Governments across the world now use software to decide who receives public_assistance, how much they receive, and whether they have committed fraud. This article examines what is lost when the moral work of #welfare_administration is transferred to automated systems. Drawing on a qualitative, interpretive analysis of documented cases from Australia, the Netherlands, the United States, the Nordic countries, the United Kingdom, Jordan and India, the article develops the concept of the #empathy_deficit: a structural, and not merely accidental, absence of recognition, responsiveness, relational accountability and moral repair in systems of #algorithmic_welfare. The analysis identifies seven mechanisms that produce this deficit: the abstraction of persons into data points; the substitution of proxies for lived need; the reversal of the burden of proof; the hollowing out of human oversight through automation bias; the diffusion of accountability across many hands; the industrialisation of error through speed and scale; and the transfer of administrative burden onto the poorest claimants. The article takes seriously the strongest arguments in favour of automation, including consistency, reach and the well documented biases of human caseworkers, and concludes that these benefits are real but do not answer the empathy critique, because they measure the wrong thing. The article closes by proposing a framework for #empathic_administration built on contestability by design, meaningful human involvement, participatory development, generous error defaults, and a renewed investment in frontline capacity. The argument is not that computation should be banished from the welfare state, but that any system which distributes survival must be judged by how it treats people when it is wrong. Keywords: algorithmic welfare, automated decision-making, empathy, administrative burden, digital welfare state, street-level bureaucracy, social policy, algorithmic accountability 1. Introduction A welfare system is a promise. It says that in a society wealthy enough to prevent destitution, no person will be left to face hunger, homelessness or untreated illness alone. The promise is honoured, or broken, in thousands of small encounters: an application form, a phone call, a home visit, a decision letter. Each encounter is a moment where a state either recognises a person as a person, or processes them as a case. For most of the twentieth century, that encounter was human. A caseworker sat across a desk, read a file, listened, asked questions, and exercised judgement within the boundaries of rules. This system had serious flaws. Caseworkers could be unkind, prejudiced, overworked or arbitrary. Discretion could shelter discrimination. The history of #public_assistance is not a history of gentle bureaucrats. Yet the human encounter carried something that rules alone cannot supply. It carried the possibility of being seen. A caseworker could notice that an applicant was shaking, that a form had been filled in by a child, that the reason for a missed appointment was a funeral. She could decide that a rule, applied literally, would produce an absurd or cruel outcome, and she could act. This capacity, imperfect and unevenly used, is what this article calls #administrative_empathy. Over the last fifteen years, governments have steadily replaced that encounter with software. Eligibility is assessed by rules engines. Fraud is detected by risk scores. Payments are calculated by systems that read tax data automatically. Applications are filed through portals with no human on the other side. Sanctions are triggered by missed digital check-ins. This transformation is often described in the language of efficiency, modernisation and fairness, and it is defended on the grounds that machines do not have bad days, do not hold grudges, and do not treat one claimant differently from another. This article argues that something has gone badly wrong in this transition, and that the failure is not simply technical. Where automated welfare systems have caused harm, the harm has rarely been a matter of a single buggy line of code. It has been a matter of design assumptions that made harm difficult to see, difficult to report, and difficult to reverse. The problem is #structural rather than incidental. The central claim of this article is that the systems now used to distribute public assistance suffer from an empathy deficit. This is not a sentimental complaint. Empathy, as used here, is not a feeling of pity. It is an administrative capacity with four components. First, #recognition: the ability to perceive the individual circumstances of a person rather than the statistical shadow they cast. Second, #responsiveness: the ability to adjust an outcome when circumstances demand it. Third, relational accountability: the presence of an identifiable human being who owes the claimant an explanation. Fourth, #moral_repair: the capacity to acknowledge error, apologise and restore. Automated welfare systems, as currently built and governed, systematically weaken all four. They abstract people into variables, they resist adjustment, they diffuse responsibility across contractors and departments, and they are extraordinarily bad at admitting they were wrong. The consequence is not that decisions become slightly less warm. The consequence, as the cases in this article show, is that hundreds of thousands of people are wrongly accused of fraud, wrongly denied care, and wrongly pushed into debt, and that when this happens the system offers them nobody to talk to. 1.1 Research questions This article addresses four questions. How do automated decision-making systems reshape the moral relationship between the state and the claimant in the distribution of public assistance? What specific mechanisms within these systems produce the empathy deficit? What is the strongest case in favour of automation in welfare, and does it survive contact with the documented evidence? What institutional, legal and design responses could restore empathic capacity without abandoning the legitimate benefits of computation? 1.2 Contribution The literature on #algorithmic_governance is now substantial, and the literature on administrative burden in social policy is older still. The two bodies of work rarely meet. Critical data studies has produced rich accounts of #datafication and its harms, but often treats the welfare state as one site among many. Public administration scholarship has produced careful empirical work on #discretion and automation, but tends to frame the question as one of accuracy and efficiency. Law has produced detailed analysis of transparency and contestation rights, but focuses on remedies after the fact. This article contributes a synthesis. It offers empathy as an organising concept that connects these strands, treats it as a measurable administrative property rather than an emotion, and uses it to explain why systems that were designed to be fair produced results that almost everyone now agrees were cruel. It also contributes a comparative reading of cases across high-income and middle-income settings, which shows that the empathy deficit is not a peculiarity of one country's politics but a recurring product of a particular design logic. 1.3 Structure Section 2 sets out the background of the digital welfare state. Section 3 reviews the relevant literature. Section 4 develops the conceptual framework. Section 5 explains the methodology. Section 6 presents the case material. Section 7 analyses the mechanisms of the empathy deficit and engages the counterarguments. Section 8 proposes a framework for empathic administration. Section 9 sets out limitations and directions for further research. Section 10 concludes. 2. Background: The Rise of the Digital Welfare State 2.1 What changed The move to automation in welfare was not a single decision. It was an accumulation of small ones, each defensible on its own terms. The first wave was #digitisation. Paper files became electronic records. This was largely uncontroversial and delivered real gains in record keeping. The second wave was #digital_by_default service delivery. Applications, reporting duties and communications moved online. Governments framed this as convenience. For claimants with reliable internet access, a device, digital literacy and stable housing, it often was. For those without, it became a barrier that had the practical effect of a denial. The third wave, and the one this article is chiefly concerned with, was #automated_decision_making. Here software does not merely record or transmit a decision. It produces one. Rules engines determine eligibility. Data matching identifies discrepancies between what a claimant reported and what another agency holds. Risk models score claimants for the likelihood of fraud. Algorithms allocate hours of home care, quantities of food assistance, or places in a queue. A fourth wave is now beginning, in which large language models and generative systems are proposed for drafting decision letters, triaging correspondence and advising caseworkers. This article treats that wave as an intensification of existing dynamics rather than a break from them. 2.2 Why governments did it Four motives recur in policy documents across jurisdictions. The first is cost. Welfare administration is labour intensive. Automation promises to reduce headcount. In many countries the adoption of automated systems coincided with austerity programmes and explicit staffing reductions. The second is #fraud_recovery. Political attention to #welfare_fraud is disproportionate to its actual scale, but it is politically potent. Automated data matching was sold as a way to find money that was allegedly being stolen. The third is consistency. Human #caseworkers decide similar cases differently. This is a genuine problem of horizontal equity, and automation appears to solve it. The fourth is speed. Claimants wait too long for decisions. Automation appeared to promise faster service, and in some narrow respects it delivered it. None of these motives is dishonourable. The argument of this article is not that officials set out to harm claimants. It is that the systems they built encoded a particular view of the claimant, as a potential fraudster to be verified rather than a citizen to be served, and that this view, once encoded, became very difficult to argue with. 2.3 The asymmetry at the heart of the system Welfare automation operates on a population that is, almost by definition, in a weak position. Claimants have low incomes, often poor health, frequently unstable housing, and limited access to legal advice. They cannot easily hire lawyers, commission expert reports, or wait months for an appeal while their rent goes unpaid. This produces a structural asymmetry. The state deploys a complex system whose logic is opaque even to many of the officials who operate it, and asks a person in crisis to challenge it. The costs of error are not borne symmetrically. If the system wrongly grants assistance, the state loses money. If the system wrongly denies assistance, a person may lose their #housing, their health or their life. Yet almost every automated welfare system studied in this article was tuned to minimise the first kind of error, not the second. That tuning decision is where the #empathy_deficit begins. 3. Literature Review 3.1 Automation and discretion in public administration The foundational insight for this field comes from Michael Lipsky's work on street-level bureaucracy, which showed that policy is not made in legislatures but enacted in the daily choices of #frontline_workers who must ration scarce resources under pressure. Discretion, in this account, is not a defect but a mechanism through which general rules meet particular lives. Contemporary public administration scholarship has taken up the question of what happens when that discretion is automated. Empirical work suggests the effect is neither total elimination nor simple preservation. Automation tends to relocate discretion rather than remove it: away from the frontline worker and towards #system_designers, procurement officers and the authors of decision rules, who make consequential choices without ever meeting a claimant. Studies of enforcement contexts show that automation changes how frontline staff use the discretion that remains, often narrowing it to the space the system permits (de Boer and Raaphorst, 2023). Work on algorithmisation in bureaucratic organisations shows that the same system produces different practices in different institutional contexts, which suggests that organisational culture matters as much as code (Meijer, Lorenz and Wessels, 2021). A parallel body of work examines the human oversight that is supposed to protect claimants from machine error. The findings are discouraging. Public officials given algorithmic advice display #automation_bias, deferring to the system even when it conflicts with available evidence, and also display selective adherence, following algorithmic advice more readily when it confirms existing stereotypes about the group in question (Alon-Barkat and Busuioc, 2023). Legal scholarship reaches a similar conclusion from a different direction, arguing that policies requiring a human in the loop frequently function as false assurance, legitimating automated outputs without meaningfully checking them (Green, 2022). 3.2 The digital welfare state and datafied poverty Critical scholarship on datafication has documented how welfare recipients are subject to forms of surveillance that wealthier citizens do not experience. #Poverty, in these accounts, is not merely a condition addressed by data systems but a condition partly constituted by them, as claimants are required to surrender information as a condition of survival. Research on the datafied welfare state in Europe traces how administrative data becomes a resource for risk prediction and how the resulting infrastructure reshapes what the welfare state is for (Dencik, 2022). Related work examines litigation as a site where these systems become visible. Kaun's study of legal challenges to automated decision-making in Swedish social services shows that court cases perform a translation function, turning an opaque technical artefact into a contestable public object (Kaun, 2022). This matters for the empathy argument, because it demonstrates that #opacity is not simply a technical property of complex models. It is often a governance choice, sustained by claims of commercial confidentiality and reversed only under legal pressure. Broader critical work situates these systems within longer histories of classification and control (Crawford, 2021; Broussard, 2023; Birhane, 2021). Hoffmann's analysis of inclusion discourse is particularly relevant, showing how efforts to make systems fairer by improving data can deepen the very violence they claim to remedy, because they leave untouched the question of whether the person should have been classified in this way at all (Hoffmann, 2021). 3.3 Empathy, care and moral perception The concept of empathy has a substantial philosophical literature, much of it concerned with whether empathy is a reliable guide to moral action at all. Critics note that #empathy is partial, biased towards the near and the familiar, and easily manipulated. This article accepts the force of that critique and therefore does not rely on empathy as an emotion. It relies on empathy as an institutional capacity: the design of processes that permit particular circumstances to change outcomes. Recent work on artificial empathy is directly relevant. Scholarship in health care argues that there are in principle obstacles to machine empathy, because empathic understanding involves a shared vulnerability and an openness to being changed by the encounter, which computational systems do not possess (Montemayor, Halpern and Fairweather, 2022). Others argue that whatever machines simulate, the recipient's knowledge that no one is actually attending to them alters the moral quality of the exchange (Perry, 2023). A dissenting line of argument holds that machine-generated empathic responses may still deliver real benefit, particularly where the human alternative is nobody at all, and that dismissing them risks a purism that leaves people unsupported (Inzlicht and colleagues, 2024). Section 7 engages this dissent seriously. Vallor's account of artificial intelligence as a mirror, reflecting human patterns back at us without understanding them, offers a useful frame for welfare systems in particular (Vallor, 2024). A fraud model trained on historical enforcement data does not perceive fraud. It perceives the pattern of who was previously investigated, and returns it with the authority of mathematics. 3.4 Administrative burden The administrative burden literature examines the costs that citizens bear when interacting with the state: learning costs, compliance costs and #psychological_costs. This framework is essential to the present argument because it makes visible the harm that eligibility statistics conceal. A system can have a defensible error rate on paper while imposing crushing burdens on the people it wrongly flags, who must gather years of documents, navigate call centres, and prove a negative under threat of penalty. Empirical research on frontline digital service encounters shows that when channels are closed, both workers and claimants develop coping strategies, and that these strategies redistribute effort onto the claimant (Breit and colleagues, 2021; Loberg, 2021). The psychological costs, including stigma, fear and exhaustion, fall hardest on those with the fewest reserves. 3.5 Law and accountability Legal scholarship provides the vocabulary of #remedy. Article 22 of the European #General_Data_Protection_Regulation restricts decisions based solely on automated processing that produce legal or similarly significant effects, and grants rights to obtain human intervention and to contest. Critics have argued persuasively that these protections are narrower in practice than in appearance, and that the concept of a decision based solely on automated processing invites evasion through token human sign-off (Veale and Zuiderveen Borgesius, 2021). Work on non-discrimination law shows that European legal doctrine is poorly matched to statistical discrimination, because it depends on a comparator and an identifiable protected characteristic, while algorithmic systems discriminate through #proxy_discrimination and intersections that the law struggles to name (Wachter, Mittelstadt and Russell, 2021). Malgieri's work on vulnerability in data protection law offers a route towards a more protective standard, one that takes the claimant's actual position of dependence as legally relevant rather than assuming a rational, well resourced data subject (Malgieri, 2023). Accountability scholarship in public administration adds a further layer, arguing that algorithmic systems disrupt the classic accountability relationship because the account giver cannot explain the decision, the forum cannot interrogate it, and the consequences cannot be attributed to any identifiable actor (Busuioc, 2021). Transparency alone does not repair this. Experimental work suggests that explanation improves perceived trustworthiness only under some conditions, and that a technically accurate explanation of an unjust decision does not make the decision just (Grimmelikhuijsen, 2023). 3.6 The gap What the literature has not done is connect these strands around a single normative question: what does a welfare system owe a person, and can a machine deliver it? Accuracy scholarship asks whether the system is right. #Fairness scholarship asks whether it is right equally across groups. #Transparency scholarship asks whether the person can see how it decided. None of these asks whether the person was treated as a person. That is the gap this article addresses through the concept of the empathy deficit. 4. Conceptual Framework: Defining the Empathy Deficit 4.1 Empathy as an administrative capacity To make empathy analytically useful, it must be separated from sentiment. This article defines #administrative_empathy as the institutional capacity of a decision system to allow the particular circumstances of an individual to change the outcome that general rules would otherwise produce, and to sustain a relationship in which the individual can be heard, believed and, if wronged, restored. This definition has four components, and each yields a diagnostic question. Recognition. Can the system perceive the claimant's situation in its particularity, including facts that were never anticipated by the designers? Diagnostic question: what happens to a true and relevant fact that has no field on the form? Responsiveness. Can the system change its output in response to that recognition, within a timeframe that matters to a person who may be about to lose their home? Diagnostic question: how long does it take, and how much must the claimant do? Relational accountability. Is there an identifiable human being who owes the claimant an explanation and who has the authority to act? Diagnostic question: who can the claimant talk to, and can that person say yes? Moral repair. When the system is wrong, does it acknowledge the wrong, apologise, compensate and correct itself? Diagnostic question: what does the institution do after it discovers it has harmed someone? An empathy deficit exists when a system's design systematically weakens one or more of these capacities. It is a structural property of the system, not a description of the character of its staff. Kind people can operate an empathy-deficient system, and frequently do, which is one reason the deficit is so hard to see from inside. 4.2 Why empathy is not a luxury A predictable objection is that welfare systems exist to apply rules accurately, not to be kind, and that empathy is a private virtue with no place in public administration. This objection misunderstands both law and rules. Rules are always underdetermined at the point of application. Every eligibility rule contains terms that must be interpreted: cohabitation, availability for work, reasonable effort, ordinary residence. Someone must decide what these mean in the messy circumstances of an actual life. If a human does not, then the interpretation is made in advance by whoever wrote the code, applied without knowledge of the case, and insulated from the corrective feedback that comes from looking a person in the eye. Empathy in this sense is not an exception to correct rule application. It is a condition of it. Furthermore, all decision systems make errors. The relevant question is not whether errors occur but how they are distributed and what happens next. Empathic capacity is precisely the machinery of error correction. A system without it does not have fewer errors. It has uncorrected ones. 4.3 The empathy deficit defined Bringing these elements together, the #empathy_deficit is defined as: the systematic reduction, through the design and governance of automated decision systems, of an administration's capacity to recognise individual circumstance, respond to it, remain answerable for it, and repair harm caused by it. The remainder of this article tests this concept against evidence. 5. Methodology 5.1 Approach This study uses a qualitative, interpretive and comparative case-oriented design. It is a work of #critical_policy_analysis rather than of primary empirical data collection. The aim is #theory_building: to derive an explanatory account of a recurring pattern from documented cases, and to test that account against alternative explanations. 5.2 Data sources The case material is drawn from publicly available documentary sources: parliamentary and commission inquiry findings, published #court_judgments, peer reviewed research on the systems concerned, and detailed investigative reporting where it has been corroborated by official inquiry. Where a claim is contested or has been disputed by a government, this article says so rather than presenting it as settled. 5.3 Case selection Cases were selected using three criteria. First, documentation. The case must have been examined by an independent body, a court, or peer reviewed research, so that the account does not rest on advocacy alone. Second, variation. The cases span different welfare functions (fraud detection, eligibility assessment, benefit calculation, resource allocation, targeting), different administrative traditions (Anglophone, continental European, Nordic, and middle-income contexts), and different technical approaches (deterministic rules engines, data matching, statistical risk models). Third, consequence. The case must involve decisions that materially affected access to subsistence, care or housing, since the argument concerns systems that distribute survival. The result is a #purposive_sample rather than a representative one. This is appropriate for theory building but limits generalisation, a point taken up in Section 9. 5.4 Analytical strategy Each case was read against the four-part framework in Section 4. The analysis then moved across cases to identify mechanisms that recurred despite differences in technology, law and politics. A mechanism was retained only if it appeared in at least three cases and if its absence in a fourth was explicable. Counterarguments were then developed in their strongest form and tested against the same evidence. 5.5 Positionality and limits of the method Documentary analysis of failure has a selection problem: systems that work quietly are not investigated, and so do not appear. This article does not claim that every automated welfare system has produced catastrophe. It claims that where catastrophe occurred, the causes were structurally similar, and that these causes are present in systems that have not yet failed publicly. That is a claim about #latent_risk, and it should be read as such. 6. Case Material 6.1 Australia: the Online Compliance Intervention Between 2015 and 2019, the Australian government operated an automated compliance programme, widely known by the name Robodebt, which sought to identify overpayments of social security benefits. The system compared annual income data held by the tax office with the fortnightly income that claimants had reported to the welfare agency. Where the two did not match, the system divided the annual figure into equal fortnightly amounts, a method known as #income_averaging, and treated the resulting gap as evidence of a debt. The method was wrong at the level of #arithmetic. Casual, seasonal and gig work is not distributed evenly across a year. A person who worked intensively for three months and had no income for nine would appear, under averaging, to have earned a steady wage throughout and therefore to have been overpaid. This was not an edge case. It was the normal working pattern of precisely the population the system targeted. Two design decisions turned an error into a catastrophe. The first was the removal of the human review that had previously preceded debt raising, which meant that a discrepancy identified by the machine became a debt notice without an official ever forming a view about the individual. The second was the #reversal_of_the_burden_of_proof. Rather than the state proving that a debt existed, the claimant was required to produce payslips, often going back years, from employers who had since closed, to prove that it did not. The scheme raised hundreds of thousands of debts, many of them unlawful. A #class_action was settled in 2020 with the government agreeing to wipe hundreds of millions of dollars in debts and to repay large sums that had been wrongly recovered, some through the use of private debt collectors and garnishment of tax refunds. A #Royal_Commission reported in 2023 and found the scheme to have been unlawful, and characterised it in language that no critic could have improved upon, describing it as a crude and cruel mechanism. Evidence before the inquiry included accounts from families who linked the receipt of automated debt notices to the deaths of relatives. Applying the framework: recognition failed, because irregular income had no representation in the model. Responsiveness failed, because the appeal channels were slow and the debt continued to accrue and be collected while a challenge was pending. Relational accountability failed spectacularly, because claimants encountered an outsourced call centre with no authority and a decision that no named person had made. Moral repair came only after years of litigation, journalism and a Royal Commission, and only after immense harm. 6.2 The Netherlands: childcare benefits and the risk classification model The Dutch childcare benefits scandal, often called the toeslagenaffaire, involved the tax administration's pursuit of alleged fraud in a subsidy that helped parents pay for childcare. A risk classification system was used to select claimants for #scrutiny. Among the factors that fed into risk scoring was information relating to nationality, with the effect that having a second nationality made a parent more likely to be flagged. Human rights organisations described the resulting system as one that produced #ethnic_profiling by design. The consequences were severe. Tens of thousands of families were accused of fraud, often on the basis of minor administrative irregularities such as a missing signature. Benefits were stopped and full repayment of past sums was demanded, in amounts that were ruinous. Families fell into debt, lost homes and marriages, and in a significant number of cases children were placed into state care in circumstances connected to the financial devastation the accusations caused. The #government_resigned in January 2021 over the affair. A related case is the Dutch System Risk Indication, a data linking tool used to identify potential welfare fraud in selected neighbourhoods. In 2020 a court in The Hague held that the system violated the right to #private_life protected by the European Convention on Human Rights, in part because the government would not disclose how it worked, leaving citizens unable to defend themselves. Applying the framework: recognition failed because a data mismatch was read as a moral fact about a person. Responsiveness failed because there was no proportionate response available to the system between doing nothing and demanding total repayment. Relational accountability failed because the risk model's outputs were treated as authoritative by officials who could not explain them. Moral repair has been attempted, through compensation schemes and public apology, and remains incomplete years later, which itself demonstrates how much harder repair is than harm. 6.3 United States: fraud detection and care allocation Two distinct American cases illustrate different faces of the same deficit. The first is the Michigan unemployment insurance system, which from 2013 used an automated system to detect fraud in benefit claims. The system issued fraud determinations, with severe financial penalties attached, largely without human review. Subsequent state review found that the overwhelming majority of the automated fraud determinations were wrong. Tens of thousands of people were falsely accused. Penalties were collected through #wage_garnishment and the seizure of tax refunds, and some of those affected reported bankruptcy and the loss of homes. Litigation continued for years and produced a settlement. The second is the use of algorithmic assessment to allocate hours of #home_care to disabled people under state Medicaid programmes. In several states, the introduction of an assessment algorithm produced sudden and steep reductions in the hours of personal assistance that people with significant disabilities received, with no change in their underlying condition. When recipients asked why their care had been cut, they were told that the computer had made the assessment. Court proceedings revealed that the models were poorly documented, that they contained errors, and that the state agencies themselves could not adequately explain the outputs on which they were relying. The second case is analytically important because it involves no allegation of fraud and no punitive intent. The purpose was #resource_allocation. Yet the empathy deficit is fully present: a person's daily capacity to eat, wash and leave their bed was reduced by a system whose reasoning nobody would explain, and the burden of challenging it lay with the person whose care had just been cut. 6.4 Nordic Europe: automation in municipal social assistance Swedish municipalities have adopted automation for the processing of social assistance applications. In the widely discussed case of one municipality, the automation of assistance decisions was accompanied by public controversy, professional resistance from #social_workers, and eventually litigation over whether the source code and decision logic constituted public documents that citizens were entitled to see. The significance of the case for this article is not that the system was uniquely harmful. It is that the question of whether the #public could inspect the logic of a system deciding their subsistence had to be settled in court (Kaun, 2022). In Denmark, proposals to use predictive models to identify children at risk, drawing on wide ranging municipal data, generated sufficient public and professional objection that they were curtailed. This case is instructive because it shows that the deficit is not inevitable. Where professional bodies retained authority and where public debate happened before deployment rather than after harm, the trajectory changed. 