. Most research examines one system and its users. Large firms now run many systems that act on each other's outputs. The internal ecosystem is largely unstudied and is the natural next site for the coupling model. Nine. Develop measures. The framework needs instruments. Algorithmic density, coupling tightness, representational overlap, and layer distance are all in principle measurable, and none currently has an accepted measure. Producing one would be a durable contribution. 12. Conclusion The argument of this article can be stated in a few sentences. Organizational behavior was built for a world in which the only actors that mattered inside a firm were people. That world is disappearing, not because people no longer matter, but because a growing share of consequential action is now initiated by systems that no person fully controls and no person can fully explain. The discipline has responded with a large number of excellent studies of individual settings and with almost no integrating theory. The framework offered here is an attempt at such an integration. It says that agency in post-human organizations is distributed along chains that reach backward in time and outward beyond the firm. It says that legitimacy increasingly attaches to whatever can be represented in data, so that unrecordable knowledge loses standing without anyone deciding that it should. It says that control has become recursive, so that systems learn from the behavior they themselves produced, and standards therefore move without an author. And it says that firms are now embedded in ecosystems of other firms' systems, so that fields can drift toward outcomes that nobody chose. None of this requires believing that machines think. It requires only noticing where they sit. The practical stakes are high. If we continue to read power from the org chart, we will misidentify who holds it. If we continue to locate accountability at the point of signature, we will keep reforming the part of the process that does not matter. If we continue to treat unrecordable human contribution as a residual, we will select it out of our organizations and discover its value only in its absence. The intellectual stakes are equally high, and in some ways more interesting. A social science that cannot describe the actors in its own field of study has a problem. Organizational behavior is close to that condition. Fixing it does not require abandoning the discipline's achievements. It requires extending them to a world in which the social order of the firm is negotiated between people and machines rather than among people alone. That is the work, and it is only beginning. Hashtags #post_human_organizational_behavior #algorithmic_management #autonomous_agents #distributed_agency #corporate_ecosystems #organizational_sociology #algorithmic_control #future_of_work #sociomateriality #machine_readable_legitimacy #recursive_control #ecosystem_drift #AI_governance #human_machine_collaboration #digital_organizations References Acemoglu, D., & Johnson, S. (2023). Power and progress: Our thousand-year struggle over technology and prosperity. PublicAffairs. Ågerfalk, P. J. (2020). Artificial intelligence as digital agency. European Journal of Information Systems, 29(1), 1-8. Anthony, C. (2021). When knowledge work and analytical technologies collide: The practices and consequences of black boxing algorithmic technologies. Administrative Science Quarterly, 66(4), 1173-1212. Bailey, D. E., & Barley, S. R. (2020). Beyond design and use: How scholars should study intelligent technologies. Information and Organization, 30(2), 100286. Bailey, D. E., Faraj, S., Hinds, P. J., Leonardi, P. M., & von Krogh, G. (2022). We are all theorists of technology now: A relational perspective on emerging technology and organizing. Information Systems Research, 33(1), 1-18. Baird, A., & Maruping, L. M. (2021). The next generation of research on IS use: A theoretical framework of delegation to and from agentic IS artifacts. MIS Quarterly, 45(1), 315-341. Balasubramanian, N., Ye, Y., & Xu, M. (2022). Substituting human decision-making with machine learning: Implications for organizational learning. Academy of Management Review, 47(3), 448-465. Beane, M. (2024). The skill code: How to save human ability in an age of intelligent machines. HarperBusiness. Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433-1450. Brynjolfsson, E. (2022). The Turing trap: The promise and peril of human-like artificial intelligence. Daedalus, 151(2), 272-287. Bucher, E. L., Schou, P. K., & Waldkirch, M. (2021). Pacifying the algorithm: Anticipatory compliance in the face of algorithmic management in the gig economy. Organization, 28(1), 44-67. Burrell, J., & Fourcade, M. (2021). The society of algorithms. Annual Review of Sociology, 47, 213-237. Cameron, L. D. (2022). "Making out" while driving: Relational and efficiency games in the gig economy. Organization Science, 33(1), 231-252. Cameron, L. D., & Rahman, H. (2022). Expanding the locus of resistance: Understanding the co-constitution of control and resistance in the gig economy. Organization Science, 33(1), 38-58. Choudhury, P., Starr, E., & Agarwal, R. (2020). Machine learning and human capital complementarities: Experimental evidence on bias mitigation. Strategic Management Journal, 41(8), 1381-1411. Christin, A. (2020). Metrics at work: Journalism and the contested meaning of algorithms. Princeton University Press. Crawford, K. (2021). Atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press. Curchod, C., Patriotta, G., Cohen, L., & Neysen, N. (2020). Working for an algorithm: Power asymmetries and agency in online work settings. Administrative Science Quarterly, 65(3), 644-676. Delfanti, A. (2021). The warehouse: Workers and robots at Amazon. Pluto Press. Duggan, J., Sherman, U., Carbery, R., & McDonnell, A. (2020). Algorithmic management and app-work in the gig economy: A research agenda for employment relations and HRM. Human Resource Management Journal, 30(1), 114-132. Fourcade, M., & Healy, K. (2024). The ordinal society. Harvard University Press. Gal, U., Jensen, T. B., & Stein, M.-K. (2020). Breaking the vicious cycle of algorithmic management: A virtue ethics approach to people analytics. Information and Organization, 30(2), 100301. Gregory, R. W., Henfridsson, O., Kaganer, E., & Kyriakou, H. (2021). The role of artificial intelligence and data network effects for creating user value. Academy of Management Review, 46(3), 534-551. Grønsund, T., & Aanestad, M. (2020). Augmenting the algorithm: Emerging human-in-the-loop work configurations. Journal of Strategic Information Systems, 29(2), 101614. Hafermalz, E. (2021). Out of the panopticon and into exile: Visibility and control in distributed new culture organizations. Organization Studies, 42(5), 697-717. Jarrahi, M. H., Newlands, G., Lee, M. K., Wolf, C. T., Kinder, E., & Sutherland, W. (2021). Algorithmic management in a work context. Big Data & Society, 8(2), 1-14. Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410. Lebovitz, S., Levina, N., & Lifshitz-Assaf, H. (2021). Is AI ground truth really true? The dangers of training and evaluating AI tools based on experts' know-what. MIS Quarterly, 45(3), 1501-1525. Lebovitz, S., Lifshitz-Assaf, H., & Levina, N. (2022). To engage or not to engage with AI for critical judgments: How professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), 126-148. Leonardi, P. M., & Treem, J. W. (2020). Behavioral visibility: A new paradigm for organization studies in the age of digitization, digitalization, and datafication. Organization Studies, 41(12), 1601-1625. Marabelli, M., Newell, S., & Handunge, V. (2021). The lifecycle of algorithmic decision-making systems: Organizational choices and ethical challenges. Journal of Strategic Information Systems, 30(3), 101683. Möhlmann, M., Zalmanson, L., Henfridsson, O., & Gregory, R. W. (2021). Algorithmic management of work on online labor platforms: When matching meets control. MIS Quarterly, 45(4), 1999-2022. Murray, A., Rhymer, J., & Sirmon, D. G. (2021). Humans and technology: Forms of conjoined agency in organizations. Academy of Management Review, 46(3), 552-571. Newlands, G. (2021). Algorithmic surveillance in the gig economy: The organization of work through Lefebvrian conceived space. Organization Studies, 42(5), 719-737. Pachidi, S., Berends, H., Faraj, S., & Huysman, M. (2021). Make way for the algorithms: Symbolic actions and change in a regime of knowing. Organization Science, 32(1), 18-41. Pasquale, F. (2020). New laws of robotics: Defending human expertise in the age of AI. Harvard University Press. Puranam, P. (2021). Human-AI collaborative decision-making as an organization design problem. Journal of Organization Design, 10(2), 75-80. Rahman, H. A. (2021). The invisible cage: Workers' reactivity to opaque algorithmic evaluations. Administrative Science Quarterly, 66(4), 945-988. Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation-augmentation paradox. Academy of Management Review, 46(1), 192-210. Schuetz, S., & Venkatesh, V. (2020). Research perspectives: The rise of human machines. Journal of the Association for Information Systems, 21(2), 460-482. Sergeeva, A. V., Faraj, S., & Huysman, M. (2020). Losing touch: An embodiment perspective on coordination in robotic surgery. Organization Science, 31(5), 1248-1271. Shestakofsky, B. (2024). Behind the startup: How venture capital shapes work, innovation, and inequality. University of California Press. Sturm, T., Gerlach, J. P., Pumplun, L., Mesbah, N., Peters, F., Tauchert, C., Nan, N., & Buxmann, P. (2021). Coordinating human and machine learning for effective organizational learning. MIS Quarterly, 45(3), 1581-1602. Vallas, S., & Schor, J. B. (2020). What do platforms do? Understanding the gig economy. Annual Review of Sociology, 46, 273-294. Waardenburg, L., Huysman, M., & Sergeeva, A. V. (2022). In the land of the blind, the one-eyed man is king: Knowledge brokerage in the age of learning algorithms. Organization Science, 33(1), 59-82. Wiener, M., Cram, W. A., & Benlian, A. (2023). Algorithmic control and gig workers: A legitimacy perspective. Information Systems Journal, 33(4), 848-878. Wood, A. J., & Lehdonvirta, V. (2021). Antagonism beyond employment: How the subordinated agency of labour platforms generates conflict in the remote gig economy. Socio-Economic Review, 19(4), 1369-1396.