6.5 United Kingdom: real time data and rigid periods The United Kingdom's Universal Credit reformed several benefits into a single monthly payment calculated automatically from employer reported earnings data. The design assumed that earnings arrive in tidy monthly units. For workers paid weekly or fortnightly, and for those whose payday shifted because of a weekend or public holiday, two salary payments could land inside a single assessment period. The automated calculation then read this as a doubling of income and reduced or eliminated the benefit award, before reversing the following month. Claimants experienced violent swings in income that made budgeting impossible. Courts eventually found the approach to be irrational in its treatment of the affected claimants. The case is valuable precisely because there was no malice, no fraud model and no ethnic profiling. There was only a #calendar_assumption embedded in code, and an administration that took years to accept that reality did not conform to it. This is the empathy deficit in its purest form: not cruelty, but the inability to hear that the world does not work the way the system says it does. 6.6 Jordan: targeting cash transfers by algorithm Jordan's cash transfer programme used an algorithmic ranking of applicant households to allocate limited assistance, drawing on indicators such as household expenditure, car ownership and utility consumption. Human rights research documented that the ranking produced results that families found arbitrary and unexplainable, and that the indicators used could misclassify need. A household owning a vehicle less than five years old, for instance, might be excluded, regardless of whether the vehicle was the family's only means of earning income or was in fact inoperable. This case matters because it shows the deficit operating in a context of genuine scarcity, where some form of rationing was unavoidable, and where the funder was an international institution. #Rationing decisions are hard, and no method will satisfy everyone. But a system that cannot explain to a family why they were ranked below their neighbour, and that offers no route to say that the model has misread their life, converts a difficult allocation into an unaccountable one. 6.7 India: biometric authentication and the ration shop India's linkage of biometric identity authentication to the distribution of subsidised food grain was intended to eliminate #ghost_beneficiaries and diversion. Research on the system has documented #exclusion_errors: authentication failures caused by worn fingerprints, common among manual labourers and the elderly, by network outages in rural areas, and by mismatches between databases. Where authentication failed, entitled people were turned away from ration shops. Journalists and researchers have reported cases of severe deprivation, including deaths in which the family attributed the loss to being cut off from rations. The government has disputed the causal attribution in specific cases, and this article records that dispute rather than resolving it. The structural point stands regardless of any individual case. A design that treats successful biometric authentication as a precondition of eating, and that provides no reliable fallback when the technology fails, has made a decision about who bears the cost of technical failure. It has decided that the cost is borne by the hungriest person in the queue. 6.8 What the cases share The cases differ in technology, wealth, legal tradition and political context. They share a pattern. In each, a system was introduced with the stated goal of accuracy or integrity. In each, the system encoded an assumption about the shape of a normal life that did not fit the population it governed. In each, the resulting mismatch was interpreted not as a defect in the model but as a fault in the claimant. In each, the person harmed was required to do the work of correcting the state's error. And in each, the harm continued for years because the system offered no channel through which the truth could travel back. 7. Analysis: The Mechanisms of the Empathy Deficit 7.1 Abstraction and the erasure of context An automated system can only act on what has been made #legible to it. Legibility requires categories, and categories are made in advance. The claimant's life must be translated into fields. #Translation is always lossy, but the losses are not random. What survives translation is what is countable and administratively convenient. What is destroyed is context: the reason, the history, the relationship, the constraint. A caseworker who is told that an applicant missed three appointments can ask why. A rules engine cannot ask why, because why is not a field. This is not a limitation that better data collection can solve. Adding fields adds surveillance without adding understanding, because the next unanticipated circumstance will fall outside the new fields as surely as it fell outside the old ones. #Human_life is not enumerable in advance. Any system that must exhaustively pre-specify relevant facts will be permanently behind reality, and the gap between them is filled with people whose true and relevant circumstances have nowhere to go. 7.2 Proxy variables and the misreading of poverty Automated welfare systems rarely measure need. They measure #proxies for need, and proxies for dishonesty. Car ownership stands in for wealth. Utility consumption stands in for household size. A data mismatch stands in for fraud. Address instability stands in for risk. Each proxy is defensible on average and wrong in the particular, and welfare is delivered in the particular. Worse, proxies encode the past. A fraud model trained on which claimants were previously investigated learns the pattern of previous investigation, which reflects where enforcement effort was historically directed. If that effort was directed at particular neighbourhoods or particular groups, the model will faithfully reproduce that direction and present it as an objective finding. The Dutch case shows this in its starkest form, where nationality entered the risk calculation directly. But the same result can be reached without ever naming a protected characteristic, through variables that correlate with it. This is the difficulty that non-discrimination law has struggled to meet (Wachter, Mittelstadt and Russell, 2021). The deeper problem is conceptual. #Fraud is a state of mind. It requires #intent. No dataset contains intent. What the data contains is discrepancy, and a system that treats discrepancy as fraud has not detected dishonesty. It has redefined it, so that being poor in a complicated way becomes indistinguishable from lying. 7.3 Reversal of the burden of proof Across the cases, the same #procedural_inversion recurs. The system asserts a debt, an overpayment or a disqualification. The claimant must disprove it. This inversion is doing enormous work, and it is rarely justified explicitly. It transforms a probabilistic signal into an accusation with legal effect, and then requires the least resourced party to rebut it, often with documents that only third parties hold and that may no longer exist. The empathy analysis reveals why this is so damaging. It is not simply unfair in the abstract. It is a statement about belief. When an administration says that its data outweighs a person's testimony about their own life, it has declared in advance that the person will not be believed. Everything that follows, the phone calls, the letters, the collection notices, communicates that declaration. The reported experiences of claimants in these systems are dominated not by financial loss alone but by the experience of being disbelieved by an institution that would not explain itself and could not be reached. 7.4 The hollow human in the loop Every system examined here had, at some stage, a human somewhere. This is the standard reassurance offered when concerns are raised: a person always makes the final decision. The evidence does not support the reassurance. #Human_oversight fails in predictable ways. Officials defer to algorithmic outputs because the output carries institutional authority and because departing from it requires extra work and creates personal risk. Where an official overrides the system and the override turns out badly, the official is accountable; where the official follows the system and it turns out badly, the system is at fault. The incentive structure guarantees deference. Selective adherence compounds this. Officials do not defer uniformly. They defer more readily when the algorithm's output aligns with existing assumptions about a claimant group, which means that automated advice can amplify rather than correct prejudice (Alon-Barkat and Busuioc, 2023). Add to this the operating conditions of most welfare administrations: high caseloads, performance targets measured in cases closed, and staff who have not been trained in the model's limitations and would not be permitted the time to investigate them if they had been. Under these conditions the human in the loop is not a safeguard. It is a signature. Legal scholarship has argued precisely this, that human oversight requirements often provide false assurance and legitimate the automated decision they were meant to check (Green, 2022). 7.5 Diffusion of accountability Ask who decided, and the question dissolves. The vendor built the model to a specification. The department wrote the specification but relies on the vendor's expertise for its behaviour. The minister sets policy but does not see the code. The frontline officer applies the output but did not choose it. The call centre is outsourced and has no authority. The debt collection is outsourced again. Each actor can truthfully say that they did not make the decision, and collectively they made it. This is the classic problem of #many_hands, sharpened by technical opacity. Accountability requires that someone can give an account, that a forum can interrogate it, and that consequences can follow. Algorithmic systems disrupt all three (Busuioc, 2021). For the claimant, the practical consequence is that there is nobody to talk to. Relational accountability is not a legal abstraction. It is a phone call in which a named person, with authority, listens and can act. Its absence is felt as abandonment. 7.6 Speed, scale and the industrialisation of error A human caseworker who holds a mistaken belief harms the people in front of her. An automated system that holds a mistaken belief harms everyone in the database, simultaneously, and continues to do so until someone stops it. This changes the character of error. Manual systems produce errors that are scattered, individual and self limiting. Automated systems produce errors that are correlated, systematic and vast. A single wrong assumption, such as that income arrives evenly across a year, becomes hundreds of thousands of wrongful debts. Scale also interacts with feedback. Because the volume is so large, complaints are received as noise. A single caseworker learns from a mistake because she meets its consequences. An institution processing millions of automated determinations receives complaints as a metric, and a rising complaint rate can be read as evidence that the system is successfully identifying fraud among people who object to being caught. The very signal that should trigger correction is reinterpreted as confirmation. This inversion appears in more than one of the cases, and it is perhaps the single most dangerous feature of automated welfare. 7.7 The transfer of administrative burden The #efficiency claim conceals a transfer. Work does not disappear when a system is automated. It moves. The state saves the labour of assessment. The claimant acquires the labour of proving, appealing, uploading, resubmitting, telephoning, and waiting. This labour is uncompensated, and it falls on people who are ill, exhausted, caring for others, working unpredictable shifts, and often without reliable connectivity or a working device. Framed in the language of administrative burden, automation frequently reduces the state's costs by increasing the citizen's learning costs, compliance costs and psychological costs. When these costs become high enough, they operate as a form of #administrative_rationing. People give up. #Non_take_up of benefits by eligible people is not a side effect of such systems. In an environment where fraud prevention is the dominant political objective, it functions as a quiet success. 7.8 The strongest case for automation It would be intellectually dishonest to present only the failures. The case for automation is serious and deserves a serious answer. Consistency. Human caseworkers are inconsistent. Identical cases receive different decisions depending on who reads the file, the time of day and the mood of the assessor. Automation eliminates this variance. This is a real gain in horizontal equity, and it should not be waved away. Freedom from human prejudice. #Discretion has sheltered racism, sexism and contempt for the poor throughout the history of welfare. The claim that human judgement is warm and machine judgement is cold ignores how cold human judgement has often been to those the caseworker despised. Some claimants prefer a machine precisely because a machine will not sneer. Reach and speed. Automated systems can process claims that would otherwise sit in queues for months. Where an administration is overwhelmed, automation may deliver assistance to people who would otherwise receive nothing. This is a strong argument in emergencies and in systems with severe capacity constraints. Detection of real fraud. Fraud exists. It diverts money from people who need it and it erodes political support for the welfare state itself. Data matching does find genuine cases. The counterfactual is not a caring caseworker. This is the most powerful version of the argument, and it parallels the debate in the artificial empathy literature (Inzlicht and colleagues, 2024). In many administrations, the realistic alternative to an algorithm is not attentive human judgement. It is an eight-week wait, a phone line that nobody answers, and a caseworker with four hundred files. Comparing automation to an idealised human encounter that no longer exists, and perhaps never existed, is a rhetorical trick. 7.9 Why the case for automation does not answer the empathy critique These arguments are real, and yet they do not defeat the thesis, for four reasons. First, they measure the wrong thing. Consistency is a property of the distribution of decisions. Empathy is a property of the handling of the individual, particularly the individual the system gets wrong. A system can be perfectly consistent and consistently wrong about the same category of person, which is exactly what income averaging achieved. #Uniform_harm is not #equity. Second, the freedom-from-prejudice argument confuses the removal of the prejudiced human with the removal of prejudice. The Dutch case demonstrates that discrimination survives automation and gains authority from it, because a biased human can be challenged, appealed and eventually dismissed, whereas a biased model is defended as objective and protected as proprietary. Machines do not sneer. They simply deny, and their denial is harder to argue with than a sneer. Third, the counterfactual argument is contingent, not structural. The absent caseworker is not a fact of nature. She was made absent by budget decisions, often the same decisions that funded the automation. To justify automation by pointing at the ruins of the human service that automation was used to defund is to complete a circle rather than an argument. Where automation is genuinely additive, expanding coverage without removing human capacity, the argument has real force and this article accepts it. Fourth, and most importantly, none of these arguments addresses error correction. Every defence of automation is a defence of its performance when it is right. The empathy critique concerns what happens when it is wrong. A system that is ninety-five percent accurate across a million determinations is wrong about fifty thousand people. The question that matters, and the question the efficiency case never answers, is what those fifty thousand people can do about it. 7.10 Empathy and generative systems The newest wave of adoption proposes conversational systems that can produce fluent, apparently sympathetic text. It may be tempting to conclude that these systems close the empathy deficit. They do not, and the reason follows directly from the framework. Fluent sympathetic language delivers the appearance of recognition without the substance of responsiveness, accountability or repair. A generated letter that says the department understands how difficult this must be, attached to a decision that cannot be changed, by an entity that cannot be reached, is not empathy. It is a more sophisticated form of the same refusal, and arguably a worse one, because it borrows the language of care to make an unaccountable decision harder to be angry with. What matters is not whether the state sounds like it understands. It is whether the state can be moved. 8. Toward Empathic Administration The critique above is not an argument for abolishing computation in the welfare state. Rules engines that calculate entitlements accurately and pay people quickly are good. Data systems that identify people who are eligible but not claiming, and reach out to them, are better than good. The argument is that any system distributing subsistence must be designed for the person it gets wrong. This section sets out how. 8.1 The generous default The foundational principle is the reallocation of the #cost_of_error. Systems should be designed on the assumption that they will be wrong about a significant number of people, and tuned so that the burden of that wrongness falls on the institution rather than on the claimant. In practice this means: no automatic recovery of alleged debts before an independent human has reviewed the individual case; continuation of payment while a dispute is pending, except where there is specific evidence of dishonesty about that individual; and a presumption in favour of the claimant's account of their own circumstances where the contrary evidence is only a data mismatch. This is not naivety about fraud. It is a decision that the state, which can absorb a wrongly paid benefit, should bear the risk rather than the claimant, who cannot absorb a wrongly withheld one. 8.2 Contestability by design #Contestation must be engineered into the system rather than bolted on as an appeal. A contestable system tells the claimant, at the moment of decision, in plain language, the specific facts and rules that produced the outcome. It tells them which fact to challenge if the decision is wrong. It provides a channel to submit a correction that is as easy as the channel through which the decision arrived. It gives a binding time limit for response. And it treats a successful challenge as information about the system, not merely as an exception granted to one person. The last point is essential. In every case studied, individual successful appeals were treated as isolated wins and did not trigger review of the model that produced the error. #Contestability that does not feed back into design is a pressure valve, not a correction mechanism. 8.3 Meaningful human involvement Given the evidence on automation bias, a human sign-off requirement is close to worthless unless four conditions hold. The reviewer must have #authority to overturn the output without seeking permission. The reviewer must have time, which means caseload standards, not aspirations. The reviewer must have information, including a plain statement of what the model relies on and what it is known to get wrong. And the reviewer must face symmetric incentives, meaning that failing to catch a false accusation is treated as seriously as wrongly overriding a correct one. Absent these four conditions, the honest course is to admit that the decision is fully automated and to accept the legal consequences that follow, rather than to maintain a fiction of human oversight that provides false assurance (Green, 2022). 8.4 Participation of claimants in design Systems that misread the lives of the poor are usually built by people who have not lived those lives and who did not ask. Participatory design is often invoked as a remedy and often performed as theatre, with a consultation held after the procurement is signed. Genuine participation means involving claimants and frontline workers early enough that their input can change the specification, including the decision not to build the system at all. Critical scholarship rightly warns that participation can be co-opted into a legitimating ritual (Sloane and colleagues, 2022). The test of genuine participation is simple: can it stop the project? If not, it is decoration. A concrete and underused form of participation is adversarial testing before deployment: taking real, complicated cases, the seasonal worker, the family with a disabled child, the woman fleeing violence with no documents, and asking what the system does to them. Every one of the failures examined in this article would have surfaced in an afternoon of such testing. 8.5 Legal architecture #Law is not sufficient, but it is necessary. The European Union's approach classifies systems used to evaluate eligibility for essential public assistance benefits and services as high risk, imposing obligations of risk management, data governance, documentation, human oversight and transparency. This is a meaningful advance, and its practical value will depend entirely on enforcement and on whether oversight obligations are implemented substantively or ritually. Data protection rights concerning solely automated decisions remain important but are vulnerable to the token-human evasion described above (Veale and Zuiderveen Borgesius, 2021). Non-discrimination law needs to develop doctrine capable of addressing proxy discrimination rather than requiring an explicit protected characteristic (Wachter, Mittelstadt and Russell, 2021). And administrative law should treat the inability of an agency to explain its own decision as a defect in the decision itself, rather than as an unfortunate technical circumstance. A further legal principle deserves attention: recognising the claimant's vulnerability as legally relevant to what the state owes them, rather than presuming an autonomous, informed and well resourced individual capable of exercising rights (Malgieri, 2023). A right to contest that a person cannot in practice exercise is not a right. It is an alibi. 8.6 Transparency that is worth something Publishing #source_code is necessary but rarely sufficient, since almost no claimant can read it. What matters is explanation at three levels. At the level of the system, the public should know that a system exists, what it does, what data it uses, how it was tested and how often it is wrong. At the level of the decision, the claimant should receive the specific reasons in plain language. At the level of the pattern, oversight bodies and researchers should have access to disaggregated outcome data sufficient to detect whether the system errs more often against particular groups. Transparency alone does not create trust and should not be expected to (Grimmelikhuijsen, 2023). Its function is to make challenge possible. 8.7 Institutional culture and frontline capacity None of the above will work in an institution that regards claimants as suspects. The Dutch, Australian and Michigan cases were not caused by code. They were caused by a political and administrative culture that had decided in advance that the population it served was probably cheating, and then commissioned tools that expressed that belief at scale. #Technology did not create the suspicion. It executed it. This means the most important reform is also the least technical. It requires reinvesting in frontline staff who have the time to listen, the authority to act and the professional standing to say that the system is wrong. It requires measuring administrations on whether eligible people receive what they are owed, not only on whether ineligible people are excluded. And it requires political leaders to stop describing the recipients of public assistance in the language of criminality, because the systems a state builds are the beliefs of that state made operational. 8.8 A checklist for evaluation Drawing the framework together, any proposed automated welfare system should be required to answer the following before deployment. What happens to a true and relevant fact for which the system has no field? If the system is wrong about a person, how will that person find out, and how long will correction take? Who is the named human being with authority to overturn the output, and what is their caseload? Does anything continue to be taken from the claimant while a dispute is pending? What is the estimated false positive rate, on whom does it fall, and how was that estimated? Was the system tested against the hardest real cases, and what did it do to them? If the system harms people, what is the mechanism by which it will be stopped, and who can trigger it? A system that cannot answer these questions is not ready, whatever its accuracy on the test set. 9. Limitations and Future Research This study has several limitations that shape how its conclusions should be read. First, it relies on documented failures, and documented failures are a biased sample. Systems that quietly work do not generate inquiries. The article therefore cannot support any claim about the proportion of automated welfare systems that are harmful. Its claims are about mechanism, not prevalence. Second, the article draws on secondary sources rather than original fieldwork. It does not include the voices of claimants gathered directly, which is a serious gap in a study about being unheard, and one that the author acknowledges as a limitation of method rather than an oversight of intent. Third, the case selection over-represents high-income countries with strong inquiry institutions, precisely because those institutions produce the documentation the method requires. This risks understating harm in settings where no commission will ever report. Fourth, the concept of the empathy deficit is proposed here as an analytical construct and has not been operationalised into a validated measurement instrument. These limitations point directly to a research agenda. Claimant-centred empirical work. Qualitative research with people subjected to automated determinations, focused specifically on the experience of contesting them, would test the four-part framework against lived accounts. Measurement. The four capacities could be operationalised into an audit instrument: time to correction, proportion of overturned decisions that trigger model review, availability of a named accountable officer, and rate of continued recovery during dispute. Comparative institutional analysis. The Danish case suggests that professional bodies and early public debate can change trajectories. Systematic comparison of cases where deployment was halted with cases where it proceeded would identify the institutional conditions of resistance. Generative systems. The introduction of language models into welfare correspondence and triage is happening now and is almost entirely unstudied. The question of whether simulated warmth changes claimant experience, and whether it makes unaccountable decisions easier to accept, is urgent. Non take-up. If administrative burden functions as rationing, then measuring the population of eligible people deterred by automated systems is the most important missing statistic in the field. 10. Conclusion The systems examined in this article were not built by villains. They were built by officials under pressure to save money, catch fraud and modernise, using tools that promised objectivity, and they were built for a population whose lives the designers did not know and did not ask about. The result was a class of administrative artefact that could not be argued with. It could accuse but not listen. It could calculate but not understand. It could decide but not explain. And when it was wrong, which it was on an enormous scale, it had no mechanism by which the truth could reach it, because the person holding the truth was the person it had already decided not to believe. This is what the #empathy_deficit means in practice. It is not the absence of a warm smile at a counter. It is the absence of any point in the machine at which a person's actual life can enter and change the outcome. The corrective is not nostalgia. The caseworker's desk was never a paradise, and there is no route back to it. The corrective is to insist that the design of any system that distributes survival must begin from a different question than the one these systems began from. They asked: how do we stop the wrong people receiving money? The right question is: how will we find out, quickly and reliably, when we have harmed someone, and what will we do then? A welfare system that cannot answer that question has not been made more efficient. It has been made unaccountable, and it has been made so in the name of the poor, at their expense, using their data, with their lives as the margin of error. The task for scholars, technologists and administrators is not to make these systems smarter. It is to make them answerable, and to remember that in public assistance, the measure of a system is never how it treats the case it gets right. It is how it treats the person it gets wrong. Hashtags #EmpathyDeficit #AlgorithmicWelfare #AutomatedDecisionMaking #DigitalWelfareState #PublicAssistance #SocialPolicy #AlgorithmicAccountability #AdministrativeBurden #StreetLevelBureaucracy #DataJustice #WelfareTechnology #AIandSociety #Robodebt #PublicAdministration #TechEthics References Alon-Barkat, S. and Busuioc, M. (2023) Human-AI interactions in public sector decision making: automation bias and selective adherence to algorithmic advice. Journal of Public Administration Research and Theory, 33(1), pp. 153-169. Birhane, A. (2021) Algorithmic injustice: a relational ethics approach. Patterns, 2(2), 100205. Breit, E., Egeland, C., Loberg, I. M. and Rohnebaek, M. T. (2021) Digital coping: how frontline workers cope with digital service encounters. Social Policy and Administration, 55(5), pp. 833-847. Broussard, M. (2023) More Than a Glitch: Confronting Race, Gender, and Ability Bias in Tech. Cambridge, MA: MIT Press. Busuioc, M. (2021) Accountable artificial intelligence: holding algorithms to account. Public Administration Review, 81(5), pp. 825-836. Crawford, K. (2021) Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press. de Boer, N. and Raaphorst, N. (2023) Automation and discretion: explaining the effect of automation on how street-level bureaucrats enforce. Public Management Review, 25(1), pp. 42-62. Dencik, L. (2022) The datafied welfare state: a perspective from the UK. In: New Perspectives in Critical Data Studies. Cham: Palgrave Macmillan, pp. 145-165. Giest, S. and Klievink, B. (2024) More than a digital system: how AI is changing the role of bureaucrats in different organizational contexts. Public Management Review, 26(2), pp. 379-398. Green, B. (2022) The flaws of policies requiring human oversight of government algorithms. Computer Law and Security Review, 45, 105681. Grimmelikhuijsen, S. (2023) Explaining why the computer says no: algorithmic transparency affects the perceived trustworthiness of automated decision-making. Public Administration Review, 83(2), pp. 241-262. Henman, P. (2022) Digital social policy: past, present, future. Journal of Social Policy, 51(3), pp. 535-550. Hoffmann, A. L. (2021) Terms of inclusion: data, discourse, violence. New Media and Society, 23(12), pp. 3539-3556. Ingrams, A., Kaufmann, W. and Jacobs, D. (2022) In AI we trust? Citizen perceptions of AI in government decision making. Policy and Internet, 14(2), pp. 390-409. Inzlicht, M., Cameron, C. D., D'Cruz, J. and Bloom, P. (2024) In praise of empathic AI. Trends in Cognitive Sciences, 28(2), pp. 89-91. Kaun, A. (2022) Suing the algorithm: the mundanization of automated decision-making in public services through litigation. Information, Communication and Society, 25(14), pp. 2046-2062. Loberg, I. M. (2021) Efficiency through digitalization? How electronic communication between frontline workers and clients can spur a demand for services. Government Information Quarterly, 38(2), 101533. Malgieri, G. (2023) Vulnerability and Data Protection Law. Oxford: Oxford University Press. Meijer, A., Lorenz, L. and Wessels, M. (2021) Algorithmization of bureaucratic organizations: using a practice lens to study how context shapes predictive policing systems. Public Administration Review, 81(5), pp. 837-846. Montemayor, C., Halpern, J. and Fairweather, A. (2022) In principle obstacles for empathic AI: why we can't replace human empathy in healthcare. AI and Society, 37(4), pp. 1353-1359. Perry, A. (2023) AI will never convey the essence of human empathy. Nature Human Behaviour, 7(11), pp. 1808-1809. Pors, A. and Schou, J. (2021) Street-level morality at the digital frontlines: an ethnographic study of moral mediation in welfare work. Administrative Theory and Praxis, 43(2), pp. 154-171. Ranchordas, S. (2022) Empathy in the digital administrative state. Duke Law Journal, 71(6), pp. 1341-1389. Sloane, M., Moss, E., Awomolo, O. and Forlano, L. (2022) Participation is not a design fix for machine learning. In: Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. New York: ACM, pp. 1-6. Vallor, S. (2024) The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. Oxford: Oxford University Press. Veale, M. and Zuiderveen Borgesius, F. (2021) Demystifying the draft EU Artificial Intelligence Act. Computer Law Review International, 22(4), pp. 97-112. Wachter, S., Mittelstadt, B. and Russell, C. (2021) Why fairness cannot be automated: bridging the gap between EU non-discrimination law and AI. Computer Law and Security Review, 41, 105567.

Social researchers increasingly want to show the people behind their numbers. Bars and dots feel cold. Photographs feel warm, but they carry a face, and a face carries a name, a family, a legal status, and sometimes a real danger. This article examines a middle path that has become technically possible in the last few years: highly realistic three dimensional human figures that are rendered without recognisable facial features. It asks whether this method can honour both the duty to protect and the duty to represent. Drawing on recent work in information visualization, visual research ethics, data protection law, and human computer interaction, the article argues that removing a face does not by itself remove risk, and that realism itself carries an ethical cost that is often ignored. Facelessness protects identity but can also strip agency, repeating the very erasure that critical scholars have documented in humanitarian and migration imagery. Photorealism increases emotional pull but also increases the chance that viewers will read the model as a document rather than as an interpretation. Generative pipelines that build such figures can quietly import stereotypes about race, gender, poverty, and disability. The article offers a five layer ethical model, a step by step production workflow, a reviewer checklist, and a set of recommendations for supervisors, ethics committees, and journal editors. It is written for students who will soon have to make these choices in their own projects, and it treats the choice of how to draw a human being as a research decision, not a decorative one. Keywords: data visualization ethics; anthropographics; anonymisation; synthetic imagery; vulnerable populations; visual methods; research governance 1. Introduction Every dataset about people is also a dataset about somebody. A row in a spreadsheet describing a fifteen year old asylum seeker who has been waiting nineteen months for a decision is not only a row. It is a person who can be found, questioned, deported, or shamed. Social researchers know this, and the standard response has been to hide the person inside an aggregate. We report percentages. We draw a bar. The bar is safe because the bar does not look like anyone. The problem is that the bar also does not feel like anyone. A growing body of research in #data_storytelling and information visualization argues that abstract encodings distance audiences from the human beings the data describe. This concern gave rise to #anthropographics, a family of visual forms in which each data point is drawn as a human figure rather than as a geometric mark. The idea is simple and old: show the people, and viewers may care more. The idea is also contested, because the way we show people is never neutral, and generic human icons can flatten difference in ways that harm the very groups being represented (Dhawka et al., 2023). Between the cold bar and the identifiable photograph sits a newer option. Modern three dimensional pipelines can produce figures that look like real human bodies, with real skin, real cloth, real light, real posture, and real weight, but with faces that are absent, obscured, turned away, or generated as a non existent composite. This article calls this approach #faceless_representation combined with #photorealistic_modeling. It appears in public health communication, in migration reporting, in reports about homelessness and domestic violence, in museum exhibitions, and increasingly in student dissertations, where free and low cost 3D tools have removed the old technical barrier. The method is attractive for an obvious reason. It seems to solve the central dilemma of visual social research at a stroke. Photographs of participants risk identification. Anonymised photographs, with black bars or blurred faces, protect identity but often produce a disturbing image of a person who has been symbolically defaced. A realistic 3D figure without a face appears to give researchers the emotional presence of the photograph and the protection of the anonymised drawing, without either cost. This article argues that the appearance is misleading, and that the method carries at least four distinct risks that are easy to miss. The first is that facelessness does not equal #de_identification. A body can be identified. Height, build, gait, hair, tattoos, prosthetics, a wheelchair, a hijab, a uniform, a room, a doorway, a landscape behind the figure, and the sheer smallness of the population being described can all point back to one individual, especially in a small community where the study is best known. #re_identification_risk lives in context, not only in faces. The second is that removing the face is itself an act with meaning. Critical work on refugee imagery has shown that obscuring or turning away human subjects is one of the standard visual moves through which people are turned into a mass rather than into individuals (Jhoti and Allen, 2025). #visual_dehumanisation is not a side effect of protection. It can be produced by protection. Research with asylum seeking and refugee children found that young participants wanted their faces shown, wanted authorship of their own images, and experienced compulsory anonymisation as a form of silencing (Korkiamaki and Kaukko, 2023). The third is that photorealism changes how an image is read. A stylised icon announces itself as a symbol. A photorealistic render whispers that it is evidence. When a viewer sees a highly realistic figure standing in a shelter queue, the viewer's default assumption is that somebody stood there. If the figure is a composite, or a plausible average, or an artistic reconstruction of a pattern in the data, then the image is making a claim it cannot support. This is a problem of #epistemic_honesty rather than of privacy. The fourth is that the tools used to build these figures are not empty. Text to image and text to 3D systems trained on internet scale image collections reproduce and amplify social stereotypes at large scale, associating poverty, criminality, and low status occupations with particular racial and national groups (Bianchi et al., 2023). A researcher who types a prompt describing a homeless person, a trafficked woman, or a migrant worker, and accepts what the model returns, has outsourced a representational decision to a system with well documented bias. #stereotype_amplification enters the figure through the pipeline, not through the researcher's intention. The article proceeds as follows. Section 2 reviews the relevant literature across four fields that rarely speak to each other. Section 3 sets out a five layer ethical model. Section 4 explains the method used to build the argument. Section 5 presents a critical analysis organised around six tensions. Section 6 translates the analysis into a practical workflow that a student can follow. Section 7 discusses the trade offs that cannot be resolved, only managed. Section 8 states the limits of the argument. Section 9 gives recommendations for researchers, supervisors, #ethics_committees, and editors. Section 10 concludes. The central claim is straightforward. A faceless photorealistic figure is not a neutral container for data. It is an argument about who the person in the data is. Students should be taught to defend that argument in the same way they defend a sampling strategy or a coding frame. 2. Background and Literature Review 2.1 From marks to bodies: the anthropographic turn For most of the history of statistical graphics, the human being disappeared into the mark. A dot was a death, a bar was a district, and the discipline treated this abstraction as a virtue. Over the past decade, visualization researchers have questioned whether that virtue has a cost. The argument is that visual abstraction may reduce the felt significance of human suffering, and that showing figures shaped like people may restore some of it. Recent research has moved past the early enthusiasm. Dhawka, He, and Willett (2022) examined anthropographic projects about marginalised groups and identified three recurring problems: the figures used are homogeneous, so that a Somali grandmother and a Guatemalan teenager are drawn identically; marginalisation is treated as a single flat attribute rather than as an intersecting set of conditions; and the design decisions are made almost entirely by a small and demographically narrow group of designers, researchers, and journalists. In a follow up study they proposed diverse anthropographics, in which visible characteristics of the figures are mapped to demographic attributes in the data, and they were explicit that this proposal is difficult and risky, because it requires representative datasets that mostly do not exist and because it invites harmful visual essentialism (Dhawka et al., 2023). Later work explored generative methods for producing such figures at scale, which is precisely the technical moment this article is concerned with (Dhawka et al., 2024). The practitioner side has been studied as well. Interviews with seventeen visualization designers working with race and gender data found deep unease about human shaped forms. One designer put the dilemma in plain terms: icons let people see the individual instead of an abstract shape, but they also open many ways to misrepresent people or to hide their experience (Dhawka and Dasgupta, 2025). The same study found that designers often held incompatible commitments, describing themselves as neutral communicators while making choices that were unavoidably political, and that designers who did not share the identities of the people in the data felt they lacked the standing to defend their own representational choices. Related work has traced how #power_asymmetry runs through visualization collaborations (Akbaba et al., 2023) and how feminist theory reframes what a visualization is for (Akbaba et al., 2024). Two further findings from this literature matter for the present argument. First, sociodemographic categories are constructed, unstable, and politically charged, and visualizing them is not a matter of applying best practice but of navigating contested ground (Cabric et al., 2023; Cabric et al., 2024). Second, public visualizations can be seized and reworked by bad actors, which means that #misinformation_risk is a design consideration and not only a communications problem (Lee et al., 2021; Lisnic et al., 2023; Lisnic et al., 2024). A humanising figure of a migrant, released as an open asset, can be re captioned by anyone. 2.2 Visual research ethics and the anonymity problem Visual methods in social research have always struggled with #participant_anonymity. Photography, photo elicitation, and #photovoice were developed partly to shift power towards participants, letting people show their own lives rather than be described by an outsider. Yet the moment a photograph exists, so does an identifiable face, and the standard machinery of research governance responds by demanding that the face be removed. Korkiamaki and Kaukko (2023) documented what happens next. Working with asylum seeking and refugee children in Finland and Australia, they found that participants wanted to be seen, wanted to be named, and understood perfectly well what they were choosing. Anonymisation, applied without thought, produced children who were not only faceless but voiceless, and it blocked the message the children were trying to send. The authors do not argue that anonymity should be abandoned. They argue that automatic anonymisation is itself an ethical act requiring justification, and that the balance between protection and recognition must be negotiated rather than assumed. This challenges a common assumption in research training, which is that more anonymity is always more ethical. It is not. Anonymity is a tool for reducing harm. Where it reduces harm, it is right. Where it silences, misrepresents, or removes credit from people who want credit, it is not automatically right, and the researcher owes an argument. The relevant concept here is #dignity, which includes both protection from exposure and recognition as an author of one's own story. There is a second body of evidence about how vulnerable groups are pictured once they leave the participant's hands. Analyses of how refugee serving organisations represent refugees on social media identify obscuring techniques, such as photographing subjects from behind or blurring faces, alongside recurring visual tropes such as the mother and child and the overloaded boat (Jhoti and Allen, 2025). The same combination of protection and cliche appears across #humanitarian_imagery, and a framework for the responsible use of imagery in global health has argued that the field routinely reproduces images that strip context, agency, and specificity from the people depicted (Charani et al., 2023). #refugee_representation is therefore not an empty category into which a researcher can drop a new technology. It is a field with a documented history of harm. 2.3 Realism, avatars, and the perception of synthetic humans If the figure is going to be realistic, the researcher enters a third literature, which studies how people perceive synthetic humans. Three findings are relevant. The first is the #uncanny_valley. As a synthetic human approaches but does not reach convincing realism, viewers report eeriness, discomfort, and reduced trust rather than increased sympathy. This effect has been documented repeatedly in virtual reality and avatar research, and it is sensitive to inconsistency: a figure whose skin is photoreal but whose motion is stiff, or whose body is detailed but whose face is a smooth blank, can trigger exactly the response the researcher was trying to avoid (Mal et al., 2024). A faceless photorealistic body is, by construction, an inconsistent human. This is an empirical risk that any project using #3D_avatars should test rather than assume away. The second is that realism interacts with personalisation and identification. Work on virtual twins found that photorealism and personalisation change embodiment, self identification, and how the figure is perceived, and that these effects are not uniformly positive (Salagean et al., 2023). Higher realism is not simply better. It shifts what the viewer thinks the figure is. The third is that realism raises the evidential temperature of an image. Concerns about #deepfake_concerns and synthetic media have made audiences aware that convincing images can be fabricated, and this awareness cuts both ways. Some viewers will believe a synthetic figure that they should not believe. Others will disbelieve genuine testimony because it could have been synthetic. Qualitative researchers have begun to describe how this environment corrodes trust in interview based work and pushes institutions towards verification practices that themselves fall hardest on marginalised participants (Santinele Martino and Hardy, 2026). A discipline that fills its journals with photoreal synthetic humans should expect a cost to its own credibility. 2.4 Synthetic data, anonymisation, and the law The fourth literature concerns whether synthetic outputs are legally and technically anonymous. The short answer from data protection scholarship is that they are not automatically anonymous. #synthetic_data is generated by a model trained on real records. The records are new. The patterns are not. Where a model overfits, it can memorise unusual cases, and the most exposed records are exactly the rare ones, which in social research are usually the records of the most vulnerable people. Legal analysis of the European framework concludes that synthetic data sits uneasily inside the binary distinction between personal and non personal data, that #re_identification_risk cannot be assumed away, and that the burden of demonstrating anonymity falls on the party claiming it (Beduschi, 2024). Under the logic of #GDPR, information is anonymous only when singling out, #linkage_attacks, and inference are all sufficiently unlikely, and a synthetic artefact that reproduces a rare combination of attributes may fail all three tests. The relevance to visual work is direct. If a researcher builds a 3D figure by scanning a real participant and then deleting the face, the underlying geometry is still derived from that person's body. If a researcher builds a figure by prompting a generative system trained on scraped photographs, the output may carry traces of individuals who never consented to anything. #model_memorisation is a live concern in both cases, and a claim that an image is anonymous because it was generated is not a claim that survives scrutiny. The governance literature adds a further layer. The CARE principles for Indigenous data governance insist that communities hold authority to control data about them, that data work must produce collective benefit, and that openness is not an unqualified good (Carroll et al., 2021). #Indigenous_data_sovereignty and the wider #CARE_principles framework offer a model that translates well to other groups who have been researched more than they have been served, because they shift the question from what the researcher may do to what the community has authorised. 2.5 The gap These four literatures rarely meet. Visualization research knows a great deal about human shaped marks and almost nothing about research governance. Visual methods scholarship knows a great deal about consent and almost nothing about rendering. Avatar research knows a great deal about perception and almost nothing about vulnerable populations. Data protection scholarship knows a great deal about identifiability and treats images as an afterthought. The result is that a student who wants to represent a vulnerable population with a faceless photorealistic figure will find no guidance that addresses the decision as a whole. This article attempts to supply it. 3. Conceptual Framework: Five Layers of Ethical Obligation The framework proposed here treats a faceless photorealistic figure as passing through five layers, each of which can fail independently. A figure is defensible only if it survives all five. Layer 1: Identifiability The first question is whether any real person can be singled out, linked, or have an attribute inferred about them from the image. This layer includes the face, but the face is only its most obvious element. The relevant checks are cumulative rather than individual. A tattoo alone may not identify. A tattoo plus a specific prosthetic plus a named district plus a sample of eleven people may. Researchers should treat the #small_numbers_problem in images exactly as they treat it in tables: when a cell is too small to publish as a number, it is too small to publish as a body. #aggregation, in the visual domain, means building figures that represent patterns across many people rather than portraits of one. #composite_figures are safer than reconstructions. Practical items belonging to this layer include body shape, gait if the figure is animated, hair, clothing, religious dress, occupational uniform, visible disability and assistive equipment, scars, background architecture, and the metadata attached to the image file. #provenance data embedded in a render can be as revealing as the render. Layer 2: Consent The second question is whether the people whose data produced the figure agreed to be depicted in this way. Standard consent forms do not cover this, because they were written for text. This article proposes #consent_layers rather than a single signature. Layer one is consent to participate. Layer two is consent to have one's data used in a visual composite. Layer three is consent to a specific visual treatment, shown to the participant as a draft before publication. Layer four is consent for reuse, which must specify whether the figure may be released as an open asset, animated, placed in a virtual environment, or licensed to third parties. #withdrawal_rights must be addressed honestly. Once a photoreal figure is published and downloaded, it cannot be recalled. Researchers should say so plainly at the point of consent rather than promising a deletion they cannot deliver. This is a limit of the medium, and hiding it is a form of deception. Layer 3: Representation The third question is whether the figure represents the population fairly. This is where #stereotype_amplification enters. Every visible attribute of the figure is a claim. #skin_tone is a claim. #phenotype is a claim. Age, weight, posture, clothing condition, and facial absence are claims. If the researcher did not decide these attributes deliberately, then a default decided, and defaults in generative systems have been shown to skew towards stereotype (Bianchi et al., 2023). A figure representing survivors of domestic violence that shows a thin, young, visibly bruised woman is making an empirical claim about who survivors are, and that claim is usually false. Two design strategies help. The first is the #abstraction_gradient: deliberately choosing how realistic each element of the figure needs to be, and reducing realism wherever realism adds no analytic value. Skin can be photoreal while clothing is neutral. Bodies can be realistic while environments are schematic. The second is #representation_fidelity: matching the distribution of visible attributes across a set of figures to the distribution in the data, and documenting the mapping. If ten figures represent one hundred people, the researcher should be able to say what each figure encodes. Layer 4: Epistemic status The fourth question is whether the audience can tell what kind of image they are looking at. A photograph is a trace of a moment. A photoreal render is a model of a claim. Audiences cannot see the difference, so the researcher must tell them. This layer is served by #disclosure_labels: a visible, plain language statement attached to every figure that says how it was made, what it encodes, and what it does not encode. It is also served by #transparency about the pipeline, including which software, which prompts, which reference images, and which decisions were made by a model rather than by a person. #research_integrity in visual work means the same thing it means in statistical work, which is that a reader should be able to reconstruct what you did. Layer 5: Authority and benefit The fifth question is who decided, and who gains. Drawing on the CARE framework (Carroll et al., 2021), this layer asks whether the community depicted had #community_authority over the representation, and whether the project produces #benefit_sharing rather than extraction. In practice this means #co_design of the figures with community members, a right of refusal that is real rather than nominal, and a plan for returning the outputs to the community in a usable form. It also means #reflexivity about who the researcher is. Designers who do not share the identities of the people they depict have reported that they lack the standing to defend their choices when challenged (Dhawka and Dasgupta, 2025), and the correct response to that is not to avoid the work but to share the decision. Together, the five layers convert a single vague obligation, do no harm, into five questions with answers that can be written down, reviewed, and criticised. 4. Methodology This article is a conceptual and critical review rather than an empirical study, and it is important to be exact about what that means and what it cannot deliver. Design. The article uses a structured narrative review combined with a design analysis. The narrative review gathers evidence from four fields that do not normally cite each other: information visualization, visual research methods, avatar and virtual human perception, and data protection scholarship. The design analysis takes the production pipeline for a faceless photorealistic figure and examines each stage for ethical failure points, in the manner of a hazard analysis. Selection of literature. Sources were selected on four criteria. First, recency: the review privileges work published within roughly the last five years, because both the technology and the regulatory environment have changed quickly. Second, relevance: sources had to address either the representation of people in data, the ethics of visual research with vulnerable groups, the perception of synthetic humans, or the identifiability of synthetic outputs. Third, standing: peer reviewed conference and journal publications were preferred, with a small number of preprints included where they represent the only substantive treatment of a topic. Fourth, criticality: the review deliberately includes work that is sceptical of humanising visualization, because an argument built only on advocates would be worthless. Analytic approach. Evidence was organised around tensions rather than themes. A tension is a place where two defensible values pull in opposite directions, such as protection against recognition, or realism against honesty. This choice reflects the nature of the object. There is no correct answer to the question of how to draw a vulnerable person, only a set of defensible positions with different costs, and a framework that pretended otherwise would mislead students. Position of the author. Following the argument about #positionality made in the visualization literature, it should be stated plainly that this article is written from within the research community that produces such images rather than from within the communities that are depicted by them. That position gives access to the production process and limits access to the experience of being represented. The recommendations in Section 9 are shaped by that limit, and they lean towards mechanisms, such as #participatory_design and community review, that transfer decisions to people better placed to make them. What this method cannot do. This article cannot tell you whether audiences respond better to faceless photoreal figures than to icons, bars, or photographs. That is an empirical question requiring controlled study, and the existing evidence on #empathy_effects in anthropographics is mixed enough that no confident prediction is warranted. What the article can do is set out what must be decided, what can go wrong, and what a defensible decision looks like. 5. Analysis 5.1 What facelessness protects, and what it does not The first finding is that the protective power of removing a face is routinely overestimated. Consider a study of street homelessness in a mid sized city, with a sample of thirty participants recruited through two shelters. A researcher builds a photoreal figure of a man in his fifties, seated, wearing a distinctive jacket, with a visible below knee prosthesis, positioned outside a recognisable building. The face is absent. Every shelter worker in that city will know who it is. So will the police. So will the man's family. #de_identification has not occurred, because identification in small communities runs through context, not through faces. Now consider the same figure with the prosthesis removed, the jacket replaced with generic clothing, and the building replaced with a schematic doorway. The figure is now safer, and it is also less true. It no longer shows a disabled man, and #disability_representation has been quietly erased in the name of protection. This is the trade off in its rawest form, and there is no way around it. Every step towards safety removes information, and some of the information removed is the information that mattered. The practical conclusion is that facelessness must be treated as one control among many, not as the control. Researchers working with #homelessness_data, #mental_health_data, trafficking, undocumented migration, or any domain where exposure carries legal or physical danger should conduct an identifiability assessment on the whole image, and should assume an adversary who knows the community. #surveillance_harm is the relevant threat model, and it does not require a face. Two further points follow. First, #blurring and #pixelation applied to photographs are weaker than they look, and have been shown to be partially reversible under some conditions, which is one reason synthetic reconstruction has become attractive in the first place. Second, if a figure is animated, #motion_capture derived from a real participant reintroduces an identifier, because gait is distinctive. Where animation is used, motion should be synthesised or drawn from a library rather than captured from the person depicted. The same logic applies to #voice_anonymisation in any accompanying audio. 