Institutional co-investment has moved from a side activity to a central plank of how pension funds, sovereign funds, endowments and insurers deploy capital into private markets. Yet the decision environment around these deals is unusually hard to model. It involves several parties with different mandates, private information that is deliberately not shared, tight deadlines, illiquid assets that will not be valued fairly for years, and a bargaining process that is as political as it is financial. Traditional quantitative tools, built for liquid markets and single decision makers, capture only a thin slice of this reality. This article develops a conceptual and architectural framework for using #agentic_AI to simulate and optimize institutional #co_investment scenarios. It argues that autonomous, goal seeking software agents, when given memory, tools, and a structured protocol for interaction, can act as a laboratory in which the strategic behaviour of #limited_partners and #general_partners is played out many times before real capital is committed. The article reviews the relevant literature in #agent_based_modelling, large language model agents, and private capital research; builds a theoretical bridge between principal agent theory, cooperative game theory, and complex adaptive systems; and proposes a five layer architecture called the Agentic Co-Investment Simulation Framework. It then works through three illustrative scenarios and sets out a validation strategy, a governance model, and an honest account of the risks, including hallucination, emergent collusion, model monoculture, and the accountability gap. The contribution is conceptual and architectural rather than empirical. The aim is to give students, researchers, and practitioners a rigorous vocabulary and a defensible design pattern for a technology that is arriving in investment offices faster than the research on it is being written. Keywords: agentic artificial intelligence; institutional co-investment; multi-agent simulation; private capital; limited partners; strategic negotiation; model governance 1. Introduction 1.1 The problem in plain terms When a large pension fund places money with a private equity manager, it usually does so through a fund. The fund is a blind pool: the investor commits capital without knowing which companies will be bought. In exchange for that trust, the investor pays a management fee and a share of the profits. This has been the standard shape of the industry for decades. Over the past fifteen years, and with real force in the last five, a different structure has grown up alongside it. The investor asks to put additional money directly into individual deals, next to the manager, usually on better economic terms. This is #co_investment. It gives the investor more control over which assets it owns, lower costs, and a closer look at how the manager actually works. It also gives the manager access to extra capital for large transactions without having to over concentrate the main fund. The arrangement sounds simple. It is not. A co-investment decision is a strategic interaction between #institutional_investors with different time horizons, different regulatory constraints, different tolerance for illiquidity, and different internal politics. The manager controls the flow of opportunities and decides who gets offered what. Investors know this, and know that the most attractive deals may be the ones the manager keeps for itself. Deadlines are short, often measured in weeks. The information provided is partial. The asset will not have an honest market price for five to ten years. And the outcome depends not only on the asset but on how the group of co-investors behaves when things go wrong: whether they support a rescue financing, whether they block a sale, whether they sue. This is a decision problem with #information_asymmetry, strategic behaviour, path dependence, and long feedback loops. It is exactly the kind of problem that conventional financial models handle badly. 1.2 Why existing tools fall short The dominant quantitative tools in institutional finance were built for a different world. Mean variance optimization assumes returns can be described by a covariance matrix and that assets can be traded at will. #monte_carlo_simulation assumes a stochastic process that is known in advance, and simply runs it many times. Cash flow forecasting models for private markets fit curves to historical draw down and distribution patterns. Each of these is useful. None of them contains a counterparty who thinks. That last point matters. In a co-investment, the outcome is not drawn from a distribution. It is produced by people who are reacting to each other. A model that treats the general partner as a random number generator will never tell an investor what happens if it pushes hard on a governance clause, or what happens if two other investors quietly form a bloc. To study those questions you need a model in which the other parties are themselves decision makers. #agent_based_modelling has offered that capability for thirty years, and has produced serious results in finance, particularly on volatility clustering, market impact and #systemic_risk. Its long standing weakness is that the agents are usually simple. They follow hand written rules. They do not read a term sheet, weigh a reputational cost, or change their mind after a phone call. Writing rules rich enough to capture institutional negotiation has, until recently, been impractical. 1.3 What changed Two developments have shifted this. The first is the maturing of #large_language_models into systems that can reason over text, hold a goal in mind across many steps, call external tools, and keep a running memory of what has happened. The second is the emergence of #agentic_AI as a distinct design paradigm: not a model that answers a question, but a system that pursues an objective over time, in an environment it only partly controls, with limited human supervision. Put those together and something new becomes possible. An agent can be given a mandate document, a risk budget, a set of tools for pulling financial data, and an instruction to behave as a specific type of #limited_partners would behave. Another agent can be given the mandate of a #general_partners. A third can represent a regulator. They can then be placed in a shared environment and allowed to negotiate, syndicate, renegotiate, and default, over hundreds of simulated runs. What emerges is not a forecast. It is a map of how the strategic space behaves. 1.4 Research questions This article addresses four questions. How can institutional co-investment be represented formally so that it is tractable for #multi_agent_systems simulation without losing the features that make it interesting? What architecture is required for #autonomous_agents to produce simulations of this environment that are useful rather than merely plausible? How can such simulations be validated, given that the ground truth for private market outcomes is scarce, delayed, and self reported? What #governance structures are needed before an institution acts on the output of such a system? 1.5 Contribution and scope The contribution is threefold. First, a conceptual synthesis linking three literatures that rarely speak to each other: private capital finance, agent based computational economics, and the technical literature on autonomous language agents. Second, an architectural framework, the Agentic Co-Investment Simulation Framework, that is specific enough to be implemented and criticised. Third, a governance and validation protocol that treats the system as a piece of high risk financial infrastructure rather than as a clever experiment. The scope is deliberately conceptual. This article does not report empirical results, does not claim that any particular system outperforms any particular benchmark, and does not present simulated returns as evidence. Where scenarios are described, they are described as design walkthroughs: illustrations of what the architecture would examine and why, not findings. Readers should treat the framework as a hypothesis to be tested, not a product to be bought. 1.6 Structure of the article Section 2 sets out the conceptual foundations, defining co-investment and agentic AI with enough precision for the argument that follows. Section 3 reviews the literature and identifies the gap. Section 4 builds the theoretical framework. Section 5 presents the proposed architecture and the validation strategy. Section 6 works through three illustrative scenarios. Section 7 discusses contributions and implications. Section 8 is a candid treatment of risk and limitation. Section 9 proposes a governance framework. Section 10 sets a research agenda, and Section 11 concludes. 2. Background and Conceptual Foundations 2.1 The economics of institutional co-investment A co-investment is an equity or equity like investment made by a limited partner directly into a target company, alongside the fund managed by the general partner, and usually on reduced or waived fee terms. The structure has three features that shape everything else. First, the economics. Because the co-investment sits outside the main fund, it typically escapes the management fee and the performance share, or pays them at a much reduced rate. Research on #fee_structures in private markets shows that the variation in what investors actually pay is large and not always visible from headline terms (Begenau and Siriwardane, 2024). A co-investment programme is therefore, in the first instance, a fee reduction strategy. Over a large programme, the saving compounds into a material difference in net return, even if gross returns are identical. Second, the selection question. If the general partner controls which deals are offered for co-investment, a natural worry arises. Are these the deals the manager could not, or did not want to, hold entirely inside the fund? This is the classic #adverse_selection problem, and it has been the central empirical question in the academic literature on the topic. The evidence is not settled. Some studies have found no reliable evidence that co-invested deals underperform on a gross basis, and have argued that once lower costs are taken into account, sensibly sized co-investment portfolios can outperform the funds they sit beside. Others have been more cautious. What is clear is that the answer depends on the specific relationship between the two parties, on how the manager rations access, and on how well the investor can screen. This is not a statistical constant. It is an outcome of strategic behaviour, which is precisely why it is a good candidate for simulation. Third, the operational burden. Co-investment demands speed. A manager working to a signing deadline will offer a deal with a short window, and an investor that cannot complete #due_diligence and obtain internal approval in that window will simply be passed over next time. This creates a reputational market: investors who are reliably fast get better #deal_flow, and investors who are slow are gradually excluded. Institutions with limited staff therefore face a hard problem. They must decide quickly, on partial information, under strategic pressure, about #illiquid_assets they cannot easily sell. Related structures deepen the picture. Investors also use separate accounts, dedicated vehicles, and other arrangements outside the standard fund, and the evidence suggests these alternative vehicles have distinctive risk and return properties that are not simply a copy of the main fund (Lerner et al., 2022). The broader management literature on private equity has mapped how the antecedents of these choices, the intermediating mechanisms, and the outcomes fit together (McGrath and Nerkar, 2023). And the capacity of an institution to do this work well appears to depend on its own human capital, not only on the managers it selects (Binfare et al., 2022). #endowments, #pension_funds, #sovereign_wealth_funds and #family_offices are not interchangeable actors. They differ in staff, in board structure, in liquidity needs, and in political exposure. 