5.2 The erasure that protection produces The second finding is that facelessness carries meaning, and the meaning is rarely the one the researcher intends. The literature on refugee imagery is unambiguous on this point. Turning subjects away from the camera, cropping heads, and blurring faces are among the standard techniques through which people become an undifferentiated mass (Jhoti and Allen, 2025). The visual grammar of the faceless body is already occupied. It says: this is not a person, this is a type. When a researcher adopts that grammar for protective reasons, the protective intention does not travel with the image. The audience sees the convention, not the ethics committee minutes. The evidence from participants themselves sharpens this. Refugee and asylum seeking children in two countries wanted their faces visible and experienced the removal of their faces as the removal of their voices (Korkiamaki and Kaukko, 2023). It is worth sitting with the discomfort of that finding. The ethical apparatus designed to protect them was experienced by them as a harm. The lesson is not that anonymity is wrong but that it is a decision with costs, and that the people who bear the costs should have a say in whether the costs are worth paying. Where the participants are #child_participants, or where legal risk is severe, or where the community itself prefers invisibility, protection will usually win. The point is that it should win an argument rather than win by default. There is a design response available. Facelessness need not mean blankness. A figure can be turned so that the face is out of frame, occluded by a natural element, or lost in shadow, and this reads differently from a smooth featureless head. The former says the moment was private. The latter says the person has been deleted. #body_language, posture, gesture, and the direction of attention can carry an enormous amount of individuality without carrying identity. A figure leaning forward, hands active, gaze directed at something outside the frame, reads as an agent. The same figure slumped and static reads as a victim. The researcher chooses. 5.3 The realism paradox The third finding is that photorealism helps and harms in the same movement. Realism helps because it defeats the flattening that generic icons produce. Homogeneous human shapes obscure real differences between groups and can misrepresent the populations they claim to show (Dhawka et al., 2022). A photoreal figure can carry age, build, dress, and condition with a specificity that an icon cannot. Realism harms in three ways. First, it invites the #uncanny_valley. A body rendered with high fidelity and a head rendered with none is precisely the kind of inconsistency that produces eeriness rather than sympathy, and the perception literature suggests that inconsistent realism is worse than uniformly low realism (Mal et al., 2024). A student who assumes that a better render produces a better emotional response may get the opposite. Second, realism raises the evidential stakes. The more an image looks like a photograph, the more the audience treats it as a record. If the figure is a composite of many participants, then the audience is being invited to believe in a person who does not exist. This is not automatically unethical, because reconstruction is a legitimate scientific practice, but it becomes unethical the moment it is not declared. Third, realism narrows interpretation. #audience_interpretation research has repeatedly shown that visual specificity forecloses readings. A schematic figure allows the viewer to fill the gap with the population they know. A photoreal figure tells them who to imagine, and if the researcher has chosen badly, the audience now has a specific and possibly false image of a whole group in their head. This produces the design principle of the #abstraction_gradient stated earlier. Realism should be spent where it earns something, and withheld everywhere else. If the analytic point concerns the physical toll of manual labour, then render the hands and the posture with care and leave the setting schematic. If the analytic point concerns overcrowding, render the space and keep the bodies simple. Realism is a budget, not a setting. 5.4 Stereotype leakage in the generative pipeline The fourth finding concerns the tools. A researcher who builds a figure with a generative system is not filling an empty vessel. Text to image systems reliably produce stereotyped output: they associate poverty and low status occupations with particular racial groups, associate whiteness with attractiveness and default humanity, and reproduce national and cultural cliches even when prompts are neutral, and mitigation strategies including counter stereotypical prompting and system level guardrails do not reliably remove the effect (Bianchi et al., 2023). The implication for our case is direct. A prompt such as a homeless man, a trafficked woman, a slum dweller, a mentally ill patient, or a refugee family will return a figure that encodes the internet's assumptions about those categories. If the researcher accepts the return, then the published figure is a picture of a stereotype wearing the authority of a research output. #generative_models are, in this respect, a hazard, and they are a hazard that scales, because the same defaults appear across thousands of projects. Three controls are available. The first is prompt discipline. Describe the figure using attributes drawn from the dataset rather than from the category label. Do not prompt for a refugee. Prompt for a woman of forty two, wearing a winter coat, carrying a rucksack, standing, because that is what the data record says, and let the category emerge from the context of the visualization rather than from the appearance of the body. The second is audit. Generate multiple candidates, compare the distribution of visible attributes across them against the distribution in the dataset, and record the mismatch. If the model produces ten figures and nine of them are young, and the population is mostly over fifty, the model has told you something about itself and nothing about your participants. The third is human decision. Every visible attribute that is not derived from data should be set by a person who can be named in the methods section. #accountability requires an author, and a prompt is not an author. 5.5 Consent that travels The fifth finding is that consent for visual synthetic representation must be built to survive the life of the image, and most consent processes are not. Text has a short reach. A quotation in a thesis is read by twelve people and then rests in a repository. A photoreal 3D figure has a long reach. It can be exported, animated, embedded in a virtual environment, used in a press release, licensed by a charity, recaptioned by a political campaign, and eventually placed in an archive where it will outlive the study, the researcher, and possibly the participant. #archival_ethics and the emerging discussion of #digital_afterlife matter here, because a figure derived from a person continues to represent that person after the person can no longer object. Three practical rules follow. Rule one is specificity. The consent document should name the medium, show an example, and state the reuse conditions. Participants cannot consent to something they have not seen. Rule two is honesty about irreversibility. If the figure will be released openly, the researcher must say that withdrawal will not be possible in practice. If withdrawal must remain possible, then the figure must not be released openly. These two commitments cannot both be honoured, and pretending otherwise is a failure of #transparency. Rule three is proportionality, which is the visual form of #data_minimisation. Ask what the figure adds that a chart, a quotation, or a schematic diagram would not add. If the answer is that it makes the report more compelling, that is a real answer, but it should be weighed against a real cost. If the answer is that it is expected, or that it looks professional, then the figure should not be made. #harm_reduction begins with not creating the artefact in the first place. 5.6 Governance, law, and the limits of the word anonymous The sixth finding is that researchers should stop describing synthetic figures as anonymous and should start describing them as de identified to a stated standard. Legal analysis of synthetic data concludes that synthetic outputs are not automatically outside the scope of #data_protection law, that identifiability is a matter of residual risk rather than of category, and that the risk is concentrated in exactly the rare and unusual records that vulnerable populations tend to occupy (Beduschi, 2024). A figure built from a scan of one participant is derived personal data. A figure built from a model trained on that scan may still carry traces of it. What follows for practice is a documentation duty. Every published figure should be accompanied by a short technical note recording the source of the geometry, whether real scans were used, whether a generative model was used and which, what identifying attributes were suppressed, and what residual risk the researchers judge to remain. This is not bureaucratic decoration. It is the visual equivalent of reporting a confidence interval, and it lets #peer_review do its job. At present, reviewers cannot assess the ethics of a figure because they are given no information about how it was made. The governance frameworks that already exist for other kinds of sensitive data are the obvious model. The CARE principles hold that communities have authority to control data about themselves and that openness must yield to that authority where the two conflict (Carroll et al., 2021). Applied to visual synthesis, this means that a community may legitimately refuse to be depicted at all, and that the refusal ends the matter. #open_data is a value, not an obligation that overrides consent. 6. A Practical Workflow This section converts the analysis into a sequence a student can actually follow. It assumes a project involving a vulnerable population, an intention to publish visual figures, and access to standard 3D or generative tooling. Step 1. Justify the figure before making it. Write one paragraph stating what the figure will show that a chart or a quotation cannot. If the paragraph is weak, stop. Most projects do not need photoreal humans, and a project that adds them for atmosphere has taken a risk for no analytic gain. Step 2. Choose the level of abstraction deliberately. Place your project on a scale running from schematic icon, through stylised figure, through non photoreal render, to photoreal render. Justify the position. Move down the scale unless there is a specific reason to move up. Remember that the #abstraction_gradient can vary within a single image. Step 3. Decide what the figure encodes. List every visible attribute the figure will have and mark each one as either (a) derived from the data, (b) required for legibility, or (c) arbitrary. Category (c) should be as small as possible, and everything in it should be set to a neutral value rather than left to the model. Write this list down. It becomes a table in your appendix. Step 4. Run an identifiability assessment. With someone who knows the community, review the draft figure and ask: could a person who knows this population identify anyone here? Consider body, dress, assistive equipment, marks, setting, and the size of the sample. Assume the reviewer is hostile. Record the result. Step 5. Build with a documented pipeline. Record the software, the model versions, the prompts, the reference material, and the #render_settings. If you used a generative system, keep the seeds. You will need all of this for the technical note, and you will need it again if a reviewer or a participant challenges the figure. Step 6. Audit for stereotype. Generate several candidates. Compare their visible attributes with the actual distribution in your data. Where the model has drifted towards stereotype, correct it by hand and say that you did. Step 7. Show the participants. Before publication, show the figures to participants or to a community panel. Ask an open question: does this look like us, and do you want it published. Record the answer. If the answer is no, do not publish. This step is the single most effective control in the whole workflow, and it is the one most often skipped. Step 8. Label the figure. Attach a caption that states, in plain words, that the figure is a synthetic composite, that no individual is depicted, what it encodes, and that faces have been omitted deliberately. Do not bury this in a footnote. #visual_literacy cannot be assumed, and a reader who mistakes a render for a photograph has been misled by you. Step 9. Set the reuse terms. Decide, in advance, whether the asset may be downloaded, animated, or reused, and license it accordingly. If in doubt, publish the image and withhold the model. Step 10. Write the technical note. Half a page in the appendix, covering source data, pipeline, suppressed attributes, residual risk, participant review, and reuse terms. This is what makes the figure reviewable, and reviewability is what separates a research image from an illustration. A short reviewer checklist follows from the same logic, and journals could adopt it as it stands. Does the paper justify the use of human figures? Does it state whether real scans were used? Does it document the generative pipeline? Does it report an identifiability assessment? Does it report participant or community review of the figures? Does the caption disclose the synthetic status? Are reuse terms stated? Seven questions, none of them onerous, and at present almost none of them are asked. 7. Discussion The tensions identified here do not resolve. They are managed, and management means choosing which cost to pay. Protection against recognition. The stronger the protection, the greater the erasure. Facelessness protects identity and removes agency. There is no setting at which both are maximised. What can be done is to move the decision to the people affected, to preserve individuality through posture and context rather than through faces, and to say openly in the text that the figures are protective devices rather than portraits. Realism against honesty. The more convincing the figure, the more it is read as evidence, and the more work the label has to do. #representation_fidelity and evidential modesty pull against each other. The answer is not to abandon realism but to spend it deliberately and to declare it loudly. Individuality against generalisation. A single vivid figure moves an audience. A single vivid figure also invites the audience to generalise from an instance the researcher invented. Multiple figures, whose distribution is documented, are safer than one, and they are less powerful. That is a real loss, and researchers should own it rather than pretend that the safest choice is also the most effective one. Technical possibility against professional norm. The tools have advanced far ahead of the guidance. A student today can produce a photoreal human in an afternoon, and no ethics form in most institutions will ask a single question about it. The gap between what is possible and what is governed is where harm accumulates, and it is closing far too slowly. There is a wider point about what visualization is for. The visualization literature has spent a decade arguing that charts are not neutral, that categories are constructed, and that designers carry #situated_knowledge into their work whether they acknowledge it or not (Akbaba et al., 2024; Dhawka and Dasgupta, 2025). Photoreal human figures make this argument concrete in a way that bars never could. When you draw a body, you cannot hide behind objectivity. You have decided what a poor person looks like. The only responsible response is to admit that you decided, to explain how, and to let the people you drew tell you whether you got it wrong. It is worth noting that a good deal of the discomfort surrounding this method is productive discomfort. The unease a researcher feels when generating a face free body of a trafficking survivor is the correct response, and it is information. Methods that feel too easy in this domain usually are. 8. Limitations Four limits apply to the argument made here. First, it is conceptual. No audiences were tested, no participants were interviewed, and no figures were evaluated. The claims about how viewers respond to inconsistent realism are drawn from adjacent literatures in virtual reality and avatar perception, and their transfer to static research figures is plausible but unverified. Second, the evidence base is uneven. The visualization literature on anthropographics is small, recent, and concentrated in a handful of research groups. The visual methods literature is rich but rarely engages with 3D synthesis. Conclusions drawn across such different bodies of work should be held loosely. Third, the framework is Western in its legal assumptions. The data protection reasoning draws on the European tradition, and the consent architecture assumes an individual rights model that does not fit every setting. The CARE principles gesture towards collective governance, but a fully pluralistic framework would require input from communities that this article has not consulted. Fourth, the technology is moving. Any specific claim about what generative systems can or cannot produce will date quickly. The ethical structure proposed here is intended to outlast the tools, but the reader should test it against whatever tools they actually have. 9. Recommendations For students and researchers. Treat the figure as a research decision. Justify it, document it, show it to the people in it, and label it. Do not describe a synthetic figure as anonymous. Describe it as de identified to a standard you can state. Assume that anything you publish will be reused by someone whose politics you would not choose, and design for that. For supervisors. Ask the seven reviewer questions before the student builds anything, not after. Teach the #abstraction_gradient as a normal design skill. Introduce the anthropographics literature alongside the standard visualization curriculum, so that students meet the critique at the same time as the technique. For ethics committees. Existing forms are built for text and for photographs. They do not ask whether synthetic human figures will be produced, whether real scans will be used, whether generative models will be involved, or what the reuse terms will be. Four additional questions would close most of the gap, and they cost nothing. #trauma_informed_research standards should extend to visual outputs, including the psychological safety of the researchers who make them, since designers working on distressing visual material have described real harm from being asked to render other people's suffering (Dhawka and Dasgupta, 2025). For journals and editors. Require a technical note for any synthetic human figure. Require disclosure in the caption. Extend image manipulation policies, which currently address the doctoring of photographs, to cover the generation of humans. A discipline that polices the brightness of a gel blot and ignores the fabrication of a refugee has its priorities inverted. For communities and participant groups. The strongest control in the entire framework is the right to say no, and to say no late. Groups working with researchers should ask to see visual outputs before publication as a standard condition, and should ask what happens to the asset afterwards. #design_justice in this domain begins with the ability to refuse. 10. Conclusion Faceless photorealistic modelling is a genuinely useful method. It offers a way to show human beings in social data without exposing them, and it responds to a real failure of abstract charts, which is that they make suffering easy to look past. Nothing in this article argues that the method should be abandoned. What the article argues is that the method has been adopted faster than it has been understood. Removing a face does not remove identifiability, because bodies, settings, and small samples identify people. Removing a face is not neutral, because the faceless body is already a figure of speech in the visual language of humanitarian reporting, and what it says is that these people are a category. Photorealism does not simply increase empathy, because inconsistent realism can repel, and because a convincing image makes a claim to be evidence that a composite cannot honour. Generative tooling does not solve representation, because it arrives pre loaded with the stereotypes of the material it was trained on. The answer is not a rule but a discipline. Justify the figure. Choose the realism you need and no more. Derive every visible attribute from data or set it deliberately. Assess identifiability against a hostile reader who knows the community. Show the figure to the people it describes and accept their refusal. Label it honestly. Document how it was made. Decide, in advance, who may use it afterwards. None of this is exotic. It is the ordinary machinery of research method, applied to a decision that has so far escaped it. The bodies we draw in our reports are arguments about who the people in our data are. Students entering this field should learn to make those arguments carefully, to write them down, and to be willing to be told that they got them wrong. #faceless_3D_modeling #ethical_visualization #vulnerable_groups_research #social_data_visualization #anonymity_in_research #synthetic_human_figures #representation_ethics #anthropographic_design #research_methods_students #privacy_by_design #visual_consent #critical_data_studies #photoreal_avatars_in_research #data_ethics_education #responsible_research_imagery References Akbaba, D., Klein, L., and Meyer, M. (2024). Entanglements for visualization: Changing research outcomes through feminist theory. IEEE Transactions on Visualization and Computer Graphics, 31(1). doi:10.1109/TVCG.2024.3456171 Akbaba, D., Lange, D., Correll, M., Lex, A., and Meyer, M. (2023). Troubling collaboration: Matters of care for visualization design study. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3544548.3581168 Andrus, M., Spitzer, E., Brown, J., and Xiang, A. (2021). What we cannot measure, we cannot understand: Challenges to demographic data procurement in the pursuit of fairness. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 249 to 260. doi:10.1145/3442188.3445888 Andrus, M., and Villeneuve, S. (2022). Demographic reliant algorithmic fairness: Characterizing the risks of demographic data collection in the pursuit of fairness. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 1709 to 1721. doi:10.1145/3531146.3533226 Bako, H. K., Liu, X., Battle, L., and Liu, Z. (2023). Understanding how designers find and use data visualization examples. IEEE Transactions on Visualization and Computer Graphics, 29(1), 1048 to 1058. doi:10.1109/TVCG.2022.3209490 Beduschi, A. (2024). Synthetic data protection: Towards a paradigm change in data regulation? Big Data and Society, 11(1). doi:10.1177/20539517241231277 Berret, C., and Munzner, T. (2024). Iceberg sensemaking: A process model for critical data analysis. IEEE Transactions on Visualization and Computer Graphics. doi:10.1109/TVCG.2024.3486613 Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J., and Caliskan, A. (2023). Easily accessible text to image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1493 to 1504. doi:10.1145/3593013.3594095 Cabric, F., Bjarnadottir, M. V., Cabouat, A. F., and Isenberg, P. (2023). Open questions about the visualization of sociodemographic data. In 2023 IEEE Workshop on Visualization for Social Good, 16 to 20. doi:10.1109/VIS4Good60218.2023.00010 Cabric, F., Bjarnadottir, M. V., Ling, M., Rafnsdottir, G. L., and Isenberg, P. (2024). Eleven years of gender data visualization: A step towards more inclusive gender representation. IEEE Transactions on Visualization and Computer Graphics, 30(1), 316 to 326. doi:10.1109/TVCG.2023.3327369 Carroll, S. R., Herczog, E., Hudson, M., Russell, K., and Stall, S. (2021). Operationalizing the CARE and FAIR principles for Indigenous data futures. Scientific Data, 8, 108. doi:10.1038/s41597-021-00892-0 Charani, E., et al. (2023). The use of imagery in global health: An analysis of infectious disease documents and a framework to guide practice. The Lancet Global Health, 11(1), e155 to e164. Chen, Y. T., Smith, A. D. R., Reinecke, K., and To, A. (2023). Why, when, and from whom: Considerations for collecting and reporting race and ethnicity data in HCI. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3544548.3581122 Dhawka, P., and Dasgupta, S. (2025). The social construction of visualizations: Practitioner challenges and experiences of visualizing race and gender demographic data. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3706598.3713243 Dhawka, P., He, H. A., and Willett, W. (2022). Representing marginalized populations: Challenges in anthropographics. arXiv preprint arXiv:2210.02660. Dhawka, P., He, H. A., and Willett, W. (2023). We are the data: Challenges and opportunities for creating demographically diverse anthropographics. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3544548.3581086 Dhawka, P., Perera, L., and Willett, W. (2024). Better little people pictures: Generative creation of demographically diverse anthropographics. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3613904.3641957 Feinberg, M. (2022). Everyday Adventures with Unruly Data. MIT Press. He, H. A., Walny, J., Thoma, S., Carpendale, S., and Willett, W. (2024). Enthusiastic and grounded, avoidant and cautious: Understanding public receptivity to data and visualizations. IEEE Transactions on Visualization and Computer Graphics, 30(1), 1435 to 1445. doi:10.1109/TVCG.2023.3326917 Jablonski, N. G. (2021). Skin color and race. American Journal of Physical Anthropology, 175(2), 437 to 447. doi:10.1002/ajpa.24200 Jhoti, A., and Allen, W. L. (2025). Visual bordering: How refugee serving organizations represent refugees on Instagram. New Media and Society. doi:10.1177/14614448231220856 Korkiamaki, R., and Kaukko, M. (2023). Faceless, voiceless child: Ethics of visual anonymity in research with children and young people. Childhood, 30(1), 55 to 70. doi:10.1177/09075682221126586 Lee, C., Yang, T., Inchoco, G. D., Jones, G. M., and Satyanarayan, A. (2021). Viral visualizations: How coronavirus skeptics use orthodox data practices to promote unorthodox science online. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3411764.3445211 Lee-Robbins, E., and Adar, E. (2022). Affective learning objectives for communicative visualizations. IEEE Transactions on Visualization and Computer Graphics, 29(1), 1 to 11. doi:10.1109/TVCG.2022.3209500 Liang, C. A., Munson, S. A., and Kientz, J. A. (2021). Embracing four tensions in human computer interaction research with marginalized people. ACM Transactions on Computer Human Interaction, 28(2), Article 14. doi:10.1145/3443686 Linxen, S., Sturm, C., Bruhlmann, F., Cassau, V., Opwis, K., and Reinecke, K. (2021). How WEIRD is CHI? In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3411764.3445488 Lisnic, M., Lex, A., and Kogan, M. (2024). Yeah, this graph doesn't show that: Analysis of online engagement with misleading data visualizations. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3613904.3642448 Lisnic, M., Polychronis, C., Lex, A., and Kogan, M. (2023). Misleading beyond visual tricks: How people actually lie with charts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3544548.3580910 Mal, D., Dollinger, N., Wolf, E., Wenninger, S., Botsch, M., Wienrich, C., and Latoschik, M. E. (2024). Am I the odd one? Exploring incongruencies in the realism of avatars and virtual others in virtual reality. Frontiers in Virtual Reality, 5, 1417066. Rose, G. (2022). Visual Methodologies: An Introduction to Researching with Visual Materials (5th ed.). Sage. Salagean, A., Crellin, E., Parsons, M., Cosker, D., and Stanton Fraser, D. (2023). Meeting your virtual twin: Effects of photorealism and personalization on embodiment, self identification and perception of self avatars in virtual reality. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. doi:10.1145/3544548.3581182 Santinele Martino, A., and Hardy, K. (2026). Authenticity in question: Navigating qualitative interviewing in the era of artificial intelligence and deepfakes. Sociological Research Online. doi:10.1177/13607804251403342 Schwabish, J., and Feng, A. (2021). Do No Harm Guide: Applying Equity Awareness in Data Visualization. Urban Institute.