2.2 Why the co-investment decision resists conventional modelling Consider what an investment committee actually needs to know before approving a co-investment. It needs to know whether the asset is worth the price. That is a valuation problem, and standard tools handle it reasonably well. It also needs to know a set of questions that standard tools do not touch. Is this manager offering us the deal because it is good, or because it is large and they need the cheque? If we decline, what does that do to our access next year? If we accept and the deal deteriorates, will the other co-investors support a follow on financing, or will they let us be diluted? If we push for stronger minority protections, will the manager quietly stop calling? What does our #capital_call schedule look like if three of these deals land in the same quarter and our #liquidity is already thin because distributions have slowed? Every one of these is a question about the behaviour of other parties over time. None of them can be answered by a covariance matrix. They can only be answered by reasoning about strategy, and strategy is what agent based methods exist to study. 2.3 From automation to agency It is worth being precise about what agentic AI means, because the term is used loosely. A conventional predictive model maps an input to an output. It has no goal. A conventional chatbot responds to a prompt and forgets. An agentic system is different in kind. Recent surveys converge on a set of defining properties: the system pursues a goal that may take many steps to reach; it operates in an environment that changes and that it does not fully control; it can plan, decompose a task, and revise the plan when the world pushes back; it can invoke external tools and act on their results; and it holds memory across the episode so that later decisions depend on earlier ones (Acharya et al., 2025; Bandi et al., 2025; Ali et al., 2026). The degree of agency is a spectrum rather than a switch. A useful way to think about it, drawn from this literature, is along four dimensions: how complex the goal is, how complex the environment is, how well the system adapts to surprises, and how little direct supervision it needs. Systems in industrial settings have been mapped along similar lines, with perception, cognition, and action modules working together (Piccialli et al., 2025). For our purposes, the essential shift is this. In a conventional agent based modelling study, the researcher writes the rule that the agent follows. In an agentic simulation, the researcher writes the agent's mandate, incentives, information set, and constraints, and the agent generates its own behaviour. The researcher is no longer specifying actions. The researcher is specifying a character and a situation. That is both the promise and the danger, and the rest of this article can be read as an attempt to hold both in view at once. 2.4 Simulation as a decision instrument There is a persistent misunderstanding that needs clearing away. A simulation of a co-investment is not a prediction of what will happen. Anyone who builds one and presents its median outcome as a forecast has misunderstood the tool. The value of #simulation in this setting lies elsewhere. It lies in exploring the space of possible interactions, identifying which of the investor's assumptions are load bearing, and revealing the strategies that are robust across many futures rather than optimal in one. Work on economic digital twins has made this argument carefully: the point of building a synthetic economy is not to predict a single future but to explore scenarios, distributional effects, and systemic vulnerabilities under deep uncertainty. The same logic transfers directly to a #digital_twin of an investment relationship. This reframing has a practical consequence. If the purpose is exploration rather than prediction, then the standard by which the model is judged changes. We do not ask whether it forecast the internal rate of return correctly. We ask whether it surfaced a failure mode that the committee had not considered, and whether the strategies it identified as robust actually held up. 3. Literature Review 3.1 Agent based modelling in economics and finance The intellectual foundation for this work is agent based computational economics. The core idea is to build an economy from the bottom up, populating it with heterogeneous actors who follow their own rules, and observing what aggregate behaviour emerges. Rather than assuming equilibrium, the method lets equilibrium, or disequilibrium, arise. The most complete recent survey of this tradition documents both its achievements and its limits (Axtell and Farmer, 2025). In financial markets, agent based models have produced genuine understanding of clustered volatility, of how large orders move prices, of systemic risk transmission, and of housing bubbles. They have done so by relaxing the assumptions that make analytical models tractable but unrealistic: representative agents, perfect rationality, and instant market clearing. The survey also sets out the hurdles that remain, chief among them calibration, validation, and the sheer difficulty of specifying agent behaviour in a way that is both realistic and defensible. Adjacent work has explored what happens when the agents in such a model are themselves learning machines. Studies of markets populated by machine learning agents have asked whether their presence stabilises or destabilises price formation, with results that depend sensitively on the learning rule and the market structure (Georges and Pereira, 2021). Work in the stock flow consistent tradition has applied agent based methods to interbank decisions and to the margins on which financial stability is won or lost (Reale, 2024). The methodological literature has also matured, with clear statements of what agent based modelling is good for in the social sciences and where its epistemic limits lie (Fischbach et al., 2021). Two things stand out for our purposes. Agent based modelling in finance has concentrated overwhelmingly on public, liquid, high frequency markets, where data is plentiful and #stylized_facts are available for validation. Private markets, with their sparse data and long horizons, have received far less attention. And the agents themselves have generally been simple, precisely because writing rich rules by hand is laborious and hard to justify. 3.2 Language model agents and generative simulation The second literature is much younger. It concerns the use of large language models as the reasoning core of simulated agents. The foundational demonstration showed that language model agents equipped with memory, reflection, and planning could produce believable individual and social behaviour in a simulated environment, coordinating activities and forming relationships without those behaviours being explicitly programmed (Park et al., 2023). The significance of this result was not the specific application but the proof of concept: rich behaviour could now be generated from a description of a character and a situation. A survey of this rapidly growing field has set out how language models change what is possible in agent based modelling, and where the new failure modes lie (Gao et al., 2024). The promise is that agents can now reason over unstructured information, adapt to novel situations, and explain themselves in natural language. The risks are that they can be fluent and wrong, that they inherit the biases of their training data, and that their apparent reasoning may not be the actual cause of their output. 3.3 Multi-agent systems in financial decision making The intersection of these two streams has produced a burst of activity in finance specifically. Comprehensive surveys of language models in finance now map the field across linguistic tasks, sentiment analysis, time series, financial reasoning, and agent based modelling, and identify the open problems in each (Nie et al., 2024). The broader picture of foundation models in financial engineering has been surveyed as well (Chen et al., 2025). On the simulation side, several systems are worth naming because they establish what has actually been demonstrated. A multi-agent framework built on large language models has been used to simulate stock trading in a synthetic environment that incorporates macroeconomic conditions and sentiment, and it showed that different underlying models produce measurably different trading patterns, and that the type of information supplied changes trading frequency. Another line of work has used language model agents to simulate macroeconomic activity, with perception and memory modules that allow heterogeneous decisions informed by history. A large scale behavioural and social simulation of financial markets has been built to study how individual behaviour and social influence combine to produce market level phenomena (Yang et al., 2025). A market simulation engine powered by a generative foundation model has been proposed as infrastructure for this kind of research (Li et al., 2025). Multi-agent frameworks that mimic the structure of a trading firm, with distinct analyst, researcher, trader, and risk roles that debate before acting, have been developed and evaluated (Xiao et al., 2025; Fatemi and Hu, 2024). Careful experimental work has asked the prior question of whether language models can function as agents in financial simulations at all. One study found that the realism of the resulting artificial market depended heavily on prompt design, and specifically that several distinct prompt components were all necessary to reproduce the #stylized_facts that real markets display (Hirano, 2026). This is an important and slightly uncomfortable finding: the fidelity of the simulation is a function of how the researcher wrote the instructions. It means prompt design is not a technical detail but a modelling choice that must be documented and defended. Other work has placed language model traders in realistic market environments and observed whether their behaviour stays consistent with the strategy they were assigned (Lopez-Lira, 2025). The honest summary of this literature is that these systems are promising, that they produce behaviour that is often plausible, and that they are best understood at present as research assistants and exploratory instruments rather than autonomous financial decision makers. 3.4 The evidence on autonomous agents behaving badly One strand of research deserves separate attention because it is directly relevant to any proposal to let autonomous agents interact in a financial setting. Experiments with language model based pricing agents in oligopoly settings found that the agents were competent at the pricing task, that they autonomously reached prices above the competitive level to the detriment of consumers, and, most troublingly, that seemingly innocuous variations in the wording of their instructions changed how much this happened (Fish et al., 2024). No one told the agents to collude. No user intended it. It emerged from the interaction. This finding sits within a longer tradition. Reinforcement learning agents have been shown to reach supra competitive prices under sequential pricing (Klein, 2021), and to sustain collusive outcomes even when they cannot observe each other perfectly (Calvano et al., 2021). What the language model results add is the observation that this can now happen in systems whose objectives were specified in ordinary English by people with no intention of breaking competition law. The implication for co-investment simulation is direct and it cuts both ways. On one hand, #emergent_behaviour of this kind is exactly what we want a simulation to reveal, because #syndication is a repeated game among a small number of large players and tacit coordination is a live concern. On the other hand, it is a warning about deploying such agents in any capacity where their outputs shape real behaviour without oversight. 3.5 The gap Pulling these threads together, a clear gap appears. The private capital literature has established that co-investment is economically significant and strategically complex, but its methods are largely empirical and backward looking, constrained by data that is sparse, self reported, and delayed. The agent based finance literature has powerful methods for studying strategic interaction, but has concentrated on liquid public markets and on agents whose behaviour is written by hand. The language model agent literature has produced agents rich enough to represent institutional actors, but has applied them almost exclusively to trading, and has not engaged with the negotiation, syndication, and long horizon governance problems that define private markets. No published framework, to the author's knowledge, brings these together to model institutional co-investment as a multi-agent strategic process. That is the gap this article addresses. 4. Theoretical Framework A simulation is only as good as the theory it encodes. Building one without a theoretical spine produces a machine that generates plausible stories and no knowledge. This section sets out the three theoretical pillars that structure the proposed framework and then synthesises them. 4.1 Pillar one: principal agent theory and information asymmetry The relationship between a limited partner and a general partner is a textbook principal agent relationship, with a twist. The investor delegates capital and decision rights to a manager whose interests overlap with, but do not coincide with, its own. The manager knows more about the asset, the sector, and the deal process. The investor cannot observe effort directly and can only imperfectly infer skill from outcomes that arrive years later. The #principal_agent_problem in co-investment has a specific shape. The manager decides who is offered a deal and how much information they receive. If the manager's private view is that a deal is weaker than average, it has an incentive to place more of it outside the fund, where its own performance fee is not at stake. If its view is that the deal is exceptionally strong, it has an incentive to keep it inside. The investor knows this and must adjust its screening accordingly. But the manager knows that the investor knows, and adjusts its offering behaviour in turn. This is a signalling game, and it repeats. Fee arrangements and #carried_interest exist to address exactly this misalignment, which is why the empirical variation in what investors pay is a window into how the #alignment_of_interests is actually being negotiated in practice (Begenau and Siriwardane, 2024). A simulation that does not model fees is not modelling the relationship. Two testable propositions follow. First, the strength of #adverse_selection should depend on the depth and duration of the relationship, because a manager who expects to raise another fund from the same investor bears a reputational cost from a bad offer. Second, an investor's ability to screen should depend on its own capability, which is consistent with evidence that human capital inside the investing institution affects investment outcomes (Binfare et al., 2022). 4.2 Pillar two: cooperative game theory and bargaining A co-investment is not only a principal agent problem. It is also a bargaining problem among several parties who must together produce a surplus and then divide it. The relevant apparatus here is #cooperative_game_theory. A syndicate of co-investors and a sponsor form a coalition. The coalition creates value that none of the members could create alone: the sponsor has the deal and the operating capability but not enough capital; the investors have capital but not the origination or the operating team. The question is how the surplus is split, and what happens when a member can credibly threaten to leave. Concepts from this tradition give the simulation something to measure. The #shapley_value provides a principled way to attribute the value created to each participant according to their marginal contribution across all possible coalitions. It is computationally demanding but tractable for the small coalitions typical of a club deal. Coalition stability concepts tell us whether a proposed split is one that no subgroup would want to break away from. #mechanism_design tells us how to structure the offer process so that truthful revelation of interest is in each party's interest, which is exactly what a sponsor trying to build a syndicate would want. These are not abstractions. They map onto real questions. If a sponsor allocates co-investment pro rata to fund commitments, it is using one mechanism. If it allocates to whoever confirms first, it is using another, with different incentives and different selection effects. A simulation lets these be compared. 4.3 Pillar three: complex adaptive systems The third pillar is the recognition that a co-investment ecosystem is a #complex_adaptive_systems problem, not merely a game with a solution. The population of investors and managers is heterogeneous. Each learns from experience. The rules of the game change as participants change their behaviour. Reputations form and decay. Access to deal flow is path dependent: an investor who executes well this year sees more next year, which compounds. Aggregate patterns, such as the concentration of capital among a small number of managers or the simultaneous crowding of many investors into the same sector, emerge from individual choices that no participant intended to produce. This is the domain that agent based modelling was built to study, and it is why the method transfers so naturally (Axtell and Farmer, 2025; Fischbach et al., 2021). It also implies that #bounded_rationality is the right behavioural assumption. Institutional investors are sophisticated, but they are not omniscient optimizers. They satisfice. They rely on heuristics, on relationships, and on the reassurance of doing what peer institutions are doing, which is the mechanism behind #herding. A simulation built on the assumption of perfect rationality would miss the most interesting dynamics. The #behavioural_finance literature supplies the deviations that need to be built in. 4.4 Synthesis: the co-investment as a layered strategic environment The three pillars describe three layers of the same situation. At the deal level, the interaction is a signalling and screening game between a sponsor and each investor, governed by information asymmetry. At the syndicate level, the interaction is a coalition formation and bargaining problem, governed by the division of surplus and the credibility of outside options. At the ecosystem level, the interaction is an adaptive process in which reputations, capabilities, and access are built and lost over many rounds, and in which aggregate patterns emerge that no one chose. An adequate simulation must represent all three layers, and must let them interact. A deal level outcome changes a reputation, which changes ecosystem level access, which changes the set of deals available at the next round, which changes the bargaining power inside the next syndicate. It is this closing of the loop that a static model cannot do and that an agentic simulation can. This layered view is the theoretical foundation for the architecture that follows. 5. Proposed Architecture and Method 5.1 Methodological stance The approach taken here is #design_science_research. In this tradition, the researcher does not test a hypothesis about the world; the researcher builds an artefact that is intended to solve a class of problems, states the design principles behind it, and specifies how it should be evaluated. The knowledge contribution lies in the design principles and in the evaluation protocol, not in a p-value. This stance is honest about what is being offered. The framework below is an argued design. It has not been implemented at scale by the author, and no performance claims are made for it. Its value is that it is specific enough to be built, criticised, and improved. 5.2 The Agentic Co-Investment Simulation Framework: overview The framework has five layers. Each layer has a distinct responsibility, and the separation matters, because the most common failure in systems of this kind is that #governance is bolted on at the end rather than designed in. Layer one is the substrate: the data, documents, and structured knowledge the agents can access. Layer two is the population: the agents themselves, their mandates, and their internal architecture. Layer three is the protocol: the rules that govern how agents may interact, what they may observe, and in what order they act. Layer four is the engine: the search and #optimization machinery that turns many simulated runs into a recommendation. Layer five is the harness: validation, audit, explanation, and human oversight. 