Generative artificial intelligence has moved from the research laboratory into the day-to-day work of social service agencies faster than the profession has been able to build rules for it. Practitioners now use large language models to summarise home visits, draft assessments, structure case notes, translate documents, and prepare reports for courts and funding panels. The appeal is easy to understand: caseloads are heavy, paperwork is heavy, and any tool that saves an hour of typing looks like a gift. But the same tools that save time can also invent facts, leak confidential information, quietly reproduce discrimination, flatten the voice of the #service_users, and blur the question of who is responsible when something goes wrong. This article examines the specific risks that arise when #generative_AI is used for #client_assessment and reporting in #social_work, and proposes a layered system of #ethical_guardrails that agencies and practitioners can adopt. Drawing on an integrative review of peer-reviewed literature published between 2021 and 2026, together with a normative analysis grounded in professional codes of ethics and emerging regulation, the article identifies ten risk domains: fabrication and inaccuracy, confidentiality and data protection, informed consent, bias and inequity, opacity and explainability, accountability and liability, automation bias and deskilling, harm to the working relationship, evidentiary and legal exposure, and the datafication of clients. It then sets out a four-tier guardrail framework operating at the regulatory, organisational, practitioner, and client levels, and offers a task-risk matrix that distinguishes low-risk, conditional, and prohibited uses of generative tools. The article argues that the profession should not choose between blanket prohibition and uncritical adoption. It should instead insist on a principle that is simple to state and demanding to apply: a machine may help a social worker write, but it must never be allowed to decide what is true about a human life. Keywords: generative artificial intelligence; social work ethics; client assessment; algorithmic accountability; professional judgement; data protection; child welfare; human oversight 1. Introduction Social work has always been a profession of writing. The assessment, the case note, the court report, the referral letter, the risk analysis, the review summary: these documents are not administrative residue. They are the instruments through which a client becomes visible to systems that hold power over them. A sentence in a #case_notes file can decide whether a child stays at home, whether a person receives housing, whether an older adult keeps control of their finances, whether an asylum seeker is believed. The written record is where #professional_judgement becomes action. This is why the arrival of #generative_AI in social work is not a routine software upgrade. When a practitioner asks a large language model to turn a recorded conversation into an assessment, they are handing part of that writing task to a system that has no understanding of the family, no duty to the client, no professional registration, and no capacity to be held to account. The system produces text that reads like a competent professional wrote it. Whether that text is true is a separate question, and one that the system itself cannot answer. Uptake has been rapid. Agencies in the United Kingdom, the United States, Australia, and across Europe have piloted or adopted tools that record practitioner-client conversations, transcribe them, and generate structured assessments and action points that can be pasted directly into case management systems. Individual practitioners, often without agency approval, paste anonymised (or, worryingly, not-so-anonymised) details into publicly available chatbots and ask for a draft. Surveys of social work educators report that the majority already use these systems in some part of their work. One study of social work educators using ChatGPT found that participants judged the tool useful in a large majority of their interactions with it, most often for teaching support. The professional literature has begun to respond. Reamer's foundational analysis noted that although artificial intelligence was already being used to conduct risk assessments, assist people in crisis, strengthen prevention, and predict service outcomes, social work's literature did not yet include an in-depth examination of the ethical implications of practitioners' use of these systems. That gap is narrowing, but slowly, and the pace of technological change continues to outrun the pace of ethical deliberation. This article addresses one slice of the problem in depth: the use of generative systems for #client_assessment and reporting. It deliberately excludes chatbots offering direct therapeutic support to clients, predictive scoring tools that produce a number rather than a narrative, and the use of AI in social work research, except where these overlap with the assessment and reporting question. The narrowing is intentional. Assessment and reporting is where the largest number of practitioners are using these tools right now, where the harms are most immediate, and where the guardrails are weakest. The article has four aims. First, to describe accurately what generative systems do and do not do, because much of the confused debate in the profession rests on a poor grasp of the underlying mechanism. Second, to map the risk domains systematically rather than anecdotally. Third, to propose a concrete, implementable framework of #ethical_guardrails rather than a list of pious principles. Fourth, to identify what the profession still does not know and needs to research. The underlying commitment is one of #accountability. Social work exists because societies decided that #vulnerable_people should not be left to the mercy of systems they cannot see or challenge. It would be a strange irony if the profession built to interrupt that dynamic became one of its newest vectors. 2. Background: What Generative Systems Actually Do 2.1 The mechanism A large language model is a statistical system trained to predict the next unit of text given the text that came before. It is trained on enormous quantities of written material and learns the patterns of how words follow other words. When a social worker types a #prompt asking for an assessment summary, the model does not consult a knowledge base, check a file, or reason about the family. It generates the sequence of words that is statistically most plausible given the input and its #training_data. This is worth stating plainly because it explains almost everything that follows. #probabilistic_text generation is very good at producing writing that has the shape, tone, and vocabulary of expert work. It is not designed to produce writing that is true. Truth and plausibility usually overlap, which is why the systems seem so impressive. When they diverge, the system has no mechanism for noticing. The distinction matters for social work more than for many other fields, because so much of what a social worker records is contested. Two family members give different accounts. A young person says one thing to the teacher and another to the practitioner. A parent's explanation is partly true. Assessment is the disciplined work of holding those tensions in view. A system optimised for plausible, coherent, fluent prose has a natural bias toward resolving tension rather than recording it. 2.2 How generative tools differ from older decision support Social work is not new to computerised assistance. #decision_support_systems, structured decision-making tools, and actuarial risk instruments have been in use for decades. More recently, #predictive_risk_modelling has been deployed in child protection screening, homelessness prioritisation, and benefits administration. The Allegheny Family Screening Tool, for example, is a machine learning based risk prediction tool designed to assist a social worker in deciding whether to screen a referral in for further investigation or screen it out, drawing on data about the child, the parents or guardians, prior child welfare history, criminal history, and use of public assistance. These older tools have their own well-documented problems, which the profession has argued about for years. But they differ from generative systems in four ways that matter for ethics. First, output type. A predictive tool produces a score or a category. A generative tool produces prose, and prose carries authority. A risk score of 14 invites the question "what does 14 mean?" A fluent paragraph describing a mother's "limited insight into the impact of her substance use on her children" invites no question at all. It reads like a professional observation, and it will be treated as one. Second, auditability. A structured decision-making tool has a documented set of items and weightings. Even when the model is complex, the inputs can be listed. With a generative system the mapping from input to output is a #black_box in a deeper sense: the same prompt run twice may produce different text, and there is no stable, inspectable rule that produced any given sentence. Third, scope. Actuarial tools were bounded to specific decisions at specific points. Generative tools are general purpose. The same system that drafts a case note will draft a court report, a complaint response, a supervision record, and an email to a client. Governance designed around a single decision point does not fit a tool that touches everything. Fourth, invisibility. Predictive tools were procured, configured, and rolled out by organisations. Generative tools are frequently adopted by individual practitioners on their own devices, outside any procurement or approval process. Much of the current use in the profession is invisible to the organisations that carry the legal liability for it. 2.3 The main use cases in assessment and reporting Five patterns of use are now common enough to be described. Ambient recording and note generation. The practitioner records a home visit or an interview. A #speech_to_text system transcribes it. A language model then converts the transcript into a structured note or assessment, often with suggested action points. This is the fastest-growing category and the one with the highest stakes, because it inserts machine processing between the client's spoken words and the official record of what they said. Drafting from practitioner notes. The practitioner writes rough notes and asks the model to expand them into a formal assessment in the agency's house style. This is lower risk than ambient recording, because a human being has already performed the act of selection and interpretation, but it is not risk-free. Summarising case files. The practitioner uploads or pastes a long history and asks for a chronology or summary. This is attractive when a worker inherits a case with hundreds of pages of history. It is also where fabrication is hardest to detect, because the whole point of the exercise is that the worker has not read the source. Translation and plain-language rewriting. Converting a report into a client's first language, or into accessible language for a person with a learning disability. Genuinely valuable, and genuinely risky, since a mistranslation in a child protection letter is not a typo. Report structuring for panels and courts. Reorganising existing material into the format demanded by a court, a funding panel, or a review board. Each of these use cases carries a different risk profile, and one of the central arguments of this article is that governance must be built at the level of the task, not at the level of the technology. A blanket policy that says "AI is permitted" or "AI is banned" is useless. What is needed is a policy that says which tasks, with which safeguards, with what human verification, and with what disclosure. 3. Review of the Literature 3.1 The optimistic strand A substantial part of the social work literature on artificial intelligence is cautiously positive, and it would be dishonest to present the field as uniformly alarmed. The argument is straightforward: administrative burden is one of the largest drivers of burnout and one of the largest thieves of time that could be spent with people. Generative tools can absorb some of that burden. Reported uses include generating draft case notes, producing psychoeducational materials, developing treatment plan goals, drafting funding applications, and preparing community assessments across micro, mezzo, and macro practice. Singer, Baez, and Rios argued in an influential editorial that it was time for social work educators to play a greater role in teaching, learning, developing, and researching AI technologies rather than standing outside the conversation. Their position, echoed since by many others, is that abstention is itself a choice with consequences, and not a neutral one. If social workers do not shape how these tools are built and used in human services, they will be built and used by people who have never sat with a family in crisis. Boetto's EPIC model, developed from a comprehensive review of the literature, proposes four components for integrating artificial intelligence into the profession: ethics and justice, policy development and advocacy, intersectoral collaboration, and community engagement and empowerment. The argument is that securing the benefits of AI requires a proactive and ethical approach rather than passive drift. The emphasis on advocacy is important. Social workers are not only users of these systems; they are also witnesses to their effects on the people they serve, and are therefore positioned to give evidence that no one else can give. 3.2 The AI literacy strand A second strand focuses on capability. Ahn and colleagues argue that AI literacy, understood as the ability to critically evaluate, collaborate with, and use AI technologies responsibly, should be embedded across the core competencies of social work rather than added as an optional extra. #AI_literacy in this framing is not a technical skill. It is a form of critical practice: understanding what a system can and cannot do, recognising where its limits fall, and being able to explain both to a client. This literature is significant because it reframes the problem. The question is not only "may a social worker use this tool?" but "does this social worker understand the tool well enough to use it safely?" A practitioner who believes that a language model retrieves facts from a database will use it very differently from one who understands that it generates plausible text. The first practitioner is dangerous. The second may be safe. Most qualifying programmes currently produce the first. Hodgson and colleagues had already argued, before the generative wave broke, that artificial intelligence in social work education needed to be problematised rather than absorbed, and that the profession should treat these systems as objects of critical inquiry. That call now looks prescient. 3.3 The critical and justice strand The third strand is the most sceptical, and draws heavily on work outside social work in the fairness and accountability research community. Its central claim is that these systems do not merely make errors; they make patterned errors, and the pattern follows existing lines of social disadvantage. The evidence base here is strong and growing. A widely cited study of a leading speech recognition system found that roughly one percent of audio transcriptions contained entire hallucinated phrases or sentences that did not exist in any form in the underlying audio, and that around 38 percent of these hallucinations included explicit harms such as perpetuating violence or inventing inaccurate associations. Crucially, the study examined why hallucinations occur by comparing rates between speakers with aphasia, who have a reduced ability to express themselves through speech, and a control group. The finding is directly relevant to social work: the people most likely to be mis-transcribed are those whose speech is disfluent, hesitant, accented, distressed, or affected by disability, which is to say a large proportion of the people social workers see. A system whose #fabrication rate rises for the people with the least power to correct the record is not a neutral tool. In the child welfare domain, critical scrutiny of algorithmic screening has been detailed. Gerchick and colleagues conducted a close interrogation of the values embedded in the Allegheny Family Screening Tool, examining what the model's design choices assume and whose interests those assumptions serve. A scoping review of predictive analytics in children's social care found that most models lack transparency, provide limited information on their development or error rates, and rarely examine #algorithmic_bias or the harms of false positives. These findings concern predictive rather than generative systems, but the lesson transfers: tools deployed in #child_welfare have consistently been introduced with far less evaluation than their consequences warrant. Ratner and Elmholdt's analysis of how predictive algorithms construct risk by calculating and anticipating children's uncertain futures shows how a technology can quietly change the meaning of professional categories, and how a public sector can drift toward treating citizens as data points suited for calculation and prediction rather than as individuals with agency and rights. 3.4 The uncertainty strand A fourth strand, less often connected to the AI debate but highly relevant, concerns the nature of risk assessment itself. Harnett argues that estimating risk probabilistically in the child protection context is unreliable and prone to error under conditions of high uncertainty, and that this holds true whether risk is estimated using professional judgement, statistical algorithms, or a combination of both. This is a sobering corrective to techno-optimism and techno-pessimism alike. It cautions against the assumption that human judgement is the reliable baseline from which machines deviate. The relevant question is not "is the machine as good as the human?" but "does the introduction of the machine make the whole system of judgement better or worse, and for whom?" 3.5 What is missing Three gaps are apparent. First, almost all empirical work on AI in social work concerns predictive scoring, not generative text. The evidence about generative tools in assessment and reporting is largely anecdotal, drawn from pilots that have not been independently evaluated. Second, there is very little research capturing the perspective of #service_users. We know almost nothing about what clients think about their words being processed by a machine, and whether they would consent if genuinely asked. Third, there is no established methodology for auditing the accuracy of an AI-drafted assessment, because there is often no ground truth to audit against. 4. Method This is a conceptual article built on an integrative #narrative_review combined with normative analysis. The purpose is not to estimate an effect size but to organise a scattered field and derive practical guidance from it. Search strategy. Peer-reviewed literature was identified through database and citation searching using combinations of terms covering artificial intelligence, generative AI, large language models, social work, child protection, case recording, assessment, and ethics. Reference lists of key articles were followed forwards and backwards. Inclusion criteria. Priority was given to peer-reviewed sources published from 2021 onwards, in English, addressing either the use of AI in social work and adjacent human services, or the technical behaviour of generative systems in ways directly relevant to professional documentation. Sources concerned solely with AI in general education, marketing, or commerce were excluded. A small number of foundational sources outside that window are referenced where they have not been superseded. Normative analysis. Identified risks were then evaluated against the ethical standards of the profession, chiefly the duties of informed consent, confidentiality, competence, integrity of the record, self-determination, and social justice as expressed in professional codes, and against the emerging regulatory instruments described in section 8. The guardrail framework in section 6 was derived by asking, for each risk domain, what control would need to be in place for a reasonable practitioner to say honestly that the duty had been discharged. Limitations of the method. A narrative review is vulnerable to selection bias, and the field is moving quickly enough that some claims will date. No formal quality appraisal instrument was applied. The framework proposed here is theoretically derived and has not yet been empirically tested, which is stated below as a research priority rather than concealed as a limitation. 5. Ten Risk Domains 5.1 Fabrication and inaccuracy The first and most fundamental risk is that the system will assert things that are not true, and will assert them confidently. The technical term is #hallucination, which is unfortunate, because it suggests a rare glitch rather than an intrinsic property. Generating plausible text is what the system is for. When the plausible text happens to correspond to reality, we call it accurate; when it does not, we call it a hallucination. The mechanism is identical in both cases. There is no internal signal that distinguishes them. In social work documentation this shows up in specific ways. Invented detail. A model summarising a transcript may add a plausible-sounding but unstated fact: a diagnosis the client never mentioned, a relative who does not exist, a date that was never given. These additions are hard to spot precisely because they fit. Smoothing. A model asked to summarise a chaotic, contradictory, emotionally messy conversation will produce something coherent. Coherence is achieved by discarding contradiction. But in assessment, the contradiction often is the finding. Attribution drift. The transcript records the practitioner asking "is he hitting you?" and the client saying nothing. The summary records "client did not deny physical violence." One sentence has migrated from silence to implication. Confabulated confidence. Models rarely write "it is unclear whether." They write declaratively. A record built from declarative sentences about a life that is genuinely unclear is a false record. Transcription-level error. Before the model even begins, the speech-to-text layer may have introduced errors, and the evidence shows those errors are unevenly distributed, falling harder on disabled, distressed, and accented speakers. The compounding effect deserves emphasis. Errors introduced at the transcription stage are inherited by the summarisation stage, which then presents them in polished professional prose. Each stage launders the previous one. By the time the text reaches the case file, the error looks like a considered clinical observation. 5.2 Confidentiality and data protection The second risk is the one most often raised and least often solved. When a practitioner types client information into a general-purpose chatbot, that information leaves the agency. Depending on the service, it may be stored, logged, reviewed by human contractors, or used to improve the model. There is a substantial literature on privacy leakage from language systems, including the unauthorised extraction of training data and the inference of personal attributes from apparently harmless input. #third_party_processing of a child protection record through a consumer service is, in most jurisdictions, a breach of #data_protection law and of the practitioner's professional duty, whether or not any harm ultimately occurs. Three specific traps deserve attention. The de-identification illusion. Practitioners often believe that removing names makes the data safe. It does not. A narrative describing a fourteen-year-old girl in a named small town, with a named school, a specific disability, and a father recently released from prison, is identifiable to anyone who knows the area, regardless of whether the name is present. Genuine #de_identification of rich case narrative is technically difficult, and the #re_identification risk in small communities is severe. The paste-in habit. The most common breach is not a sophisticated attack. It is a tired practitioner at eight in the evening pasting three paragraphs of a case record into a free chatbot to get a draft. No policy on its own prevents this. Only culture, training, and the provision of a safe alternative do. The vendor promise. Enterprise tools frequently promise that data is not retained or used for training. These promises may be sincere and may be true. They are also contractual claims that most agencies lack the technical capacity to verify, and that can change with an update to the terms of service. 5.3 Informed consent and client autonomy A client who sits in their kitchen and speaks to a social worker has consented to speak to a social worker. They have not necessarily consented to have their voice recorded, transcribed by a commercial system, processed by a language model, and rendered into a document by a machine. Reamer highlighted the importance of obtaining appropriate informed consent from service users so that the benefits, limitations, and risks associated with AI are clarified, and stressed that protecting privacy in relation to third-party access to service user data, and informing service users about risks and breaches, are essential for transparent practice. Responsible practice on this view includes clients being fully informed about how AI is being used in their care and having a genuine #right_to_opt_out of AI-assisted processes. The difficulty is that consent in social work is frequently not free. A parent in child protection proceedings is not in a position to refuse a request from the practitioner who will write the report that goes to court. A person applying for scarce housing support is not going to risk being seen as difficult. Consent obtained under those conditions is a formality, not a safeguard, and treating it as a safeguard is a form of self-deception. This has a direct implication for #client_autonomy. Where the power imbalance is steep, consent should not be the primary control. The primary control should be a limit on what the agency does at all. Some things should not be done, even with a signature. 5.4 Bias, discrimination, and inequity Language models learn from text produced by human societies, and human societies are unequal. The systems absorb, compress, and reproduce those inequalities, sometimes obviously and more often subtly. In assessment writing, several mechanisms operate. Register and description. Models trained largely on published text carry patterns about how different groups are described. A drafting system may reach more readily for the language of deficit when the input concerns a poor family and the language of stress when the input concerns an affluent one, mirroring the patterns in its training material. Differential error rates. As noted, transcription and comprehension errors are not evenly distributed. Non-standard speech, second-language speakers, people with speech impairments, and distressed speakers are all more likely to be misrepresented. Proxy variables. Even when protected characteristics are excluded, correlated features carry the same information. Postcode, benefit history, and prior service contact function as #proxy_variables for race and class. Feedback loops. If AI-generated assessments influence decisions, and those decisions generate data, and that data trains future systems, initial bias is amplified rather than corrected. Families already known to services generate more data, appear riskier, and are watched more closely, which generates more data. These #feedback_loops are the engine of algorithmic injustice. The justice concern here is not merely that a model may be rude about a group. It is that the technology may make discriminatory judgement more efficient, more consistent, and much harder to challenge, because a biased sentence written by a machine looks exactly like a considered sentence written by a professional. 5.5 Opacity, transparency, and explainability A client has a right to know how a judgement about them was reached. A tribunal has a right to test the basis of an assertion. Both rights depend on someone being able to explain where a statement came from. With a generative system, this is often impossible in any deep sense. The practitioner can say "I asked the tool to summarise the transcript and it produced this." They cannot say why the tool chose the word "resistant" rather than "frightened," or why it foregrounded one incident and omitted another. #explainability at the level required for meaningful #contestability does not currently exist for these systems. The practical response is not to demand explanations the technology cannot give. It is to insist that no AI-generated statement enters the record unless a named human being has independently verified it against the source and is prepared to defend it as their own professional statement. This converts an unexplainable machine output into an explainable human one. It also, correctly, makes the tool less of a time-saver, because verification takes time. That is not a flaw in the guardrail. That is the guardrail. 5.6 Accountability and liability If a fabricated sentence in an AI-drafted assessment contributes to a child being wrongly removed, or wrongly left at home, who is responsible? The answer, under every current professional and legal framework, is the social worker and the employing agency. The vendor's terms of service will disclaim responsibility for outputs. The model has no legal personality. The practitioner signed the report. This is the right answer, and practitioners should understand its full weight. Delegating the drafting of a document does not delegate #liability for its contents. A worker who signs an assessment containing a fact they did not verify has made a professional assertion they cannot support, and that is a #malpractice exposure independent of any technology. There is a further complication in the emerging pattern of blame. When an AI-assisted decision goes wrong, organisations face a temptation to locate fault in the individual worker who "failed to check," while continuing to promote the tool that made checking feel unnecessary and continuing to set caseloads that make checking impossible. An honest governance framework must place duties on the organisation as well as the individual, and must treat unrealistic workload as a cause of AI-related error rather than an excuse for it. 5.7 Automation bias and deskilling #automation_bias is the well-documented human tendency to over-trust a machine's output, to under-scrutinise it when it agrees with an initial impression, and to be swayed by it when it does not. It is not a failure of intelligence. It is a predictable feature of how people work with confident-seeming systems under time pressure. In assessment work this bias has a specific shape. Editing a draft is cognitively different from writing one. When a practitioner writes an assessment from scratch, they are forced to decide what matters, what is evidenced, and what remains unknown. When they edit a fluent draft, they are anchored to a structure and a set of claims that something else has already made. The path of least resistance is to accept the frame and correct the surface. Over time, the concern extends to #deskilling. Assessment is not merely reporting; it is a mode of thinking. The discipline of writing "I do not know" is part of how a practitioner learns to notice that they do not know. If the machine never writes it, and the practitioner never has to, the capacity to tolerate uncertainty may weaken. The literature on #professional_curiosity in child protection makes clear how much depends on that capacity, and how easily it is lost. The risk is most acute for students and newly qualified practitioners, who may never develop the capability in the first place. A senior worker using a drafting tool is delegating a skill they possess. A student doing the same is skipping the apprenticeship. 5.8 Harm to the working relationship Social work is done through relationship, and relationship depends on attention. A practitioner who knows that the conversation is being recorded and will be auto-summarised may, in theory, be freed to listen. In practice, several effects run the other way. Clients who know they are being recorded change what they say, and the people most likely to withhold are those with the most reason to fear the record: people with insecure immigration status, people with prior involvement in the justice system, people who have been let down by services before. The chilling effect falls exactly where trust is most fragile. There is also the question of what the client experiences. Being told that a machine will write the account of your life is not a neutral communication. It carries a message about how much your account matters and how much of a person is required to receive it. The #therapeutic_relationship depends partly on the client's sense that a human being is holding their story in mind. A workflow that makes the story machine-readable can make the client feel processed rather than heard. #empathy is not a formatting feature. None of this means that recording and drafting tools are always harmful to relationship. It means the effect on relationship is an outcome that must be measured, not assumed. 5.9 The record as legal evidence Social work documents are legal documents. They are disclosed in care proceedings, quoted in tribunals, examined in serious case reviews, and used in litigation. An assessment is, in effect, evidence, and its author may be cross-examined on it. This creates exposure that many practitioners have not yet thought through. Provenance. If a report contains an assertion that the author cannot trace to a source, the report is vulnerable, and so is the author's credibility on everything else in it. Consistency. Generative systems are non-deterministic. Two summaries of the same transcript may differ. If both exist in the record, the inconsistency is disclosable. Disclosure of method. A court may reasonably ask whether AI was used, how, and with what verification. An agency that has not recorded the answer cannot give it. The transcript itself. If a conversation was recorded, the recording and transcript may be disclosable material. Agencies that record home visits without a retention and disclosure policy are creating evidence they do not control. The prudent position is that any use of generative tools in a document that may become #court_evidence should be disclosed, logged, and independently verified, and that the underlying source material should be retained and retrievable. The integrity of the #legal_record is not a technical detail. It is the difference between a professional opinion and an unsourced allegation. 5.10 Datafication and surveillance The final risk is structural rather than individual. Every technology that makes recording easier tends, over time, to increase the amount of recording. If assessments become cheap to produce, more will be produced. If conversations are routinely recorded, the recording becomes the norm and refusal becomes suspicious. If summaries are machine-generated, they become machine-searchable, and searchable records invite aggregation, linkage, and analysis that were never contemplated when the client spoke. This is the slow slide toward #surveillance, and it does not require anyone to intend it. It requires only that each individual step be locally convenient. The people it affects are, as always, those with the least ability to opt out of the systems that record them. Social work has a professional obligation to notice this, because the profession sits at the point where the state meets the private life of the citizen. Practitioners are not merely users of these systems. They are, whether they wish it or not, among its instruments. 