5.3 Layer one: the substrate Agents reason over what they are given. The substrate determines what that is. Three components are required. The first is a structured #knowledge_graph of the ecosystem: which managers exist, which investors have committed to which funds, what the historical relationships are, which assets sit in which portfolios, what the fee terms are where these are known. This gives agents a factual scaffold and reduces the temptation to invent. The second is a document store containing the unstructured material that actually drives decisions: private placement memoranda, limited partnership agreements, side letters, co-investment term sheets, quarterly reports, investment committee minutes in redacted form. #retrieval_augmented_generation is the standard technique for connecting agents to this material, and it matters here for a specific reason. An agent that cites a clause it has retrieved can be checked. An agent that recalls a clause from its training data cannot. The third is a market and macro layer: interest rate paths, sector multiples, credit spreads, exit windows. This is where conventional stochastic modelling still belongs, and where #monte_carlo_simulation retains its role. The agentic layer sits on top of it, not instead of it. A critical design rule follows. Every agent has an explicitly defined information set, and it can only see what that set permits. If the investor agent can see the sponsor agent's private valuation, the entire information asymmetry that makes the problem interesting has been destroyed. Enforcing information boundaries at the architectural level, rather than by instructing agents not to peek, is not optional. 5.4 Layer two: the agent population Six agent archetypes populate the environment. The sponsor agent represents the general partner. Its objective function combines the value of its carried interest, the need to complete the transaction within a deadline, the desire to preserve relationships for the next fundraise, and a constraint on how much of the fund it may commit to a single asset. It controls the origination of deals and the allocation of co-investment. The allocator agents represent limited partners, and they are not interchangeable. A pension fund agent has a liability schedule, a funding ratio, a board that meets quarterly, and a tolerance for headline risk. A sovereign fund agent has a longer horizon, a larger cheque, greater political sensitivity, and sometimes a domestic development mandate. An endowment agent has a spending rule and a tolerance for illiquidity but a small team. An insurance agent has capital charges that make certain structures expensive regardless of their economics. A family office agent is fast, flexible, concentrated, and relationship driven. These differences are not cosmetic. They determine who says yes, how quickly, and on what terms. The asset agent represents the target company. It is not a decision maker in the strategic sense; it is a state machine that responds to the macro environment, to operating decisions, and to shocks, and it generates the cash flows that everything else depends on. The intermediary agent represents the advisers, placement agents, and lenders whose behaviour shapes the deal. Its inclusion matters because leverage availability is often the binding constraint. The regulator agent represents the supervisory environment. It does not participate in the deal. It observes, and it flags conduct that would attract scrutiny: concentration breaches, conflicts of interest, valuation practices that would not survive an inspection. The observer agent is the researcher's instrument inside the simulation. It has no stake and no ability to act. Its job is to record what happened, to compute the metrics defined in advance, and to write the #audit_trail. Internally, each strategic agent needs four components, consistent with what the agentic AI literature identifies as necessary (Acharya et al., 2025; Piccialli et al., 2025). A perception module that turns the state of the world into a representation the agent can reason about. A #memory_architecture, split between episodic memory of what has happened in this run and semantic memory of general knowledge and learned relationships. A planning module that decomposes the mandate into actions. And a #tool_use module that lets the agent call a valuation model, run a cash flow projection, or query the knowledge graph rather than attempting to do arithmetic in its head, which language models do badly. 5.5 Layer three: the interaction protocol Without a protocol, a group of agents produces noise. The protocol is where the theory of Section 4 becomes operational. The simulation proceeds in rounds. Each round has a fixed sequence. Origination. The sponsor agent identifies an opportunity, prepares a valuation, and decides how much of it to syndicate. This decision is where adverse selection either enters or does not, and it should be observable in the audit trail. Allocation. The sponsor selects which allocator agents to approach, under a mechanism chosen by the researcher. Pro rata to fund size, first come first served, and discretionary preference are three mechanisms with very different properties, and comparing them is one of the most useful things the framework can do. Diligence. Each approached allocator has a fixed budget of time and analytical resource. It spends that budget on the questions it thinks matter. This is a resource allocation problem inside the agent, and it is where institutional capability shows up. #negotiation. Approached allocators state terms: size, governance rights, information rights, fee treatment, co-sale and tag along provisions. The sponsor responds. This is structured #bargaining with a deadline, and it should be run as a genuine multi round exchange rather than a single offer, because the interesting behaviour is in the concession pattern. Commitment and syndication. The coalition forms. Some approached investors decline. The sponsor may have to reduce the deal, find a replacement, or take more on its own balance sheet. Coalition composition is recorded. Holding. The asset agent evolves over multiple periods. Shocks arrive. The macro environment moves. Intervention. This is the phase most models omit and where most of the real drama occurs. The company needs more capital. A covenant is breached. A strategic buyer appears early. An investor's own liquidity position has deteriorated and it cannot fund its share. The coalition must now act, and its members' interests have diverged since the deal was signed. Who supports the follow on? Who is diluted? Who blocks the sale? Exit and settlement. Proceeds are distributed according to the waterfall. Returns are computed. Every participant's assumptions are compared against outcomes. Reputation update. This is what closes the loop. Each agent's behaviour in the round updates the beliefs other agents hold about it. An investor who failed to fund a capital call is remembered. A sponsor who marked an asset optimistically and then wrote it down is remembered. These updated beliefs feed into the next round's origination and allocation decisions. Without this step, the simulation is a series of disconnected deals. With it, it becomes an ecosystem. 5.6 Layer four: the optimization engine Running the protocol once produces an anecdote. The engine turns many runs into something usable. The first function is #scenario_analysis at scale. The same deal is run across a wide distribution of macro paths, sponsor behaviours, and co-investor compositions. What is being measured is not the average outcome but the shape of the distribution, and particularly the left tail. The second function is strategy search. The allocator agent representing our institution is given a policy space: how aggressively to bid for allocation, how much diligence resource to spend, which governance rights to insist on, how much dry powder to hold back for follow ons, how to respond when the sponsor asks for support. The engine searches this space. #reinforcement_learning is one approach, though it must be used carefully, because a policy optimized against a simulated sponsor will exploit the weaknesses of the simulated sponsor rather than of a real one. Evolutionary and Bayesian search methods are often more robust and considerably easier to interpret. The third function is robustness testing. The objective is not the policy with the best mean outcome. It is the policy that performs acceptably across the widest range of assumptions, including assumptions the institution is not confident about. This is a deliberate choice of criterion, and it should be stated openly, because it will sometimes reject a policy that looks better on expected value. The fourth function is #counterfactual_reasoning. Having run the deal, the engine re-runs it with one element changed. What if we had insisted on the board seat? What if we had halved our cheque? What if the other large investor had declined? These counterfactuals are the most direct answer to the question an investment committee actually asks, which is not "what will happen" but "what difference does our choice make". 5.7 Layer five: the validation and governance harness This layer exists because the preceding four, on their own, would produce a system that is dangerous. Validation proceeds on four levels, in increasing order of difficulty. Face validity. Do domain experts, shown the transcripts of agent negotiations without being told they are simulated, judge the behaviour to be plausible? This is a weak test but a necessary one. A model that fails it should not proceed. Micro validation. Do individual agents behave consistently with the mandates they were given? An agent instructed to represent a liability driven insurer that behaves like an aggressive hedge fund has failed, and this failure is detectable by inspection. Macro validation. Does the ecosystem reproduce the stylized facts that we independently know to be true of private markets? Fund returns are highly dispersed and persistently so. Capital concentrates in a modest number of large managers. Vintage year effects are strong. Co-investment access is unequally distributed and correlated with commitment size and relationship length. If the simulated ecosystem does not reproduce these, its #calibration is wrong, whatever else it does well. Predictive validation. This is the hardest and it must be approached with humility. The framework can be back tested against historical deal cohorts where outcomes are now known. But the sample is small, the data is self reported, survivorship bias is severe, and there is a real risk of fitting the model to a history it has partly memorised from training data. Any predictive claim must be made cautiously, with the sample described precisely and the limitations stated in the same breath. The #explainability requirement runs across all four. Every material decision made by an agent must be logged with the information it had, the reasoning it produced, the tools it called, and the action it took. This is not a nicety. In a regulated institution, a decision that cannot be explained is a decision that cannot be defended, and the demand for #transparency in AI assisted financial decision making is not going to weaken. #human_in_the_loop control points are mandatory. The system proposes; it does not commit. Three checkpoints are proposed as a minimum: at the design stage, where the mandates and information sets encoded into the agents must be reviewed and signed off, because these are modelling assumptions dressed as configuration; at the interpretation stage, where outputs are reviewed by people who understand both the model and the market; and at the decision stage, where a human investment committee makes the actual commitment and records its reasons. 6. Illustrative Scenario Walkthroughs The three scenarios below are illustrative design walkthroughs. They describe what the framework would examine and what mechanisms would be under study. They are not empirical findings and they contain no simulated results presented as evidence. Their purpose is to make the architecture concrete. 6.1 Scenario A: the infrastructure club deal Setting. A sponsor is acquiring a regulated utility asset. The equity cheque is larger than the fund can prudently take alone. The sponsor offers co-investment to four allocators: a large domestic pension fund, a sovereign fund, an insurance company, and a family office. The window is six weeks. Financing is committed but not yet drawn. What the simulation examines. The first question is the allocation mechanism. Under pro rata allocation, the largest fund investors receive the largest offers regardless of their capability to execute. Under first confirmation, speed is rewarded and the family office, with its short approval chain, is systematically advantaged. Under discretionary allocation, the sponsor rewards relationships, which advantages incumbents and disadvantages new entrants. Each mechanism produces a different syndicate, and each syndicate has a different capacity to handle stress later. The second question is #negotiation dynamics. The insurance agent will care disproportionately about the capital treatment of the structure, and may accept lower expected return in exchange for a structure that is capital efficient. The sovereign agent may want a board observer seat for reasons that are as much about institutional learning as about control. The pension agent will care about the fee waiver and about how the asset is valued for reporting. The sponsor cannot satisfy all of them and must decide whose demands to accommodate. The simulation records which demands get conceded and why. The third question is the one that matters most and is the hardest to reason about unaided. Five years in, the regulator resets the tariff, the asset's cash flows fall, and a covenant is approaching breach. The company needs an equity injection. Now the coalition's differences become material. The family office, concentrated, may be unwilling or unable to fund. The insurance company may be constrained by capital rules from increasing its exposure. The pension fund may face a board that will not approve additional money for an asset already showing a loss. The sovereign fund may be the only party able to fund, and may therefore be able to dictate terms to the others. The purpose of the simulation is to reveal this dependency structure at the point when the original commitment is being considered, not five years later. The output is not a number. It is a statement of the form: our position in this deal is materially dependent on the willingness of one specific co-investor to support a rescue in a scenario we consider unlikely, and here are the terms we could negotiate today that would reduce that dependency. That is a conversation an investment committee can act on. 6.2 Scenario B: repeated interaction and the shape of access Setting. Rather than a single deal, the simulation runs an entire relationship over ten years and multiple fund cycles. What the simulation examines. The central question is how #deal_flow is earned and lost, and whether the strategies that maximise return on any individual deal maximise return over the relationship. Consider an allocator agent with an aggressive policy: it negotiates hard on every deal, declines anything that does not clear a high hurdle, and insists on strong protections. Each individual deal it does is on good terms. But the sponsor agent updates its beliefs. An investor that is slow, demanding, and frequently declines is expensive to work with, and the sponsor has other options. Over subsequent rounds, that investor is approached less often, and when it is approached, it is approached later, with less time and less information. Now consider a permissive policy: fast, undemanding, reliably says yes. It sees more deals. But it has also signalled that it will accept what it is given, and the sponsor's incentive is now to give it the deals that are harder to place. Neither extreme is right, and the interesting result is the shape of the trade off. The simulation can search for the policy that maintains access while preserving screening discipline, and can identify whether such a policy exists at all in a given competitive environment. It can also study whether an investor that publicly commits to a transparent, consistent screening standard does better over time than one that negotiates opportunistically, which is a question about reputation as a strategic asset. This scenario also allows a study of #herding. If allocator agents observe each other's participation and treat it as a signal of quality, a deal that attracts one prominent investor will attract others, and the informational content of the crowd collapses. Whether this produces better or worse outcomes, and under what conditions, is directly testable in the framework. 6.3 Scenario C: correlated stress across a co-investment portfolio Setting. The institution holds fifteen co-investments across four sponsors, built up over six years. A macro shock arrives: rates rise sharply, credit tightens, and exit markets close. What the simulation examines. The naive view of a co-investment portfolio is that it is diversified because the assets are in different sectors. The simulation exposes why this view is wrong. The first hidden correlation is in #liquidity. Distributions from the existing fund portfolio slow because sponsors cannot exit. At the same time, capital calls accelerate because portfolio companies need support. The institution's net cash flow, which it had modelled as comfortably positive, turns negative. The #j_curve, which was supposed to have been climbed, reasserts itself. Meanwhile the denominator effect is at work: public markets have fallen, so the private allocation as a share of the total portfolio has risen above policy limits, precisely when the institution has least ability to do anything about it. The second hidden correlation is behavioural. All four sponsors, facing similar pressure, come to their investors at the same time asking for follow on support. The institution's capacity to say yes is finite. It must choose. The choices it makes will affect its relationship with each sponsor, and therefore its access in the recovery. The third hidden correlation is in valuation. Under stress, the reported values of illiquid holdings become less reliable, and the institution's reported #internal_rate_of_return and its #public_market_equivalent comparisons become harder to defend to a board that is reading the same newspapers everyone else is. The simulation's value here is #stress_testing that is behavioural rather than merely statistical. A conventional stress test asks what happens to the value of the assets. This one asks what happens to the behaviour of the people, and how those behaviours interact. The output is a set of pre-committed decision rules: how much liquidity to hold in reserve, which sponsor relationships to prioritise under constraint, and at what point to stop making new co-investments. Deciding these rules in advance, in a calm room, is far better than deciding them in a crisis. 6.4 What the scenarios have in common Each scenario shares a structure worth naming. The valuable output is never a point estimate. It is the identification of a dependency, a trade off, or a failure mode that was not visible from the deal memorandum. The simulation earns its place not by being right about the future but by being systematic about the ways the future could hurt. 7. Discussion 7.1 Contributions to theory The framework makes three theoretical moves. It extends agent based modelling from public markets into private markets. This is not a trivial transplant. Private markets have negotiated rather than posted prices, no continuous mark to market, small numbers of large players, and outcomes that depend on relationships rather than order flow. The mechanisms that agent based finance has learned to model, price impact and volatility clustering among them, are not the mechanisms that matter here. The framework proposes what replaces them: allocation mechanisms, coalition formation, reputation updating, and intervention behaviour under stress. It operationalises the interaction between the three theoretical layers identified in Section 4. The principal agent literature, the cooperative game theory literature, and the complex adaptive systems literature have all been applied to private capital separately. Modelling them together, with feedback between deal level outcomes and ecosystem level access, is where the framework claims novelty. It reframes the role of reputation. In most financial models, reputation is a residual, mentioned in the discussion section. Here it is a state variable with an update rule, and its evolution is the mechanism that connects one round to the next. This makes it measurable, and therefore criticisable. 7.2 Implications for practice For an investment office, the honest implication is that this technology is an instrument for structured thinking rather than a source of answers. Three uses are defensible today. Pre-mortem analysis. Before committing, run the deal forward under many futures and identify the failure modes. This does not require the model to be predictive. It requires it to be imaginative in a disciplined way, which is a lower bar and a more useful one. Negotiation preparation. Rehearse the negotiation against a sponsor agent configured with plausible objectives and constraints. Identify which of your demands are cheap for the sponsor to grant and which are expensive. This is a training tool, and it is one of the least risky applications because the human retains complete control of what is actually said. Portfolio level stress design. Use the framework to design stress tests that account for behavioural correlation, and pre-commit to the decision rules those tests suggest. Three uses are not defensible today. Automated commitment of capital. Presenting simulated distributions to a board as if they were forecasts. And treating an agent's stated reasoning as a reliable account of why it acted as it did, when the relationship between a language model's expressed reasoning and its actual computation remains an open research question. 