6. A Four-Tier Guardrail Framework Principles are cheap. Every organisation now has an AI policy that says the words "transparency," "fairness," and "human oversight," and almost none of them tell a practitioner at half past four on a Friday what they are actually allowed to do. This section proposes something more concrete. The framework has four tiers, and the argument is that a control at one tier cannot substitute for a control at another. Client consent does not fix a badly procured tool. A good tool does not fix an untrained practitioner. A well-trained practitioner does not fix an absent regulator. Tier 1: Regulatory guardrails These are the constraints that come from outside the agency, and that the agency cannot waive. Classification of risk. Any system used in the assessment of a person's needs, risks, or eligibility for a public service should be treated as high risk by default, and should attract the obligations that come with that classification: conformity assessment, technical documentation, logging, and mandated human oversight. Legal basis for processing. Client data may only be processed by a system that has a lawful basis, a data processing agreement, and a completed data protection #impact_assessment. This is not a formality to be completed after adoption. It is a precondition of adoption. Professional standards. Registration bodies should state explicitly, rather than by implication, that a registrant who signs a document containing unverified machine-generated assertions has committed a fitness-to-practise breach, in the same way that a registrant who signs a report they did not write has always done. Prohibitions. Certain uses should be off the table entirely, regardless of consent or safeguards. These are set out in the task matrix below. Tier 2: Organisational guardrails Approved tools only. The agency should maintain a short, explicit list of approved systems, and use of any other system with client data should be a disciplinary matter. This only works if the approved list actually contains something usable. A policy that bans everything and provides nothing produces shadow use, not safety. #procurement discipline. Before adoption, the agency should require from the vendor a description of the training data, known error rates, evidence of testing on populations resembling the agency's clients, data retention terms, subprocessor lists, and a #model_card or equivalent. If the vendor will not provide these, the answer is no. #data_minimisation by design. The system should receive the minimum data required for the task. Ambient recording of an entire home visit in order to generate a two-paragraph note is not minimisation; it is maximal collection justified by minimal benefit. Mandatory #audit_trail. Every AI-assisted document should carry metadata recording which tool was used, which version, what prompt or configuration, who verified it, and when. This is not bureaucratic decoration. It is the only thing that makes retrospective investigation possible. #version_control and retention. The original source material, the raw transcript, the machine draft, and the final human-verified document should all be retained and distinguishable. An agency that keeps only the final document has destroyed its own ability to investigate an error. #incident_reporting. A specific, low-friction route for practitioners to report AI errors, near-misses, and inappropriate outputs, with an explicit no-blame framing for honest reporting. Without this, the organisation learns nothing until a case collapses in court. Caseload realism. If verification is mandatory, time for verification must be in the workload model. An organisation that adopts a drafting tool and simultaneously raises caseloads on the strength of the expected time saving has converted a safety control into a fiction. Tier 3: Practitioner guardrails The verification rule. No sentence enters the record unless the practitioner has checked it against the source and is willing to defend it as their own. This is the single most important rule in the framework and everything else supports it. The authorship rule. The practitioner is the author. The tool is an instrument. If the practitioner cannot explain a sentence, the sentence comes out. The uncertainty rule. The practitioner is responsible for putting back the uncertainty the model removed. Every AI-assisted assessment should be read specifically for false confidence, and hedges should be reinstated wherever the source does not support a firm claim. The voice rule. Where the client's own words matter, the client's own words go in. A machine paraphrase of a disclosure is not a disclosure. The disclosure rule. A #disclosure_statement should appear in any document where generative tools contributed materially to the drafting, naming the tool and the nature of its contribution. The competence rule. A practitioner who does not understand what the tool does should not use it. Competence includes knowing that the model does not retrieve facts, that it can fabricate, that it is non-deterministic, and that its errors are not randomly distributed. The supervision rule. Use of AI in assessment should be a standing item in professional #supervision, not an occasional topic. Supervisors should ask to see the source and the draft, not just the polished output. Tier 4: Client guardrails Plain-language notification. Clients should be told, in language they can understand, in their own language, what is being used, what it does, what could go wrong, and what happens if they say no. A real right of refusal. The refusal must be costless. If declining AI processing means a slower service, a worse-tempered worker, or a note in the file describing the client as uncooperative, the right does not exist. A right to see and correct. Clients should be able to read what has been written about them and challenge it. This right exists in most jurisdictions already; AI makes exercising it more urgent, because the volume of text is larger and its authorship is murkier. #co_production of policy. Agencies should involve people with lived experience of services in deciding how these tools are used. The people whose records these are have a legitimate claim to a say in how they are made, and their questions are usually sharper than the ones the professionals ask. 6.1 The task-risk matrix The framework is operationalised through #risk_tiering of tasks rather than of technologies. Green: generally permissible with standard verification. Rewriting a practitioner's own completed draft for grammar, clarity, or reading level. Producing a plain-language version of an already-approved document for a client, subject to human check by a qualified translator where language is changed. Generating template structures, headings, or question prompts that contain no client data. Drafting non-client-specific material: training content, policy summaries, funding narratives. Assisting with the practitioner's own reflective learning, using fictional or fully synthetic scenarios. Amber: permissible only with specific safeguards. Summarising a case file that the practitioner will also read, with the summary treated as an index rather than a substitute. Expanding practitioner-written notes into a formal assessment, where the practitioner has already made every substantive judgement and the model only assists with form. Transcription of a recorded session, with the transcript treated as a draft requiring human correction and never as the record itself. Amber tasks require: an approved enterprise tool with contractual data protections; documented client notification; a completed verification against source; an audit trail; and supervision oversight. Red: #prohibited_uses. Generating a risk judgement, a recommendation, or a conclusion. The model may help you write what you have decided. It may not decide. Any use that inserts machine processing between a disclosure of harm and the record of that disclosure. Processing identifiable client data through any non-approved or consumer tool, under any circumstances. Generating text for a court report, expert statement, or statutory assessment conclusion without full source verification and explicit disclosure. Using AI output to fill a gap in the practitioner's own knowledge of the case. If you did not know it, the model does not know it either. Covert use. Undisclosed reliance on a machine to author a professional document is a form of misrepresentation. The matrix is intentionally strict at the red line and generous in the green zone. The profession's instinct to ban everything is as unhelpful as the sector's instinct to buy everything. What is needed is a clear boundary in the right place. 7. Implementation: Making Guardrails Real A framework that exists only in a policy document is a liability, because it creates the appearance of governance without the substance. Four implementation questions decide whether guardrails actually hold. 7.1 Testing before deployment Agencies should not adopt these tools on the basis of a vendor demonstration. They should test them, and they should test them adversarially. #red_teaming. Deliberately attempt to make the system fail in the ways that would matter most. Feed it a transcript containing a hesitant disclosure and see whether the summary strengthens or weakens it. Feed it an accented or disfluent speaker and check the transcription. Feed it two versions of the same case that differ only in the family's name, address, or ethnicity, and compare the language of the outputs. Feed it a transcript containing a contradiction and see whether the contradiction survives. #sandbox_testing. Test with synthetic cases that resemble the agency's real caseload before any real case is processed. Ratner and Elmholdt's work on sandbox approaches to predictive modelling in child and family welfare shows how simulated environments can surface the assumptions embedded in a tool before those assumptions do damage. Publish the results internally. Staff should be told what the tool got wrong in testing. A practitioner who has seen the system fabricate a detail in a test case will read its outputs very differently from one who has only seen the marketing. 7.2 Training that changes behaviour #staff_training in this area is frequently a slide deck about being careful. That is not training. Effective training should include: a mechanical explanation of what the model does, delivered without mystification; hands-on exposure to the tool failing, so that fallibility is experienced rather than described; explicit practice in verification, including timed exercises where staff must check a draft against a source and find planted errors; and structured discussion of the ethical duties engaged, using the agency's own case types. Training should be repeated, because models change. A tool that behaved one way last year may behave differently after an update, and staff confidence built on the old behaviour is a hazard. 7.3 Continuous monitoring #quality_assurance sampling. A proportion of AI-assisted documents should be routinely audited against their source material, and the error rate should be reported to the board. If nobody is measuring the error rate, nobody knows whether the guardrails work. Equity monitoring. Error rates should be disaggregated. If fabrications and mis-transcriptions are concentrated among particular groups, the tool is producing discriminatory outcomes and must be withdrawn or fixed, regardless of its average performance. Outcome monitoring. Does the tool actually save time? Does the time saved go to clients or to more cases? Do clients report feeling more or less heard? These are answerable questions and they are rarely asked. 7.4 Speaking up Practitioners need a protected route to raise concerns, including concerns about the tool itself and about pressure to use it in ways that feel wrong. #whistleblowing protection is a technical safeguard as much as a moral one, because in most technology failures the people who knew first were junior and had no safe way to say so. The professional obligation runs both ways. A social worker who sees a tool producing unjust outcomes has a duty to report it, and an organisation that punishes that report has forfeited any claim to ethical governance. 8. The Regulatory Landscape 8.1 Statutory regulation The most developed regulatory instrument at present is the European Union's artificial intelligence regulation, which takes a risk-tiered approach: some practices are prohibited outright, a defined set of #high_risk_systems attract substantial obligations, and lower-risk systems attract transparency duties. Systems used by public authorities to evaluate eligibility for essential public services and benefits fall within the high-risk category, which places much of the social work technology stack squarely inside the regime. The obligations include risk management, data governance, technical documentation, record-keeping, transparency to users, and effective #human_oversight. The #EU_AI_Act matters beyond Europe because vendors build one product, and a compliance standard in a large market tends to become the global default. In the United States, there is no equivalent federal statute, and the most influential instrument is a voluntary framework. The #NIST_AI_RMF organises AI risk management around four functions: govern, map, measure, and manage. Being voluntary, it has no teeth, but it provides a common vocabulary and a defensible standard of care, which matters in negligence litigation. Health-sector guidance from the World Health Organization on large multi-modal models is directly relevant to social workers in health settings, and its emphasis on transparency, human oversight, and independent evaluation aligns with the framework proposed here. 8.2 Professional regulation Professional codes were mostly written before generative AI existed, but they are not silent. The duties of competence, informed consent, confidentiality, integrity, and avoidance of misrepresentation apply straightforwardly. The relevant provisions of the #Code_of_Ethics, in most national versions, already prohibit a practitioner from claiming as their own work that they did not do, from disclosing confidential information without authorisation, from practising beyond their competence, and from misrepresenting their qualifications or the basis of their opinions. Every one of these is engaged by the careless use of generative tools. The profession does not need entirely new ethics. It needs to apply the ethics it has, with sufficient specificity that a practitioner knows what compliance looks like on a Tuesday afternoon. That said, two areas do require new drafting. First, an explicit standard on disclosure of AI assistance in professional documents. Second, an explicit standard on the retention of source material where AI has been used, so that verification is auditable after the fact. 8.3 The gap There is a structural mismatch between the pace of regulation and the pace of adoption. Statutes take years. Model updates take weeks. This gap will not close, and any governance model that depends on regulators keeping up will fail. The implication is that #data_governance must be built at the organisational level, in a form that does not require a new rule for every new product. The task-risk matrix above is an attempt at exactly this: it regulates what the practitioner does, not what the vendor sells, and it therefore survives the next model release. #regulatory_compliance is a floor, not a ceiling, and the profession should say so. 9. Implications for Education and Research 9.1 Education If AI literacy is a competence, it must be taught, assessed, and examined. Teach the mechanism. Students should be able to explain, in plain language, why a language model can produce a confident falsehood. This is not a computer science lecture. It is a five-minute explanation that changes how a person uses a tool for the rest of their career. Teach verification as a skill. Give students a transcript and a machine-generated summary with planted errors, and mark them on whether they find them. This is directly transferable to practice and it is not currently taught anywhere. Protect the apprenticeship. #social_work_education should require students to write assessments unaided before they are permitted to use drafting tools. The purpose is not purity. It is that you cannot supervise a machine at a task you have never performed. Teach the politics. #curriculum should locate these tools in the history of technology and social control, so that students understand that the question "does it work?" is inseparable from the question "for whom, and against whom?" Model honest use. Educators who use these tools should say so, and should demonstrate verification openly. A programme that bans student use while quietly relying on the same tools for marking has taught a lesson, but not the one it intended. The deeper educational aim is to protect #critical_thinking. The most dangerous outcome of the generative era for social work would not be that practitioners use these tools. It would be that they stop noticing when the tools are wrong. 9.2 Research The evidence base needed to govern these tools well does not yet exist. #future_research should prioritise the following. Accuracy studies. Rigorous, independent comparison of AI-assisted and human-written assessments against an agreed source, measuring fabrication, omission, and distortion rates. This is difficult but not impossible, and no responsible deployment should proceed without it. Equity studies. Disaggregated error analysis by race, language, disability, age, and class. If the tool fails differently for different people, that must be known before it is used, not after. Service user studies. What do clients actually think and feel about machine processing of their words? Do they understand what they are consenting to? What would they choose if the choice were real? This is the largest and most shameful gap in the current literature. Relationship studies. Does ambient recording change what clients disclose? Does the promise of time saved actually translate into time with people? Decision studies. Does exposure to an AI draft change the judgement a practitioner would otherwise have made, and in what direction? #practitioner_research. Frontline workers are well placed to document what these systems do in the real conditions of the job, and their observations should be systematically collected rather than treated as anecdote. 10. Limitations Several limitations should be acknowledged. The article is conceptual, and its central framework has not been empirically tested. It may prove unworkable in some settings, particularly small agencies without governance capacity, and it may prove insufficient in others. The literature base is thin in precisely the area of greatest concern. Most of the empirical evidence cited concerns predictive systems or general-purpose language models, not generative tools used specifically for social work assessment. The transfer of those findings is reasoned, not demonstrated. The analysis is written largely from an Anglophone, high-income perspective, drawing on regulatory instruments and professional codes from Europe, North America, and Australasia. The risks and the appropriate guardrails may look different in contexts with different legal traditions, different resource constraints, and different histories of state intervention in family life. Finally, the technology is moving. Claims about what these systems can and cannot do have a short shelf life. The risk domains identified here are likely to prove durable because they follow from the structure of the technology and the structure of the profession. The specific mitigations may date quickly. 11. Conclusion The question facing social work is not whether to use generative artificial intelligence. That question has been answered in practice, by thousands of practitioners, mostly without permission and mostly without training. The question is whether the profession will govern that use or be governed by it. The case for use is real. Paperwork is crushing the profession, and a tool that returns time to practitioners could return that time to clients. Rejecting these systems outright would be an act of professional self-harm, and would leave social workers with no voice in how tools that will shape their clients' lives are built. The case for caution is also real, and it is sharper than the enthusiasts allow. These systems generate plausible text without any capacity to know whether it is true. Their errors are not random; they fall hardest on the people who are already least able to correct the record. They erode the professional habits of uncertainty and curiosity that make good assessment possible. They can leak the most sensitive information a state holds about a family. And they arrive wrapped in a fluency that makes all of these problems invisible. The resolution is not a compromise between these positions but a boundary between them. Generative systems may assist with the expression of professional judgement. They may not perform it. A machine may help a social worker write, but it must never be allowed to decide what is true about a human life. Every guardrail in this article is a way of making that boundary hold under the pressure of a heavy caseload on a difficult day, which is the only condition under which guardrails are ever tested. Social work's authority rests on a promise: that a real person, accountable and named, has looked at a life with care and told the truth about what they saw. #ethical_practice in the age of generative AI means keeping that promise. The profession's commitment to #social_justice and to #human_dignity is not a set of values that sits above the technology. It is a set of obligations that must be built into every workflow, every procurement decision, every supervision session, and every sentence that enters a client's file. The tools will keep improving. The obligation will not change. Hashtags #EthicalAIinSocialWork #GenerativeAI_Ethics #SocialWorkTechnology #ClientAssessmentIntegrity #AIGovernanceInHumanServices #ResponsibleAI #DigitalSocialWork #AlgorithmicAccountability #HumanOversight #DataProtectionInPractice #SocialWorkResearch #AIandChildProtection #ProfessionalJudgementFirst #TechEthicsForPractitioners #STULIB_Research References Ahn, E., Choi, M., Fowler, P., & Song, I. H. (2025). Artificial intelligence (AI) literacy for social work: Implications for core competencies. Journal of the Society for Social Work and Research, 16(1), 9-26. https://doi.org/10.1086/735187 Alowais, S. A., Alghamdi, S. S., Alsuhebany, N., Alqahtani, T., Alshaya, A. I., Almohareb, S. N., Aldairem, A., Alrashed, M., Bin Saleh, K., Badreldin, H. A., Al Yami, M. S., Al Harbi, S., & Albekairy, A. M. (2023). Revolutionizing healthcare: The role of artificial intelligence in clinical practice. BMC Medical Education, 23, 689. https://doi.org/10.1186/s12909-023-04698-z An, R., & Lindsey, M. A. (2026). Ethical foundations of AI in social work. In Artificial intelligence in social work (pp. 19-43). Springer. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610-623). https://doi.org/10.1145/3442188.3445922 Boetto, H. (2025). Artificial intelligence in social work: An EPIC model for practice. Australian Social Work, 78(3), 292-305. https://doi.org/10.1080/0312407X.2025.2488345 Chuang, Y.-S., Sarkar, A. R., Hsu, Y.-C., Mohammed, N., & Jiang, X. (2025). Robust privacy amidst innovation with large language models through a critical assessment of the risks. Journal of the American Medical Informatics Association, 32(5), 885-892. https://doi.org/10.1093/jamia/ocaf037 Coeckelbergh, M. (2022). The political philosophy of AI: An introduction. Polity Press. Floridi, L. (2023). The ethics of artificial intelligence: Principles, challenges, and opportunities. Oxford University Press. Gerchick, M., Jegede, T., Shah, T., Gutierrez, A., Beiers, S., Shemtov, N., Xu, K., Samant, A., & Horowitz, A. (2023). The devil is in the details: Interrogating values embedded in the Allegheny Family Screening Tool. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (pp. 1292-1310). https://doi.org/10.1145/3593013.3594081 Hall, S. F., Sage, M., Scott, C. F., & Joseph, K. (2024). A systematic review of sophisticated predictive and prescriptive analytics in child welfare: Accuracy, equity, and bias. Child and Adolescent Social Work Journal, 41(6), 831-847. Harnett, P. H. (2024). Managing risk and uncertainty in the context of child protection decision making. The British Journal of Social Work, 54(6), 2435-2449. https://doi.org/10.1093/bjsw/bcae056 Hodgson, D., Goldingay, S., Boddy, J., Nipperess, S., & Watts, L. (2022). Problematising artificial intelligence in social work education: Challenges, issues and possibilities. The British Journal of Social Work, 52(4), 1878-1895. https://doi.org/10.1093/bjsw/bcab168 Hodgson, D., Watts, L., & Gair, S. (2023). Artificial intelligence and implications for the Australian Social Work journal. Australian Social Work, 76(4), 425-427. https://doi.org/10.1080/0312407X.2023.2247833 Huang, J., Yang, F., & Lee, J. (2026). Ethical challenges and AI governance in social work. In Artificial intelligence in social work (pp. 519-544). Springer. Jacobi, C. B., & Christensen, M. (2023). Functions, utilities and limitations: A scoping study of decision support algorithms in social work. Journal of Evidence-Based Social Work, 20(3), 323-341. https://doi.org/10.1080/26408066.2022.2159777 Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. https://doi.org/10.1145/3571730 Kemshall, H. (2021). Bias and error in risk assessment and management. HM Inspectorate of Probation Academic Insights. Kithulgoda, C. I., Vaithianathan, R., & Culhane, D. P. (2022). Predictive risk modeling to identify homeless clients at risk for prioritizing services using routinely collected data. Journal of Technology in Human Services, 40(2), 134-156. https://doi.org/10.1080/15228835.2022.2042461 Koenecke, A., Choi, A. S. G., Mei, K. X., Schellmann, H., & Sloane, M. (2024). Careless whisper: Speech-to-text hallucination harms. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. https://doi.org/10.1145/3630106.3658996 Mokander, J., Schuett, J., Kirk, H. R., & Floridi, L. (2024). Auditing large language models: A three-layered approach. AI and Ethics, 4, 1085-1115. https://doi.org/10.1007/s43681-023-00289-2 Molala, T. S. (2025). Social work in the digital age: Towards adopting artificial intelligence in social work. International Social Work. National Association of Social Workers. (2021). Code of ethics of the National Association of Social Workers. NASW Press. National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0), NIST AI 100-1. U.S. Department of Commerce. Ratner, H. F., & Elmholdt, K. (2023). Algorithmic constructions of risk: Anticipating uncertain futures in child protection services. Big Data and Society, 10(2). https://doi.org/10.1177/20539517231186120 Reamer, F. G. (2023). Artificial intelligence in social work: Emerging ethical issues. International Journal of Social Work Values and Ethics, 20(2), 52-71. https://doi.org/10.55521/10-020-205 Reamer, F. G. (2025). Artificial intelligence in the behavioral health professions. NASW Press. Sebastian, R., Kottekkadan, N. N., Thomas, T. K., & Niyas, K. K. M. (2025). Generative AI tools (ChatGPT) in social science research. Journal of Information, Communication and Ethics in Society, 23(2), 284-290. Segal, M. (2025). Confronting and managing ethical dilemmas in social work using ChatGPT. European Journal of Social Work, 28(1). Singer, J. B., Baez, J. C., & Rios, J. A. (2023). AI creates the message: Integrating AI language learning models into social work education and practice. Journal of Social Work Education, 59(2), 294-302. https://doi.org/10.1080/10437797.2023.2189878 Victor, B. G., Perron, B. E., Sokol, R. L., Fedina, L., & Ryan, J. P. (2021). Automated identification of domestic violence in written child welfare records: Leveraging text mining and machine learning to enhance social work research and evaluation. Journal of the Society for Social Work and Research, 12(4), 631-655. Weidinger, L., Uesato, J., Rauh, M., Griffin, C., Huang, P.-S., Mellor, J., Glaese, A., Cheng, M., Balle, B., Kasirzadeh, A., Biles, C., Brown, S., Kenton, Z., Hawkins, W., Stepleton, T., Birhane, A., Hendricks, L. A., Rimell, L., Isaac, W., ... Gabriel, I. (2022). Taxonomy of risks posed by language models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 214-229). https://doi.org/10.1145/3531146.3533088 World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. World Health Organization. Yang, F., & Liechty, J. M. (2026). Building AI literacy and competency in social work. In Artificial intelligence in social work (pp. 545-570). Springer. Yu, M.-H., & Rose, R. A. (2026). Algorithmic-assisted decision-making tools in child welfare practice: A systematic review. Research on Social Work Practice. https://doi.org/10.1177/10497315251350933

This article offers a full academic review of The Hero with a Thousand Faces, the 1949 book in which Joseph Campbell proposed that heroic stories from every part of the world share a single underlying pattern. The study has four aims. First, it explains the #monomyth clearly and completely, stage by stage, in language that undergraduate students can follow. Second, it places the book inside the intellectual history that produced it, including psychoanalysis, analytical psychology, ritual theory, and early folklore study. Third, it traces how the book travelled out of the university and into film, publishing, games, therapy, and classroom teaching. Fourth, it reviews the serious scholarly criticism that has gathered around the book, especially the criticism published in the last five years, which questions the claim of universality and asks who is served by a story shape built on one lonely male figure. The method is qualitative, interpretive, and text based: close reading of Campbell's own argument, combined with a thematic synthesis of recent peer reviewed literature. The findings suggest that Campbell's framework is best understood not as a scientific description of all human storytelling but as a powerful heuristic, that is, a thinking tool with real analytical value and real blind spots. The article argues that students should learn the #hero_journey and learn its limits at the same time, because the second lesson protects them from the first. It closes with practical recommendations for teaching, for research design, and for the responsible use of archetypal models in an age of interactive and machine generated narrative. Keywords: Joseph Campbell; monomyth; hero's journey; comparative mythology; archetypal criticism; narrative structure; myth criticism; literary theory 1. Introduction 1.1 Why this book still matters Very few academic books escape the library and enter ordinary conversation. The Hero with a Thousand Faces is one of the few that did. Published in 1949 by an American professor of literature who taught for most of his working life at Sarah Lawrence College, the book argued something bold and simple: that beneath the enormous surface variety of the world's myths there lies one story, told again and again, with different names, different landscapes, and different gods, but the same deep shape (Campbell, 2008). Campbell borrowed the word #monomyth from James Joyce and used it to name this shared pattern. A hero leaves the ordinary world, enters a strange region where hard things happen, wins something there, and comes back changed, carrying a gift for the community. That sentence is the whole book in miniature. It is also, at this point, part of the mental furniture of global popular culture. Students who have never heard of #Joseph_Campbell can still recognise the shape when they see it in a film. The reason is that Campbell's model did not stay in #comparative_mythology. It was picked up by screenwriters, by studio executives, by novelists, by game designers, by advertising agencies, by management consultants, and by psychologists, and each of those groups passed it on further (Hanney, 2024). This double life, half scholarly and half commercial, is exactly what makes the book worth studying carefully. A student who reads The Hero with a Thousand Faces is not only reading a book about myths. That student is reading the source code of a very large part of the entertainment industry, and also reading a document that has been accused of flattening the world's cultures into one convenient Western shape (Nield, 2022; Hanney, 2024). Both things are true at once, and holding both in mind is the beginning of good literary thinking. 1.2 Statement of the problem The problem this article addresses is a teaching problem as much as a scholarly one. In many universities and secondary schools, the #hero_journey is introduced as if it were a settled fact, a neutral map of story that anyone can pick up and apply. Students are given the circle diagram with its numbered stages, asked to plot a novel or film onto it, and rewarded when the plot fits. This practice has two weaknesses. The first is that it teaches pattern matching rather than interpretation. If the model is assumed to be universal, then finding it in a text proves nothing, because a sufficiently flexible model can be found anywhere. The second weakness is ethical and cultural. Recent scholarship argues that the model was assembled in a particular time and place, mid twentieth century North America, from fragments of other people's stories, and that it centres one figure, usually male, usually alone, in a way that quietly pushes aside stories built on community, on kinship, on cycles rather than conquests (Aston, 2024; Hambly, 2021; Nield, 2022). So the problem is this: how should students engage with a framework that is genuinely illuminating and genuinely partial at the same time? This article answers that question by doing three things at once: explaining the framework properly, showing what it does well, and showing where it fails. 1.3 Research questions The study is guided by five questions. What exactly does Campbell claim, and what is the internal logic of the #monomyth as he sets it out? Which intellectual traditions shaped the book, and how do those traditions limit what the book can see? How did the model move from academic #comparative_mythology into the film, publishing, and games industries, and what changed in the move? What are the strongest criticisms of the model in recent peer reviewed scholarship, and how strong are they? What is a responsible way for students and teachers to use the model today? 