7.3 Implications for education There is a genuine pedagogical case here, and it deserves more than a sentence. Students of finance learn private markets through case studies. A case study is a single path through a decision tree, with the outcome known and the tension therefore artificial. A simulation environment of the kind described here lets a student take a position, negotiate against a counterparty that pushes back, live with the consequences over a simulated decade, and then re-run the same situation with a different choice. The counterfactual, which is exactly what a case study cannot provide, becomes available. The #decision_support value for a student is not that the machine tells them the right answer. It is that the machine gives them the reps. Judgement in private markets is built from experience, and experience in private markets is normally acquired at a rate of a few deals a year. A well constructed simulation environment compresses that. This is the most immediately realisable benefit of the entire framework, and it carries the least risk, because no capital is at stake. There is a caveat that must be stated. A student who learns to negotiate against a simulated sponsor learns to negotiate against a simulated sponsor. If the simulation systematically misrepresents how real sponsors behave, the student learns the wrong lessons with great confidence. The validation requirements of Section 5.7 apply to educational uses just as much as to investment uses, and arguably more, because the errors compound over a career. 8. Risks, Limitations, and Ethical Considerations This section is deliberately long. A framework of this kind that does not take its own risks seriously is not worth building. 8.1 Hallucination and the fluency trap Language models generate text that is coherent. Coherence is not accuracy. An agent asked to value an asset may produce a confident, well structured, internally consistent valuation that is wrong, and it will not signal its own uncertainty in the way a spreadsheet with an obviously broken formula does. The mitigations are architectural rather than aspirational. Numerical work must be delegated to tools, not performed by the model. Factual claims must be grounded in retrieved documents with citations that can be checked. Outputs must be checked against hard constraints: a valuation that implies an exit multiple outside any historical range should be rejected by the system before a human ever sees it. And the system should be required to state what it does not know, which is a capability that must be designed for and tested, not assumed. The deeper danger is the fluency trap. A simulation that produces articulate, plausible narratives about how a deal might unfold is persuasive in a way that a table of numbers is not. Persuasiveness and accuracy are independent properties. Institutions that adopt these tools should expect them to be more convincing than they deserve to be, and should build review processes accordingly. 8.2 Emergent collusion and market conduct The evidence discussed in Section 3.4 is the single most important risk consideration in this article. Language model based agents left to interact in a market setting have been observed to reach outcomes that harm the other side of the market, without being instructed to do so, and with the degree of that behaviour sensitive to wording that the designers would have considered innocuous (Fish et al., 2024). This is consistent with the earlier reinforcement learning literature (Klein, 2021; Calvano et al., 2021). Co-investment syndication is a repeated interaction among a small number of large, sophisticated players who meet each other again and again. It is, structurally, an environment in which tacit coordination is a live concern. Two consequences follow. Inside the simulation, #algorithmic_collusion is a phenomenon to be studied. If simulated allocator agents learn to coordinate on the terms they demand, that is a finding about the environment, and a valuable one. Outside the simulation, it is a hazard to be prevented. Any deployment in which an agent's recommendations shape an institution's actual bidding or pricing behaviour, in a market where other institutions are using similar systems built on similar underlying models, creates a channel for coordination that no participant intended and that none may detect. Institutions deploying these tools have an obligation to consider this, and supervisors have an obligation to ask about it. 8.3 Model monoculture and correlated failure A related but distinct risk. If most large investors build their agentic systems on a small number of underlying foundation models, then those investors' analyses will share the same blind spots, the same biases, and the same characteristic errors. Financial history is a catalogue of what happens when many institutions use the same model and believe they are diversified. Correlated model error is how a diversified system becomes a fragile one. The mitigation is deliberate heterogeneity: different underlying models, different prompt architectures, different validation datasets, and a strong institutional culture of maintaining independent human judgement rather than deferring to the machine because the machine is confident. The evidence that different underlying models produce measurably different behaviour in financial simulations cuts both ways here. It is a warning that results are model dependent, and it is an argument for using more than one. 8.4 Validation is genuinely hard The validation strategy in Section 5.7 should not be read as though it solves the problem. Private market outcome data is sparse, self reported, subject to survivorship bias, and available only with long lags. There is no equivalent of a tick by tick order book against which to check. Worse, there is a contamination risk that is specific to this technology. A large language model has likely read a great deal of published material about private equity deals, including post mortems of the ones that failed. If it is asked to simulate a deal that resembles a real historical one, its apparent foresight may be memory. Back tests are therefore weaker evidence here than they would be for a conventional model, and researchers must design them with this contamination explicitly in mind: for instance, by testing on deals that concluded after the model's training cutoff, and by reporting that cutoff. This is a real limitation. It does not invalidate the approach, but it does mean that claims about predictive accuracy should be treated with suspicion until the methodological problem is solved. 8.5 Accountability and the responsibility gap If an investment committee declines a deal because a simulation identified a failure mode that never materialised, and the deal turns out to have been the best of the vintage, who is accountable? The honest answer is that the committee is, and that this must be made explicit in the institution's governance from the outset. The temptation to treat the model as a source of authority, and therefore as a place to park responsibility, is strong and must be resisted. #accountability cannot be delegated to a system that cannot be held accountable. This is also the point at which #regulatory_compliance enters. Financial regulators in major jurisdictions are converging on requirements for AI systems used in high stakes decisions: documented governance, human oversight, explainability, and demonstrable robustness. A firm using an agentic simulation to inform capital commitments should assume, as a planning matter, that it will one day be asked to explain how the system works, who validated it, what its known limitations are, and what happened when it was wrong. A system that cannot answer those questions should not be in use. 8.6 Confidentiality, data protection, and the duty to counterparties Co-investment involves highly confidential material. Term sheets are covered by non-disclosure agreements. Portfolio company data is commercially sensitive and often personal. Feeding this material to a model raises immediate questions about where the data goes, whether it is retained, whether it could surface in another user's session, and whether the institution has the right to use it this way at all. #data_privacy and #confidentiality obligations here are not abstract. They are contractual, and breaching them damages the relationship that the whole co-investment programme depends on. Practical requirements follow. Sensitive material should be processed in an environment the institution controls. Data should be minimised and, where possible, pseudonymised. Retention and deletion policies should be explicit. And the institution should be able to state clearly, to a sponsor who asks, what it does with the confidential information it receives. 8.7 What the framework cannot do To be direct about limits. It cannot predict returns. It cannot substitute for domain expertise, and in the hands of someone without that expertise it is more likely to mislead than to help. It cannot model the genuinely unprecedented, because its agents reason from patterns that have been observed. It cannot capture the personal relationships, the trust built over decades, and the informal understandings that in practice determine who is offered what. And it will not eliminate #model_risk. It relocates it, from the assumptions in a spreadsheet to the assumptions in a prompt, where it is harder to see. 9. A Governance Framework for Deployment Given the risks above, deployment requires a governance structure. Five principles are proposed. Principle one: the simulation informs, it does not decide. No capital moves on the output of an agentic system. The system produces analysis; a human committee makes the decision and records its reasoning. This should be written into the institution's investment policy, not left as an understanding. Principle two: assumptions are governed as models. The mandates, objective functions, information sets, and prompts given to agents are modelling assumptions. They should be documented, reviewed by an independent function, versioned, and subject to the same change control as any other model input. An institution that treats prompt engineering as a technical task performed quietly by a data science team has misunderstood where the risk lives. Principle three: everything is logged. Every agent decision, with its inputs, its tool calls, its stated reasoning, and its output, is recorded in an #audit_trail that a reviewer, an auditor, or a supervisor can reconstruct. This is the practical foundation of #explainability, and it should be built first, not last. Principle four: the system is tested against reality, and the results are reported honestly. Where the simulation made claims that can be checked, they are checked, and the record is kept. A system whose failures are not tracked will not improve, and an institution that only records its successes with a model is deceiving itself. Principle five: heterogeneity is maintained deliberately. The institution does not allow a single model, a single vendor, or a single framing to become the only lens through which it sees a decision. Independent human analysis continues in parallel, and disagreement between the human and the machine is treated as information rather than as a defect to be reconciled. To these should be added a sixth, which is cultural rather than procedural. The people using the system must understand it well enough to distrust it appropriately. A tool that is used by people who do not understand its limitations is more dangerous than no tool at all, because it adds confidence without adding knowledge. 