1.4 Significance of the study The significance of this study is chiefly pedagogical and critical. It is written for students, and it assumes that a student deserves the full picture rather than a simplified one. It brings together the classical presentation of the model, which is widely available but often thin, with the recent critical literature, which is rigorous but often written for specialists and locked behind the vocabulary of media theory. Placing the two side by side gives readers something they rarely get in one place: the case for Campbell and the case against him, presented fairly, with the evidence for each. The study is also timely. In 2024 a full special issue of a peer reviewed journal, Media Practice and Education, was devoted to what its editors called unravelling the hero's journey, gathering critiques from film education, feminist filmmaking, queer narrative studies, theme park design, tabletop role playing, and virtual reality (Hanney, 2024). In 2023, a major psychology journal published eight studies showing that people who retell their own lives in the shape of a hero's journey report more #meaning_in_life (Rogers et al., 2023). These two bodies of work pull in different directions. One says the model is culturally narrow and should be decentred. The other says the model, whatever its origins, does something real to human beings who use it. A serious student of literature should be able to sit with that tension. 1.5 Structure of the article Section 2 reviews the literature, both the older sources that Campbell drew on and the newer sources that criticise him. Section 3 sets out the method. Section 4 gives the necessary background on Campbell himself and on the intellectual world of the 1940s. Section 5, the longest, walks through the seventeen stages of the monomyth in order and in plain language. Section 6 examines the archetypes that populate the journey. Section 7 traces the model's migration into popular culture. Section 8 presents the criticism in full. Section 9 examines the psychological research. Section 10 considers educational use. Section 11 looks at interactive and digital media. Section 12 discusses what all of this means. Sections 13 to 15 give limitations, future directions, and conclusions. 2. Literature Review 2.1 The ground Campbell stood on No book appears from nowhere. The Hero with a Thousand Faces rests on at least four earlier bodies of work, and understanding them explains both the book's power and its problems. The first is psychoanalysis. Sigmund Freud had argued that myths and dreams speak the same language, a language of disguised wish and buried fear, and that the family drama of childhood is replayed in the symbolic material of culture. Campbell accepted a great deal of this. When he reads the hero's confrontation with a monstrous father figure as a psychological event rather than a historical one, he is thinking in a Freudian key. The influence of #Sigmund_Freud is visible on almost every page where Campbell moves from the outer adventure to the inner one. The second is analytical psychology. Carl Jung proposed that below the personal unconscious lies a deeper layer shared by all human beings, the #collective_unconscious, populated by inherited patterns of imagination that he called archetypes (Jung, 1968). This idea is the engine of Campbell's universalism. If every human mind carries the same deep images, then it follows almost automatically that every human culture will produce similar stories. Campbell's #archetype is Jung's archetype, adapted for literary use. It is worth noticing, and we will return to this, that Campbell inherited Jung's conclusion along with Jung's evidence, and the evidence was never as strong as the conclusion required. The third is ritual theory and anthropology. Arnold van Gennep had described the three part structure of #rites_of_passage: separation from the old status, a middle period of ambiguity and testing, and reincorporation into the community with a new status (van Gennep, 1960). Campbell's three great movements, departure, initiation, and return, are the same three movements. The middle phase, which later anthropologists called #liminality, is where the hero is neither what he was nor what he will become, and it is where almost everything interesting in Campbell's model happens. The fourth is folklore and early narratology. Vladimir Propp had shown, working on a corpus of Russian wonder tales, that a limited set of functions recurs across many stories in a stable order (Propp, 1968). This is a structural claim, and it is in some ways more rigorous than Campbell's, because #Vladimir_Propp restricted himself to a defined corpus from one tradition and counted what he found. Campbell ranged across the whole world and counted nothing. The contrast between these two methods is one of the most useful things a student can learn from this literature. 2.2 The reception: two streams After 1949, the reception of the book split into two streams that barely spoke to each other. The popular stream ran through the arts. Screenwriters and directors read the book as a manual. Christopher Vogler, working as a story analyst in Hollywood, condensed Campbell's seventeen stages into a twelve stage template and circulated it inside the studios, later publishing it as a book for writers (Vogler, 2007). Through this channel the monomyth became an industrial standard, a shared vocabulary in which a producer and a writer could argue about a script (Hanney, 2024). The academic stream was cooler. Historians of religion, folklorists, and classicists raised objections early and have repeated them since. The objections concern method: that Campbell chose examples that fit, ignored those that did not, stripped stories from the cultures that made them, and defined his categories so loosely that almost any narrative could be squeezed into them (Lefkowitz, 1990; Segal, 1978). This critical position never had the cultural reach of the popular one, but it has never gone away either. 2.3 Recent scholarship, 2021 to 2025 The last five years have seen the critical stream gather real force, and this is the literature that most matters for a current student. Hambly (2021) examined the claim of universality directly, arguing from screenwriting practice that the journey is not so universal as its promoters believe, and that many successful narratives, particularly those written by women and by writers outside the Anglo American mainstream, do not follow it. Nield (2022) coined the phrase #Hollywood_McMyth, describing the model as a mass produced narrative formula stitched together from decontextualised fragments of world mythology, sprinkled with Jungian archetypes and psychoanalytic readings of dreams, and then sold back to the world as though it were the world's own inheritance. Hanney (2024), introducing the special issue of Media Practice and Education devoted to the topic, developed this argument into a full critique. The charge is not merely that Campbell was sloppy. It is that the framework functions as a monolithic dogma which has crowded out other voices and other ways of telling, and that its eclectic method reflects a colonial inheritance in which other people's stories are appropriated to prove the centrality of European culture. This is a strong claim and it deserves careful examination rather than automatic agreement, which Section 8 attempts to give it. Within the same issue, Aston (2024) turned to teaching, arguing that film and media departments continue to hand the monomyth to students as if it were the only available architecture of story, and proposing multiperspectival alternatives. Gough-Brady (2024) connected the heroic protagonist to the figure of the productive neoliberal citizen and asked what a feminist filmmaker is supposed to do with a template built for that figure. Taylor (2024) examined how heroes' journeys and the straight line of time they assume can be queered, and what happens to the model when the tidy forward march of the narrative is refused. Bowman (2024) studied how players in role playing games weave myth, narrative, and identity together, showing that participants do use mythic structures but bend them constantly. Gröppel-Wegener (2024) followed the hero into theme parks, where the visitor is invited to become the immersive protagonist. Remedios, Mathew, and Schleser (2024) considered what happens to a linear journey inside virtual reality, where the audience can look away, wander off, or refuse the call entirely. Alongside this critical wave, a different kind of study appeared. Rogers and colleagues (2023), working in social and personality psychology, distilled the journey into seven elements, protagonist, shift, quest, allies, challenge, transformation, and legacy, built a measurement scale from them, and ran eight studies. They found that people whose life stories contained more of these elements reported more meaning in life, and that an intervention teaching people to retell their lives in this shape causally increased meaning and #resilience. Whatever one thinks of Campbell as a mythographer, this is evidence that the pattern does something when human beings apply it to themselves. Applied literary studies have continued to use the model as an analytical instrument. Wu and Wang (2022) applied Campbell's structure to Spirited Away and read Chihiro's story as a #heroine_journey. Dewi and Yusrina (2023) mapped archetypal symbols and the journey stages onto the film adaptation of Alice in Wonderland. Arvas (2025) reworked the journey through an ecological and feminist lens in children's literature, replacing conquest with reciprocity and dispersing agency among several protagonists rather than concentrating it in one. Ambasciano (2021) examined how Disney and Lucasfilm have handled, or failed to handle, the more troubling parts of Campbell's legacy in the Star Wars franchise. 2.4 The gap this article fills Reading across this literature, one gap is obvious. The explanatory accounts of the monomyth are mostly uncritical, and the critical accounts mostly assume the reader already knows the model. Students are therefore offered either a diagram without a warning or a warning without a diagram. This article deliberately supplies both, at length, in one place, so that the reader can judge for themselves. 3. Methodology 3.1 Research design This study uses a qualitative, interpretive design. It is a work of #textual_analysis and critical synthesis rather than an empirical investigation. No human participants were involved and no new data were generated. The object of study is a body of texts: Campbell's book itself, the intellectual sources behind it, and the scholarly literature that has responded to it. Two complementary methods are used. The first is close reading. Campbell's argument is examined in its own terms, following his logic, reconstructing his categories, and testing whether his stages hold together as a coherent structure. The purpose here is charitable interpretation: to present the model as strongly as its author would present it, before any criticism is made. A critique of a weakened version of an argument is worth nothing. The second is #thematic_analysis of the secondary literature. Recent peer reviewed sources were read and coded for recurring concerns. Five themes emerged and organise Section 8: universality and #source_selection_bias, cultural appropriation and #Eurocentrism, gender and the position of women, ideological content and #individualism, and industrial standardisation. These themes were not imposed in advance. They surfaced repeatedly across independent authors, which is itself a form of evidence. 3.2 Selection of sources Sources were selected on three criteria. Recency was the first: the critical and applied literature reviewed here is drawn almost entirely from the period 2021 to 2025, so that students encounter the current state of the debate rather than a fossilised version of it. Peer review was the second: journal articles and scholarly books were preferred over commentary, with a small number of exceptions where an influential idea, such as the McMyth argument, first appeared outside the peer reviewed press and was subsequently taken up within it. Relevance was the third: sources had to address Campbell's model directly, either by applying it, by criticising it, or by testing it. A small number of older works are cited. These are the primary and foundational texts, Campbell himself, Jung, Propp, van Gennep, Vogler, Murdock, Le Guin, and the early critical responses. It is not possible to write honestly about a 1949 book while citing only work from the last five years, and no attempt has been made to disguise that. 3.3 Analytical framework The article treats the monomyth as a heuristic rather than a law. A heuristic is a tool that helps you notice things. It is judged by what it reveals and what it hides, not by whether it is true in the way a law of physics is true. This framing matters, because much of the argument between Campbell's defenders and his critics comes from a confusion of categories. Defenders often praise the model for its usefulness and then slide into claiming it is universally true. Critics often demonstrate that it is not universally true and then conclude it is useless. Both moves are mistakes. Treating the model as a heuristic keeps the two questions apart, and this article keeps them apart throughout. 3.4 Limitations of the method Three limitations should be stated at the outset. First, interpretive work is not replicable in the way experimental work is; another careful reader might weight the same sources differently. Second, the review is confined to sources available in English, which is a real problem for an article about a theory accused of Western narrowness, and the irony is not lost on the author. Third, the article discusses Campbell's readings of myths from many traditions but is not itself a work of primary comparative mythology; where it reports what a myth contains, it reports what Campbell and his commentators say it contains. 4. Campbell and His Moment 4.1 The man and the method Joseph Campbell was born in New York in 1904 and died in 1987. He was not a field anthropologist. He did not collect stories from living communities, did not learn the languages of the peoples whose myths he retold, and did not do the slow, unglamorous work of establishing what a given story meant to the people who told it. He was a reader, a comparatist, and a synthesiser of extraordinary range, working almost entirely from translations, collections, and the scholarship of others (Campbell, 2008). This is not a small point and it is not a smear. It defines the kind of book The Hero with a Thousand Faces is. It is a work of pattern recognition performed on a library. Its strength is the strength of a wide reader who has noticed something that specialists, buried in one tradition each, could not see. Its weakness is the weakness of any argument built from secondary sources: the reader has no way to check whether a story has been reported accurately, whether it has been trimmed to fit, or whether its meaning in its own culture bears any resemblance to the meaning Campbell gives it. Critics have pressed hard on precisely this point, and they are not wrong to press (Nield, 2022; Hanney, 2024). 4.2 The intellectual climate of 1949 The book appeared four years after the end of the Second World War. This matters more than it first appears. A generation had just watched industrial civilisation produce industrial slaughter, and a hunger for something deeper than politics, something that connected human beings below the level of nation and ideology, ran through the intellectual culture of the period. Campbell's message, that all peoples dream the same dream, that beneath our quarrels lies one shared imagination, answered that hunger directly. It is a generous message, and part of the book's enduring appeal is that generosity. But recent scholarship insists that we notice what else was in the air. Campbell wrote from within a confident postwar American culture, at the beginning of the period when that culture would export its stories to the entire planet. A theory which says that the American way of telling a story is in fact everybody's way of telling a story is, at minimum, convenient. Aston (2024), following Bond and Christensen, describes the framework as a particular take on myth, constructed in a postwar North American context, focused on one character and serving one point of view. The word to hold on to is particular. Campbell's claim is universal; his standpoint was not. 4.3 The central thesis in Campbell's own terms Campbell's summary of his own thesis is famous and worth restating in plain form. A hero goes out from the everyday world into a region of wonder. Strange powers are met there. A decisive victory is won. The hero comes back with the power to give gifts to others (Campbell, 2008). Everything else in the book is an elaboration of that sentence. The three movements are #departure, #initiation, and #return. Inside them Campbell places seventeen stages. He is clear that not every myth contains every stage, that stages can be combined, doubled, or omitted, and that the order can vary. This flexibility is both his defence and his vulnerability. It is a defence because it means the model is not a rigid checklist. It is a vulnerability because a model that permits any stage to be missing and any order to be rearranged is very hard to falsify, and a claim that cannot fail a test cannot really pass one either. 5. The Structure of the Monomyth: The Seventeen Stages What follows is a full walk through the model. Each stage is explained in ordinary language, illustrated, and, where relevant, connected to the psychological meaning Campbell attaches to it. Students should read this section as an exposition of Campbell's view, not as the author's endorsement of it. The criticism comes later, and it comes hard. 5.1 Act One: Departure Stage 1: The Call to Adventure The story begins with a disturbance. Something arrives to break the ordinary world open: a message, a wound, a stranger, an accident, a death, an invitation. The hero is living a life that seems settled, and the settled life stops being possible. The #call_to_adventure is not merely a plot device. In Campbell's reading it marks the moment when a person's centre of gravity shifts from a known world to an unknown one. The call may be dazzling or it may be terrible. What matters is that it names something the hero has been avoiding or has not yet imagined. In psychological terms it is the moment when the unconscious sends a signal that the current arrangement of the self has become too small. Examples are everywhere once you look. A letter arrives for a boy living under the stairs. A message is found in the memory of a droid on a desert farm. A frog speaks to a princess who has lost her golden ball down a well. Stage 2: Refusal of the Call Very often the hero says no. The #refusal_of_the_call is Campbell's acknowledgment that human beings do not walk cheerfully into transformation. Fear, duty, comfort, obligation, or simple disbelief holds them back. Campbell treats refusal seriously and gives it a dark reading. To refuse the call is to convert the adventure into its opposite. The person who refuses becomes, in his phrase, a victim to be saved: their world becomes a wasteland, their life dries out, and everything they cling to loses meaning precisely because they clung to it. This is one of the most morally loaded moments in the whole framework, and it tells us a great deal about Campbell's values. The refusal is not neutral. Growth is presented as obligation. Stage 3: Supernatural Aid Once the hero commits, help appears. The #supernatural_aid stage introduces the protective figure, often old, often strange, who gives the hero something for the road: a talisman, a piece of advice, a weapon, a warning, a map. Campbell reads this figure as a projection of the benign, protecting power of destiny, the sense that once a person genuinely commits to a difficult path, resources appear that were invisible before. Whether one believes that as metaphysics or merely as psychology, it is a reliable narrative fact: the mentor arrives after the decision, not before it. Stage 4: The Crossing of the First Threshold Now the hero leaves the known world. The #crossing_the_threshold is guarded, and the guardian is there to enforce the boundary of the ordinary. The guardian may be a person, a beast, a rule, a bureaucrat, a wall, a customs officer, a locked gate. The threshold marks the edge of the map, the place where the community's protection ends. Beyond it, the rules that the hero has relied on all their life no longer apply. To cross is to accept that the hero's own competence, not the community's rules, will now have to do the work. Stage 5: The Belly of the Whale The final stage of departure is the strangest and, for many readers, the most memorable. The #belly_of_the_whale is the moment when the hero is swallowed. The passage of the threshold turns out to be a form of self annihilation. The hero appears to die. Campbell gathers the examples: Jonah in the fish, the hero descending into a cave, the initiate buried in the earth, the shaman dismembered in a vision. The point is that the hero does not simply walk into the new world. The hero is consumed by it, and what emerges is not quite the same being that went in. This stage is the clearest expression of the whole model's underlying claim: that transformation requires a death, and that the death is real, even when it is symbolic. 5.2 Act Two: Initiation Stage 6: The Road of Trials The middle of the story is a sequence of tests. The #road_of_trials is where the hero is broken down and rebuilt: repeated ordeals, repeated failures, gradual competence. Campbell notes that the hero is often helped here by the advice or the tools acquired earlier, and that the trials frequently come in threes. The psychological content is straightforward. The old personality, the one that fitted the ordinary world, is not adequate to the new one. The trials strip it away piece by piece. It is worth noticing that Campbell places most of the story's actual content in this stage, and that it is therefore the least precisely described part of his framework. A model whose middle act is simply named "trials" is not offering much guidance about what should happen there. Stage 7: The Meeting with the Goddess Here the modern reader begins to feel the age of the book. The #meeting_with_the_goddess describes an encounter with a figure of total, unconditional love, whom Campbell casts as feminine, maternal, and cosmic. She is the world, the nourishing mother, the queen of the universe. The hero who is worthy is united with her. Read symbolically, this is a stage about the hero's capacity to receive love and to recognise a power that is not conquered but joined. Read literally, it is a stage in which the woman is a landscape the hero arrives at. Both readings are available in Campbell's text, and it is honest to say that the second is not merely a hostile misreading. The criticism this attracted is examined in Section 8.3. Stage 8: Woman as Temptress The stage that follows makes the difficulty explicit. #woman_as_temptress presents temptation, the pull of the material and the physical, the desire that would keep the hero from the spiritual goal. Campbell means something broader than sex, and he says so: the temptation is to settle for the world of appearances. But he chose to name that temptation with the figure of a woman, and the naming has consequences. This is a moment where students should slow down and think carefully about the difference between what an author means and what an author's language does. Campbell's intention was symbolic. His symbol was a person. Other people are members of that category, and they have to live in a culture shaped by stories built on it. Stage 9: Atonement with the Father The centre of the model, and the stage Campbell invests with the most weight. #atonement_with_the_father is the confrontation with whatever holds ultimate power in the hero's life: the father, the god, the law, the source of authority and judgement. Campbell splits the word: at one ment, becoming at one with. The hero must abandon the self that was formed in fear of this power. The father is initially seen as an ogre, a destroyer, a monster; the atonement occurs when the hero sees past the mask and recognises that the terrifying father and the merciful mother are two faces of a single reality. The most famous popular version of this is the moment when a masked villain in black tells a young man that he is his father, and the young man must find a way to be his own person while carrying that knowledge. Campbell's framework explains why that scene has the force it does. It is not a plot twist. It is the hinge of the entire structure. Stage 10: Apotheosis Having faced the source of authority, the hero enters a state of expanded understanding. #apotheosis means becoming divine, and Campbell uses it for a moment of rest and clarity in which the hero sees the world whole and stops being afraid of it. The opposites that structured the old life, good and evil, life and death, self and other, are perceived as belonging together. This is the most explicitly mystical stage in the model and the one that owes most to Campbell's lifelong reading of Buddhist and Hindu philosophy. It is also the stage that film adaptations of the model almost always drop, because it is a state rather than an event and states are hard to film. Stage 11: The Ultimate Boon The goal is achieved. The #ultimate_boon is the thing the hero came for: the elixir, the treasure, the fire, the knowledge, the rescued person, the cure. Campbell's point is that the boon is almost always smaller than the transformation that made it possible. The gods have it easily; the hero has to be remade in order to hold it. The object is the story's excuse. The change in the hero is the story's subject. 5.3 Act Three: Return Stage 12: Refusal of the Return Having won the boon, the hero may not want to go home. #refusal_of_the_return is the temptation to stay in bliss, to remain in the world of wonder, to keep the enlightenment private. Campbell has no patience with this. The journey is not complete until the gift comes back. A transformation that benefits only the transformed is, in his moral scheme, a failure. This is one of the places where the model is not merely descriptive but is quietly making an ethical argument about what a life is for. Stage 13: The Magic Flight Sometimes the return is a chase. The #magic_flight occurs when the boon has been taken against the will of its guardian and the hero must escape, throwing obstacles behind them, transforming, tricking, running. Stage 14: Rescue from Without Sometimes the hero cannot get back alone and must be brought back. #rescue_from_without is the stage that most directly contradicts the myth of the self made hero, and it is striking how rarely popular retellings of Campbell mention it. The hero, exhausted or unwilling or lost in bliss, is retrieved by the world, by allies, by the community that needs them. Campbell includes this stage explicitly. The individualist reading of his model, which the critics attack, is not entirely faithful to his own text. That is a point in Campbell's favour, and it should be made. Stage 15: The Crossing of the Return Threshold The hard part. #crossing_the_return_threshold is the problem of coming home. How does the hero carry the wisdom of the other world back into the ordinary one, where nobody has seen what they have seen and nobody wants to hear about it? This stage describes something every returning soldier, every graduate, every migrant, and every person who has survived a serious illness recognises immediately. The wisdom does not translate. The ordinary world has continued without you and does not know that you are different. The hero must find a language for what happened, or the journey remains locked inside them. Stage 16: Master of Two Worlds The resolution. The #master_of_two_worlds is the person who can move freely between the ordinary world and the world of wonder, between the outer life and the inner one, without being destroyed by either. Stage 17: Freedom to Live The end. #freedom_to_live describes a state in which the hero no longer fears death and no longer clings to the future. They are able to be fully present in the moment they are actually in. Having died symbolically and come back, they have nothing left to defend. 5.4 What the structure is really about Read as a whole, the seventeen stages describe one process expressed three times: a death and a rebirth. The hero dies to the ordinary world at the threshold, dies to the self at the atonement, and dies to the fear of death at the end. Each death is followed by a return with something gained. This is why the model resonates. It is not primarily a plot formula. It is a description of #transformation, and it maps onto real experiences that are otherwise very hard to describe: adolescence, grief, migration, addiction and recovery, conversion, illness, the end of a marriage, the loss of a belief. Whatever its scholarly failings, and they are serious, the model survives because a great many people read it and recognise their own lives. 6. The Cast: Archetypes in the Journey Campbell describes stages. The figures who appear during those stages were systematised later, most influentially by Vogler (2007), whose Hollywood adaptation gave each of them a job. Students meet these figures constantly and should be able to name them and to see the machinery behind them. The Hero. Not necessarily brave, strong, or good. In Campbell's terms the hero is the one who leaves, who accepts the loss of the known world. The defining property is willingness to be changed. Vogler adds a functional definition useful for writers: the hero is the character the audience is invited to learn through. The Mentor. The #mentor_archetype gives the hero what is needed to survive the threshold: training, a gift, a truth. The mentor's second and less noticed function is to die or to depart, because the hero cannot complete the journey while the mentor is still available to be leaned on. The mentor is a scaffold, and scaffolds come down. The Threshold Guardian. The #threshold_guardian tests whether the hero is serious. Guardians are usually not the real enemy, which is why so many stories turn them into allies once they have been passed. In psychological terms they are the internal voices that ask whether you really want this. The Shadow. The #shadow_archetype is the antagonist, but in the Jungian reading that Campbell inherits, the shadow is what the hero has refused to acknowledge in themselves. This is why the best villains are recognisable to the hero, why the confrontation feels like a mirror, and why defeating the shadow so often means accepting rather than destroying it. The Trickster. Disorder in person. The trickster breaks the story's rules, punctures the hero's self importance, and prevents the whole apparatus from becoming pompous. Trickster figures are among the strongest counter evidence to the neatness of the monomyth, because in many traditions, particularly in West African, Native American, and Norse #folklore, the trickster is not a supporting character but the central one, and trickster stories do not have the moral shape that Campbell's model requires. The Shapeshifter. The character whose loyalty cannot be established. The shapeshifter generates doubt, which is a narrative resource. The Herald. The one who delivers the call. The Ally. The companion. Note how thin this category is in Campbell compared with how thick it is in actual practice. Recent work argues that this thinness is not accidental but ideological: a framework centred on one transforming individual is structurally unable to give proper weight to companionship, and the alternative paradigms proposed in the 2024 critical literature deliberately foreground community, companionship, and commonality instead (Hanney, 2024). 7. From Library to Box Office: The Migration of the Monomyth 7.1 The Hollywood transmission The single most consequential event in the reception history of The Hero with a Thousand Faces was its adoption by filmmakers. #George_Lucas has said repeatedly that Campbell's book helped him give shape to the story that became Star Wars, and the influence is not subtle: a young man on a farm, a call, a refusal, an old mentor with a weapon and a secret, a threshold, a descent, a masked father, a return (Ambasciano, 2021). It is reported that Stanley Kubrick gave a copy of the book to Arthur C. Clarke while they were working on 2001: A Space Odyssey. The transmission was formalised by #Christopher_Vogler. Working as a development executive, Vogler wrote an internal memo summarising Campbell's structure for practical use by screenwriters, reducing seventeen stages to twelve and renaming them in language a production meeting could handle: ordinary world, call to adventure, refusal, meeting with the mentor, crossing the threshold, tests and allies and enemies, approach to the inmost cave, the ordeal, the reward, the road back, the resurrection, return with the elixir (Vogler, 2007). The memo circulated, then became a book, and the book became a fixture of #screenwriting education worldwide. Something important happened in that translation, and students should see it clearly. Campbell was describing. Vogler was prescribing. Campbell asked what myths have in common. Vogler answered the question of what a commercially successful script should contain. The mystical stages, apotheosis in particular, were dropped or converted into something filmable. The model became a template, and a template is a machine for producing sameness. 7.2 Saturation Once inside the industry, the model spread everywhere. It structures large parts of contemporary #popular_culture, including franchise cinema, animated features, young adult fiction, superhero narratives, and the tutorial arc of a great many #video_games, where the player is literally called, threshold crossed, trained by a mentor, tested, and rewarded. It is visible in the wizarding school novels, in the story of a hacker who learns the world is a simulation, and in a thousand smaller works that borrowed the shape without knowing where it came from. It also escaped fiction altogether. Marketing departments describe the customer as the hero and the brand as the mentor. Corporate leaders are trained to narrate their careers as journeys. Politicians borrow the arc. This is worth noticing because it changes the ethical stakes: a story shape that is used to sell things and to legitimise people is no longer only an aesthetic matter. 7.3 What saturation cost Recent scholarship treats this saturation as the central problem rather than as a triumph. Hanney (2024) argues that the framework has become a monolithic dogma which drowns out other voices and other ways of storytelling. Nield's phrase, the #Hollywood_McMyth, captures the objection exactly: a mass produced formula, standardised, portable, and available everywhere, which crowds out local and specific alternatives in the way that a global chain crowds out local kitchens. The complaint is not that any single film using the structure is bad. It is that when one structure is taught as the structure, the range of stories a culture is capable of telling narrows, and the narrowing is invisible to the people inside it. Aston (2024) shows how this happens institutionally. Film and media departments, pressed to align with industry, teach the monomyth because industry uses it. Students learn it as the professional standard. They graduate, enter the industry, and use it. The loop closes. #narrative_plurality, in this account, does not fail on the merits; it simply never gets taught. 8. The Critical Case Against the Monomyth This section presents the case against Campbell as strongly and fairly as possible. Students should engage with it seriously. Some of it is decisive; some of it, in the author's judgement, overreaches; the article says which is which. 