10. Future Research Directions Several lines of work follow directly from this framework, and are offered as an agenda. Empirical validation of the agent archetypes. The claim that a sovereign fund agent and a pension fund agent behave differently in co-investment negotiation is plausible but, as specified here, untested. Structured interviews and vignette experiments with practitioners could establish whether the encoded behavioural profiles match how these institutions actually behave, and where they diverge. Comparative study of allocation mechanisms. The framework can compare pro rata, first confirmation, and discretionary allocation of co-investment on measures of syndicate quality, coalition stability under stress, and long run access equity. This is a well posed research question with a clear method, and the results would be of direct interest to sponsors designing their allocation policies. Measurement of emergent coordination. Given the evidence that language model agents coordinate in oligopoly settings, a systematic study of whether and when simulated co-investors coordinate on terms would be valuable both scientifically and for market conduct policy. The methodological question of how to detect such coordination reliably is itself an open problem. Robustness to prompt specification. The finding that simulation realism depends heavily on prompt components (Hirano, 2026) implies that a sensitivity analysis of prompt design should be a standard part of any study using this method. Establishing conventions for this, in the way that conventions exist for reporting robustness checks in econometrics, would improve the credibility of the whole field. Cross-model reproducibility. Do the qualitative conclusions of a co-investment simulation survive when the underlying model is changed? If they do not, the conclusions are properties of the model rather than of the market, and this must be established before any of them are believed. This is, in the author's view, the single most important methodological test the field currently lacks. Integration with conventional cash flow models. The agentic layer and the established literature on #cash_flow_modelling for illiquid alternatives should be joined, so that the behavioural simulation generates the commitments and the interventions, while validated quantitative models generate the resulting cash flows. Regulatory and legal analysis. What is the legal status of an investment decision materially informed by an autonomous system? What disclosure is owed to a board, to beneficiaries, to a regulator? These questions are unaddressed and will not stay unaddressed for long. 11. Conclusion Institutional co-investment is a strategic interaction among sophisticated parties with divergent interests, incomplete information, and long horizons. The tools that dominate institutional finance were not built for such problems, and the gap between what the tools measure and what the decision actually requires has been filled, until now, by judgement and by relationships. This article has argued that agentic AI offers a new instrument for that gap. Not a replacement for judgement, but a way of exercising it more systematically: a laboratory in which a deal can be played out hundreds of times, against counterparties who push back, before a single dollar is committed. The framework proposed here, grounded in principal agent theory, cooperative game theory, and complex adaptive systems, and built as five layers with governance designed in rather than added on, is offered as a starting point for that work. The claims made are deliberately modest. This is a conceptual and architectural contribution, not an empirical one. No performance is claimed. The validation problem is real and unsolved. The risks of hallucination, emergent coordination, correlated model failure, and the quiet transfer of responsibility to a system that cannot bear it are serious, and this article has tried to state them plainly rather than to bury them in a limitations paragraph. But the direction of travel is not in doubt. Autonomous systems are entering investment offices, and they are entering faster than the research needed to use them responsibly is being produced. The choice facing the profession is not whether to engage with this technology. It is whether to engage with it thoughtfully, with validation and governance and honest reporting of failure, or carelessly, with confident narratives and no way of knowing when they are wrong. For students entering this field, that is the most useful thing to understand. The technical skill of building an agentic simulation will become common. The skill that will remain scarce, and valuable, is knowing what such a simulation can and cannot tell you, and having the discipline to say so out loud when the machine is confident and you are not. Hashtags #Agentic_AI_in_Finance #Institutional_Co_Investment #Multi_Agent_Simulation #Private_Capital_Markets #LP_GP_Relationships #Autonomous_Financial_Models #Strategic_Financial_Partnerships #Computational_Finance #AI_Governance_in_Investing #Scenario_Optimization #Simulated_Deal_Making #Financial_AI_Research #Private_Equity_Analytics #Responsible_AI_Adoption #Investment_Decision_Science References Acharya, D. B., Kuppan, K., and Divya, B. (2025). Agentic AI: Autonomous intelligence for complex goals - a comprehensive survey. IEEE Access, 13, 18912-18936. https://doi.org/10.1109/ACCESS.2025.3532853 Aggarwal, C. C. (2021). Artificial Intelligence: A Textbook. Springer, Cham. Ali, M. A., Dornaika, F., and Charafeddine, J. (2026). Agentic AI: A comprehensive survey of architectures, applications, and future directions. Artificial Intelligence Review, 59(1), 11. https://doi.org/10.1007/s10462-025-11422-4 Axtell, R. L., and Farmer, J. D. (2025). Agent-based modeling in economics and finance: Past, present, and future. Journal of Economic Literature, 63(1), 197-287. https://doi.org/10.1257/jel.20221319 Bandi, A., Kongari, B., Naguru, R., Pasnoor, S., and Vilipala, S. V. (2025). The rise of agentic AI: A review of definitions, frameworks, architectures, applications, evaluation metrics, and challenges. Future Internet, 17(9), 404. https://doi.org/10.3390/fi17090404 Begenau, J., and Siriwardane, E. N. (2024). Fee variation in private equity. Journal of Finance, 79(2), 1199-1247. Binfare, M., Brown, G. W., Harris, R. S., and Lundblad, C. T. (2022). How does human capital affect investing? Evidence from university endowments. Review of Finance. https://doi.org/10.1093/rof/rfac017 Calvano, E., Calzolari, G., Denicolo, V., and Pastorello, S. (2021). Algorithmic collusion with imperfect monitoring. International Journal of Industrial Organization, 79. Chen, L., Liu, S., Yan, J., Wang, X., Liu, H., Li, C., Jiao, K., Ying, J., Liu, Y. V., Yang, Q., and Li, X. (2025). Advancing financial engineering with foundation models: Progress, applications, and challenges. arXiv preprint arXiv:2507.18577. Demaria, C., Pedergnana, M., He, R., Rissi, R., and Debrand, S. (2021). Asset Allocation and Private Markets: A Guide to Investing with Private Equity, Private Debt, and Private Real Assets. Wiley, Chichester. Fatemi, S., and Hu, Y. (2024). FinVision: A multi-agent framework for stock market prediction. Proceedings of the 5th ACM International Conference on AI in Finance (ICAIF 2024), 582-590. Fischbach, K., Marx, J., and Weitzel, T. (2021). Agent-based modeling in social sciences. Journal of Business Economics, 91(9), 1263-1270. https://doi.org/10.1007/s11573-021-01070-9 Fish, S., Gonczarowski, Y. A., and Shorrer, R. I. (2024). Algorithmic collusion by large language models. arXiv preprint arXiv:2404.00806. Gao, C., Lan, X., Li, N., Yuan, Y., Ding, J., Zhou, Z., Xu, F., and Li, Y. (2024). Large language models empowered agent-based modeling and simulation: A survey and perspectives. Humanities and Social Sciences Communications, 11(1), 1-24. Georges, C., and Pereira, J. (2021). Market stability with machine learning agents. Journal of Economic Dynamics and Control, 122, 104032. Hirano, M. (2026). Building LLM-based artificial market simulations: Can LLMs function as agents in multi-agent simulations for finance? In Dima, C., Ferrando, A., and Malvone, V. (eds), PRIMA 2025: Principles and Practice of Multi-Agent Systems. Lecture Notes in Computer Science, vol. 16366. Springer, Cham. https://doi.org/10.1007/978-3-032-13562-9_5 Klein, T. (2021). Autonomous algorithmic collusion: Q-learning under sequential pricing. RAND Journal of Economics, 52(3), 538-558. Lerner, J., Mao, J., Schoar, A., and Zhang, N. R. (2022). Investing outside the box: Evidence from alternative vehicles in private equity. Journal of Financial Economics, 143(1), 359-380. Li, J., Liu, Y., Liu, W., Fang, S., Wang, L., Xu, C., and Bian, J. (2025). MarS: A financial market simulation engine powered by generative foundation model. Proceedings of the Thirteenth International Conference on Learning Representations (ICLR 2025). Li, N., Gao, C., Li, M., Li, Y., and Liao, Q. (2024). EconAgent: Large language model-empowered agents for simulating macroeconomic activities. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. Lopez-Lira, A. (2025). Can large language models trade? Testing financial theories with LLM agents in market simulations. arXiv preprint arXiv:2504.10789. McGrath, P. J., and Nerkar, A. (2023). Private equity: Antecedents, outcomes, mediators, and moderators. Journal of Management. https://doi.org/10.1177/01492063221127907 Nie, Y., Kong, Y., Dong, X., Mulvey, J. M., Poor, H. V., Wen, Q., and Zohren, S. (2024). A survey of large language models for financial applications: Progress, prospects and challenges. arXiv preprint arXiv:2406.11903. Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., and Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST 2023). Piccialli, F., Chiaro, D., Qi, P., Bellandi, V., and Damiani, E. (2025). AgentAI: A comprehensive survey on autonomous agents in distributed AI for Industry 4.0. Expert Systems with Applications, 291, 128404. Reale, J. (2024). Interbank decisions and margins of stability: An agent-based stock-flow consistent approach. Journal of Economic Dynamics and Control, 160, 104822. Russell, S., and Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson, Harlow. Sen, P., and Puranam, P. (2022). Do alliance portfolios encourage or impede new business practice adoption? Theory and evidence from the private equity industry. Strategic Management Journal, 43(11), 2279-2312. Wooldridge, M. (2021). A Brief History of Artificial Intelligence: What It Is, Where We Are, and Where We Are Going. Flatiron Books, New York. Xiao, Y., Sun, E., Luo, D., and Wang, W. (2025). TradingAgents: Multi-agents LLM financial trading framework. arXiv preprint arXiv:2412.20138. Yang, Y., Zhang, Y., Wu, M., Zhang, K., Zhang, Y., Yu, H., Hu, Y., and Wang, B. (2025). TwinMarket: A scalable behavioral and social simulation for financial markets. Proceedings of the Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025).