8.1 Source selection bias The first and most damaging criticism is methodological. Campbell claimed to find a pattern in the myths of the world. To establish such a claim you would need a defined corpus, a sampling procedure, explicit criteria for what counts as an instance, and an honest count of the cases that do not fit. Campbell provided none of these. He selected the stories that supported his thesis and did not systematically report those that did not. This is source selection bias, and it is the standard objection raised by folklorists, who have generally regarded the monomyth as a non scholarly approach for exactly this reason. The point can be made sharper by comparison. Vladimir Propp worked on a bounded corpus of one hundred Russian tales, stated his method, and derived his functions from what was actually there (Propp, 1968). One can check Propp. One cannot check Campbell, because there is nothing to check against. This criticism is, in the author's view, correct and not answerable. It does not show that the model is useless. It shows that the model has never been tested and that Campbell's universal claim was not earned. 8.2 Eurocentrism, appropriation, and the colonial inheritance The second criticism is cultural and political, and it has become the dominant strand in recent work. The charge is that Campbell's eclectic method takes stories out of the cultures that made them, strips away the ritual, linguistic, and social contexts that gave them meaning, and reassembles the fragments into a pattern that turns out to look remarkably like a Western romance of individual self realisation. The result, on this reading, is a #decolonial_critique waiting to happen: other people's sacred narratives are appropriated in order to demonstrate that European culture is the universal yardstick (Nield, 2022; Hanney, 2024). The evidence for Eurocentrism is not merely theoretical. Many traditions organise narrative around principles that Campbell's model cannot represent. Collectivist storytelling in many #collectivist_cultures foregrounds group survival, ancestral obligation, and cyclical rather than linear time; the individual protagonist who leaves the community in order to be transformed and then returns to save it is not a neutral shape but a culturally loaded one. Trickster cycles resist the moral arc. Oral traditions that exist to be performed and re performed do not aim at the closure that the model assumes. Le Guin (2019) offered a well known alternative image: the story as a carrier bag, a container for gathering, rather than a spear, a weapon for killing, and observed that the spear shaped story has been mistaken for the only shape there is. Where should a careful student land? The strong form of the charge, that Campbell was engaged in colonial appropriation, is contestable; his motive appears to have been the opposite, a sincere attempt to show human unity. But motives do not settle effects. The effect of a universal claim advanced from a particular standpoint, and then exported globally through the most powerful entertainment industry in history, is to make one culture's story shape look like nature. That effect is real, whatever Campbell intended, and the criticism stands. 8.3 The problem of gender The third criticism is the one students notice first and fastest. In Campbell's model the hero is grammatically and conceptually male, and women appear overwhelmingly as functions of his development: the goddess who loves him, the temptress who distracts him, the mother he must separate from. There is no stage in which a woman departs, is tested, and returns. #gender_bias is not incidental to the structure; it is written into the names of the stages. The response to this began early. Maureen Murdock, who had studied with Campbell, proposed an alternative heroine's journey built on a different problem: not separation from the mother and confrontation with the father, but the wounding produced by identifying with a masculine ideal, the descent, and the eventual healing of the split between the masculine and feminine within (Murdock, 1990). Campbell reportedly told her that women do not need to make the journey because they are the place the hero is trying to reach. It is difficult to imagine a reply that could more completely confirm the criticism. Recent work has developed this line considerably. Gough-Brady (2024) connects the heroic protagonist to the ideal of the productive neoliberal citizen and asks what remains for the feminist filmmaker who does not wish to reproduce that figure. Arvas (2025) demonstrates an alternative in practice, reworking the journey in children's literature so that agency is dispersed across several protagonists and heroism is redefined as relational attentiveness and shared transformation rather than conquest. Wu and Wang (2022) take a different route, showing that a female protagonist can be read through the monomyth productively, which suggests that the framework can be adapted even if it was not designed for the purpose. #feminist_criticism of Campbell is therefore not monolithic, and students should notice that it contains both a rejection and a repair. 8.4 Ideology: the lonely individual The fourth criticism concerns what the model teaches without saying it. A structure in which one exceptional person separates from the group, is transformed by private ordeal, and returns to bestow gifts on a passive community is a structure that encodes individualism as the natural form of a meaningful life. Community is background. Allies are equipment. Institutions are obstacles. The 2024 critical literature responds to this by proposing narrative paradigms that emphasise community, companionship, and commonality (Hanney, 2024), and by exploring what happens when the straight forward line of heroic time is refused altogether, which is the burden of Taylor's (2024) work on #queer_narrative and heteronormative timelines. It should be said in fairness that Campbell's own text is less individualist than its industrial descendants. He insists on the return; he includes rescue from without, in which the hero is saved by others; he says explicitly that a boon kept private is a journey failed. The lonely hero of the Hollywood template is a simplification of Campbell, not a faithful transcription of him. But the simplification is the version that travelled, and the version that travelled is the one doing the cultural work. 8.5 Unfalsifiability and #reductionism The fifth criticism is philosophical. Campbell's categories are so elastic, and his permission to omit or reorder stages so generous, that it is unclear what a myth would have to look like in order to count as a counterexample. A theory that cannot be contradicted by any possible evidence is not a strong theory; it is a lens that always shows you what you brought with you. This is the core of the older academic objection (Lefkowitz, 1990; Segal, 1978), and it has not been satisfactorily answered. Related to this is the charge of reduction. To say that the Buddha's enlightenment, the crucifixion, the descent of Inanna, and the flight of a folk hero are all instances of one pattern is to say something true at a very high level of abstraction and, at the same time, to lose almost everything that makes each of them what it is. The abstraction is purchased with the specific, and the specific is where the meaning of a myth lives for the people who hold it sacred. 8.6 A balanced verdict Taking the criticism as a whole, three conclusions seem defensible. The universal claim fails. Campbell did not demonstrate that all cultures tell one story, and there is good reason to believe that they do not. The cultural criticism largely succeeds. The model carries a specific, Western, individualist, masculine set of assumptions, and its global spread has narrowed rather than widened the range of stories that get made. The analytical value survives, in reduced form. As a heuristic, a set of questions to bring to a text, the model still does real work. It draws attention to thresholds, to mentors, to the structure of transformation, and to the relationship between an outer plot and an inner change. Used as a question, it is useful. Used as an answer, it is a cage. 9. The Psychological Turn: Does the Journey Do Anything? A curious development complicates the picture. While literary and media scholars were dismantling Campbell's claims about myths, psychologists were finding that his pattern has measurable effects on people. Rogers and colleagues (2023) reduced the journey to seven elements: a protagonist, a shift, a quest, allies, a challenge, a transformation, and a legacy. From these they constructed a scale to measure how strongly a person's own life story resembles the pattern. Across eight studies, including community samples of older adults, they found that people whose self narratives contained more of these elements reported greater meaning in life. They then designed a #restorying intervention that guided participants to retell their own lives in this shape, and found that it causally increased meaning, increased the perception of meaning in an unrelated task, and increased #resilience in the face of difficulty. What should we make of this? Three readings are possible and students should be able to argue each. The strong reading says this vindicates Campbell: the pattern works on people because it is written into people. This reading is too quick. The studies show that a narrative structure, once supplied, is useful to the people who adopt it. They do not show that the structure is innate, and they were conducted largely in cultures already saturated with the structure. A story shape that a culture has spent seventy years installing in its members will naturally feel meaningful to those members. The weak reading says the finding is purely cultural: the participants are Western, the story is Western, and the effect is the sound of a culture recognising itself. This reading is too dismissive. Even a culturally supplied narrative that produces genuine resilience in real people is doing something worth understanding, and the elements the researchers isolated, purpose, social support, growth through challenge, contribution to others, correspond to variables that #narrative_psychology has independently linked to wellbeing. The middle reading, which this article favours, is that the journey is a useful cultural technology whose usefulness does not depend on its universality. It helps people organise a messy life into a shape with direction. That is a real service. It is not evidence for the collective unconscious, and it does not rescue the claim that every people on earth tells this story. There is also a caution to state plainly. A framework which insists that suffering must be redemptive, that the ordeal must yield a boon, can be cruel to people whose suffering yields nothing. Not every illness teaches. Not every loss is a threshold. #self_narrative work that forces every experience into a growth arc risks shaming those who are simply enduring, and any clinical or educational use of the model should hold that risk in view. 10. The Monomyth in Education 10.1 Why it is taught The model is popular with teachers for good reasons. It gives students a vocabulary. It makes an abstract subject, narrative structure, concrete and diagrammable. It connects the canon to films students already love, which is a genuine pedagogical gift. It supports comparison across cultures and periods, and comparison is where #literary_education begins. 10.2 What goes wrong The trouble starts when the model is presented as fact and applied as a checklist. Three failures recur. The first is confirmation. Given a flexible template and an instruction to find it, students will find it. They then believe they have discovered something, when they have only demonstrated the elasticity of the template. This teaches the opposite of #critical_thinking. The second is flattening. Students who learn to see every story as a hero's journey stop seeing what makes each story different, which is the only thing worth seeing. The third is exclusion. When one structure is presented as the structure, texts that do not use it, and there are many, appear defective rather than different. Aston (2024) documents this happening at university level, where the industrial dominance of the monomyth reproduces itself through the curriculum. 10.3 A better classroom method A more defensible #pedagogy treats the model as one lens among several and puts it under pressure deliberately. Teach the model fully, including the stages that the popular version drops. Students should know that Campbell included rescue from without and insisted on the return, because those stages complicate the individualist reading they will otherwise absorb. Teach the criticism at the same time, not as an afterthought. A student who learns the diagram in week three and the objections in week eleven has already internalised the diagram as truth. Ask students to find a counterexample. This single assignment does more than any lecture. Give them the task of locating a narrative that resists the model, and require them to explain what structure it uses instead. Oral traditions, collectivist narratives, trickster cycles, modernist novels, ensemble stories, and much contemporary fiction will serve. This converts the class from textual analysis by template into genuine inquiry. Ask what the model cannot see. This is the most valuable habit a literature student can acquire, and it generalises far beyond Campbell. 11. Interactive, Immersive, and Machine Made Stories The most interesting current pressure on Campbell's framework comes from forms of storytelling he could not have imagined. In tabletop and live action role playing, the participant is not receiving a story but producing one, and Bowman (2024) shows that players do weave mythic material into their play but continually reshape it, sharing the protagonist role, refusing closure, and using play as a site for the negotiation of identity rather than the completion of an arc. The single hero dissolves into a group. In theme parks, Gröppel-Wegener (2024) traces the design logic by which the visitor is invited to become an immersive protagonist, walking a physical threshold into a themed world. Here the journey becomes architecture, and the boon is a purchased experience, which raises questions about what is left of a myth when it has been fully commercialised. In #virtual_reality, the linear ordering that the monomyth depends upon simply breaks. Remedios, Mathew, and Schleser (2024) examine parallel interactive narratives in which the participant's attention cannot be controlled and multiple story threads run at once. A model built on a fixed sequence of stages has little to say about a form in which the audience can turn its back on the call to adventure and go and look at something else. The critical literature therefore proposes #immersive_storytelling and #transmedia forms as evidence that narrative structure is not one thing and never was. #artificial_intelligence adds a further complication that current scholarship is only beginning to address. Systems that generate stories are trained on the corpus of existing stories, and that corpus is saturated with the Hollywood template. There is a plausible risk of a feedback loop in which the machine learns the monomyth because the monomyth dominates the training data, then produces more of it, which becomes training data in turn. If the twentieth century installed one story shape through the studio system, the twenty first may harden it through automation. This is a hypothesis rather than an established finding, and it is exactly the kind of question that the next generation of researchers is well placed to test. 12. Discussion Bringing the threads together, four propositions can be defended. First, the monomyth is a heuristic, not a law. Campbell's own hedging, that stages may be omitted, combined, or reordered, gives the game away. The framework is a set of prompts for noticing, and it should be evaluated as such. Judged as a scientific description of human narrative, it fails. Judged as an instrument for reading, it earns its place. Second, its cultural effects are separable from its scholarly merits and are more consequential. Whether or not Campbell was right about Polynesian myth matters to a small number of specialists. Whether the entertainment industry of the entire planet organises itself around one Western, individualist, masculine story shape matters to everyone. The recent literature is right to move the argument onto this ground (Hanney, 2024; Aston, 2024; Nield, 2022). Third, the model's psychological usefulness is real but does not prove its universality. The Rogers et al. (2023) findings are robust and important, and they establish that this narrative shape, when adopted, changes how people experience their lives. They do not establish that the shape is native to the human mind, and the distinction is not a quibble; it is the entire question. Fourth, the honest position is neither devotion nor dismissal. Campbell noticed something. Human beings do tell stories about leaving, suffering, changing, and coming back, and they tell them a great deal. He then overclaimed, generalising a pattern he had assembled from a partial and unexamined sample into a law of the human spirit, and the overclaim was industrialised. A student can hold both facts. Holding both facts is what a literary education is for. 13. Limitations This study is limited in four ways that should be stated plainly. It is interpretive rather than empirical, and its conclusions are arguments, not measurements. It relies on English language sources, which is a genuine weakness in an article that criticises Western narrowness. It does not undertake primary work in comparative mythology, and therefore reports at second hand what particular myths contain. And it addresses a moving target: the critical conversation about Campbell is currently active, and the balance of scholarly opinion may look different in five years than it does now. 14. Recommendations for Future Research Several projects follow naturally from the gaps identified here. A properly sampled corpus study is overdue. With computational text analysis it is now feasible to take a defined, culturally diverse corpus of narratives, code them against Campbell's stages with explicit criteria and independent coders, and report how many actually fit. Campbell's claim has never been tested. It could be. Cross cultural replication of the psychological findings is needed. The restorying intervention should be tested in strongly collectivist settings, and, critically, against a rival intervention built on a communal or cyclical narrative shape. If a community centred restorying produces equal or greater meaning in those settings, the universality claim loses its last support. Classroom research would be valuable. Does teaching the monomyth alongside its critique produce better literary reasoning than teaching it alone? This is measurable and nobody has measured it. Work on generative systems is urgent. Researchers should test whether large language models default to the monomyth when asked for a story, whether they do so across languages, and whether this constitutes a mechanism for the further homogenisation of world #storytelling. Finally, the alternatives deserve development. The carrier bag, the spiral, the ensemble, the cycle, the trickster arc, and the relational models emerging in recent work (Arvas, 2025; Taylor, 2024) need the same careful, teachable articulation that Campbell gave his own model. Criticism has done its work. Construction is the next task. 15. Conclusion The Hero with a Thousand Faces is a flawed book of real importance, and both halves of that description need to be taken seriously. Campbell's achievement was to notice, across an enormous range of reading, that stories of departure, ordeal, and return recur with striking frequency, and to describe the pattern with a clarity that made it usable by other people. That description has helped readers see structure where they saw only incident, helped writers build stories that move audiences, and helped ordinary people find a shape for experiences that were otherwise shapeless. These are not small things. Campbell's error was to mistake a pattern he had assembled for a law he had discovered. He gathered fragments from cultures he did not study in their own terms, arranged them according to a psychological theory he did not test, and announced the result as the shared inheritance of humankind. The announcement was believed, and the pattern was then exported by an industry with unprecedented reach, until a particular way of telling a story came to look like the only way there is. The right response to this is neither to burn the book nor to teach it as scripture. It is to read it carefully, to use what it offers, and to keep asking the question it cannot answer for itself: what stories does this shape make it impossible for me to hear? A student who leaves this material able to explain the seventeen stages, able to explain why the framework has been called a Hollywood McMyth, and able to point to a story that will not fit inside it has learned something more valuable than either the model or its refutation. They have learned how to hold a powerful idea without being held by it. That, in the end, is its own kind of initiation. The framework calls; the critical reader crosses a threshold; the comfortable certainty of the diagram is lost; and what returns is a better reader, carrying something useful for everyone else. Hashtags #Joseph_Campbell_Studies #The_Hero_With_A_Thousand_Faces #Monomyth_Theory #Hero_Journey_Stages #Comparative_Mythology_Research #Archetypal_Criticism #Narrative_Structure #Myth_And_Literature #Literary_Theory_For_Students #Campbell_Criticism #Story_Structure_Analysis #World_Mythology #Narrative_Studies #Academic_Literature_Review #Humanities_Research References Ambasciano, L. (2021). The trials and tribulations of Luke Skywalker: How the Walt Disney Co. and Lucasfilm have failed to confront Joseph Campbell's troublesome legacy. Implicit Religion. Arvas, I. D. (2025). Reimagining the hero's journey through SDG lenses: Myth, ecology, and emotional literacy in The Three Girls Who Changed the World. Journal of Lifestyle and SDGs Review. Aston, J. (2024). There's more to life than the monomyth: Multiperspectival approaches to teaching narrative and story in university film and media departments. Media Practice and Education, 25(2), 123-136. doi:10.1080/25741136.2024.2331342 Bowman, S. L. (2024). Finding the self in role-playing games: Weaving myth, narrative, and identity. Media Practice and Education, 25(2), 160-173. doi:10.1080/25741136.2024.2324085 Campbell, J. (2008). The hero with a thousand faces (3rd ed.). New World Library. (Original work published 1949) Dewi, R. F., & Yusrina, D. A. (2023). Archetypal symbols and the hero's journey of Alice in Wonderland film. Surakarta English and Literature Journal. doi:10.52429/selju.v6i1.112 Gough-Brady, C. (2024). The heroic character, the neo-liberal productive citizen, and the feminist filmmaker. Media Practice and Education, 25(2), 149-159. doi:10.1080/25741136.2024.2324076 Groppel-Wegener, A. (2024). Immersive protagonists: Exploring the notion of the hero in theme parks. Media Practice and Education, 25(2), 137-148. doi:10.1080/25741136.2024.2324089 Hambly, G. (2021). The not so universal hero's journey. Journal of Screenwriting, 12(2), 135-152. Hanney, R. (2024). One myth to rule them all and in the darkness bind them: A critical examination of Joseph Campbell's The Hero's Journey. Media Practice and Education, 25(2), 113-122. doi:10.1080/25741136.2024.2333649 Jung, C. G. (1968). The archetypes and the collective unconscious (2nd ed., R. F. C. Hull, Trans.). Princeton University Press. Le Guin, U. K. (2019). The carrier bag theory of fiction. Ignota Books. Lefkowitz, M. R. (1990). The myth of Joseph Campbell. The American Scholar, 59(3), 429-434. Murdock, M. (1990). The heroine's journey: Woman's quest for wholeness. Shambhala. Nield, C. P. (2022). The hero's journey is Hollywood's McMyth. The Evolution of Story. Propp, V. (1968). Morphology of the folktale (2nd ed., L. Scott, Trans.). University of Texas Press. Remedios, D. J., Mathew, D. J., & Schleser, M. (2024). Navigating parallel interactive narratives in virtual reality. Media Practice and Education, 25(2), 188-201. doi:10.1080/25741136.2024.2324088 Rogers, B. A., Chicas, H., Kelly, J. M., Kubin, E., Christian, M. S., Kachanoff, F. J., Berger, J., Puryear, C., McAdams, D. P., & Gray, K. (2023). Seeing your life story as a hero's journey increases meaning in life. Journal of Personality and Social Psychology, 125(4), 752-778. doi:10.1037/pspa0000341 Segal, R. A. (1978). Joseph Campbell's theory of myth: An essay review of his oeuvre. Journal of the American Academy of Religion, 46(1), 67-80. Taylor, S. (2024). Not getting your story straight: Queering heroes' journeys and heteronormative timelines. Media Practice and Education, 25(2), 174-187. doi:10.1080/25741136.2024.2324120 van Gennep, A. (1960). The rites of passage (M. B. Vizedom & G. L. Caffee, Trans.). University of Chicago Press. Vogler, C. (2007). The writer's journey: Mythic structure for writers (3rd ed.). Michael Wiese Productions. Wu, M.-H., & Wang, Y.-H. (2022). Using mythic structure of Campbell's monomyth to analyze Spirited Away: A heroine's journey. Metathesis: Journal of English Language, Literature, and Teaching.

Forced displacement has reached the highest level ever recorded, with more than 120 million people living away from their homes because of conflict, persecution, disaster, and violence. Humanitarian and health systems have responded with a growing family of #trauma_informed approaches that promise to make services safer, more respectful, and less likely to repeat the harm that people have already survived. Yet a persistent tension sits at the centre of this work. Practitioners must meet urgent needs today, such as food, shelter, safety, and psychological first aid, while also confronting the legal, economic, and political structures that produce and prolong suffering. This article examines that tension. Using an integrative narrative review of peer reviewed research and grey literature published mainly between 2019 and 2025, it synthesises what is known about the mental health burden carried by #displaced_populations, the evidence base for scalable psychological and psychosocial interventions, and the critiques raised by cultural, ecological, and decolonial scholarship. The analysis identifies eight cross-cutting themes: the primacy of #daily_stressors alongside war exposure, the effectiveness and limits of brief scalable interventions, the neglected role of legal status and work rights, the tension between short funding cycles and long-term change, the demands of cultural adaptation and #community_participation, the wellbeing of the workforce, the specific needs of children and families, and the promise and risk of digital delivery. From these themes the article proposes a dual-track model in which relief and #systemic_advocacy are treated as parallel obligations rather than sequential stages. The model is supported by five propositions and a set of practical implications for practitioners, organisations, policy makers, and students. The article argues that trauma-informed practice which stops at the clinic door is incomplete, and that advocacy without competent, immediate care is hollow. Both are required, and they must be designed together from the beginning of a response rather than bolted together at the end. Keywords: forced displacement, trauma-informed care, refugee mental health, psychosocial support, structural advocacy, humanitarian practice, social determinants of health 1. Introduction The number of people living in forced displacement has grown almost every year for more than a decade. By 2024 the global figure had passed 120 million, a total that includes refugees, asylum seekers, internally displaced people, stateless persons, and others in refugee-like situations (UNHCR, 2024). These are not abstract numbers. They describe families sleeping in tents in the Levant, students learning in shipping containers in East Africa, mothers waiting years for an asylum decision in Europe, and men working without contracts in cities that will not recognise their qualifications. Displacement is not a single event. It is a long sequence of losses, waits, and negotiations that reshapes a person's relationship with time, place, and other people. Health and humanitarian systems have responded to this reality with a rapidly expanding architecture of mental health and psychosocial support, usually shortened to #MHPSS. Within that architecture, the language of trauma has become dominant. Agencies advertise #trauma_informed_care. Donors fund trauma programmes. Universities teach trauma modules. Job advertisements request trauma expertise. The spread of the concept has been remarkably fast, and much of it has been valuable. It has pushed services to ask what happened to a person rather than what is wrong with them, and it has made ordinary aspects of service design, such as waiting rooms, appointment letters, and interpreter arrangements, into matters of clinical importance. But rapid adoption has costs. Sweeney and Taggart (2018) warned that trauma-informed approaches risk being reduced to a slogan, adopted in name while leaving the underlying power relations of services untouched. In displacement settings this risk is sharper still. A service can be exquisitely gentle with a client for fifty minutes and then send that client back to a detention centre, an overcrowded shelter, or an immigration process that treats their account of persecution as a credibility test. The service has been kind. The system has not. If #trauma_informed_practice means anything, it must have something to say about the second of these facts as well as the first. This is the problem the present article addresses. The field has developed strong tools for #immediate_relief. Psychological first aid, brief scalable interventions delivered by non-specialists, group support, and crisis stabilisation all have a growing evidence base (Bangpan, Felix, & Dickson, 2019; Purgato et al., 2021; Tol et al., 2020). The field has also developed a strong critical literature on the structural determinants of distress, showing that ongoing stressors such as poverty, insecure legal status, discrimination, and family separation often predict mental health outcomes as strongly as past exposure to violence (Hou et al., 2020; Miller, Jordans, Tol, & Galappatti, 2021). What is less developed is a practical account of how a single organisation, team, or worker can hold both of these truths at once, in real programmes, with real budgets and real time pressure. The article therefore asks three questions. First, what does the recent evidence say about the mental health needs of #displaced_people and about the interventions designed to meet them? Second, where and why does the tension between relief and #structural_change appear in practice? Third, what would a model look like that treats the two as parallel commitments rather than competing priorities? The contribution is threefold. It offers a synthesis of recent literature written in accessible language for students and early-career practitioners. It names and analyses a tension that is frequently felt in field settings but rarely theorised in a usable way. And it proposes a dual-track framework, with supporting propositions and practical implications, that can be taught, applied, and tested. The article is organised as follows. Section 2 sets out conceptual foundations, defining displacement, trauma, and trauma-informed practice, and distinguishing trauma-informed from trauma-focused work. Section 3 reviews the empirical and critical literature. Section 4 presents the theoretical framework. Section 5 describes the review methodology. Section 6 presents eight thematic findings. Section 7 develops the dual-track model. Section 8 draws out practical implications. Sections 9 and 10 address limitations and future research, and Section 11 concludes. 2. Background and Conceptual Foundations 2.1 Who Counts as Displaced Displacement is a legal category before it is a psychological one, and the legal category shapes the psychology. A person recognised as a refugee under the 1951 Convention has rights to work, education, and residence that an asylum seeker awaiting a decision may not have. An internally displaced person who has never crossed a border may have citizenship but no protection, no compensation, and no realistic route home. A stateless person may have neither. These distinctions are not bureaucratic details. They determine whether a person can rent a flat, enrol a child in school, open a bank account, or report a crime without fear. They therefore determine much of what a mental health service can actually achieve. For the purposes of this article, #displaced_populations includes refugees, asylum seekers, internally displaced people, stateless persons, and people displaced by disaster and climate-related events. The category is deliberately broad because the psychological mechanisms discussed here, including loss, uncertainty, loss of control, and the erosion of social roles, cut across legal labels. Where the evidence is specific to one group, this is stated. It is worth stressing at the outset that most displaced people are hosted in low and middle income countries, often close to the country they left, and often for many years. The image of displacement as a short journey ending in a wealthy country is misleading. Protracted displacement, lasting five years or more, is the norm rather than the exception. Any model of care that assumes a rapid transition from crisis to settlement is therefore designed for a minority of cases. 2.2 Trauma: A Contested and Useful Word Trauma has at least three meanings in circulation, and confusing them causes trouble. The first is the #event meaning. Here trauma refers to an experience that overwhelms a person's capacity to cope, such as torture, sexual violence, bombardment, or witnessing a killing. This is the meaning embedded in diagnostic criteria for post-traumatic stress disorder. The second is the #response meaning. Here trauma refers to the lasting effects of such experiences on the nervous system, memory, emotion regulation, identity, and relationships. This is the meaning used when clinicians speak of the body remembering what the mind has tried to put away. The third is the #structural meaning, developed largely by anthropologists, critical psychiatrists, and community practitioners. Here trauma refers to patterned, ongoing harm produced by systems, such as war, occupation, racism, hostile immigration policy, or extreme poverty. Collective and continuous adversity is treated as a condition rather than an episode. Much of the confusion in the field comes from sliding between these meanings without noticing. A programme may be designed around the second meaning, treating trauma responses in individuals, while the population it serves is being harmed mainly in the third sense, by conditions that no individual therapy can undo. This article treats all three meanings as legitimate, and argues that a trauma-informed programme must be explicit about which one it is addressing at any given moment. It is also important to say what trauma is not. It is not a universal or inevitable outcome of terrible events. A substantial share of people exposed to atrocity do not develop a diagnosable disorder, and many find meaning, growth, or new purpose in the aftermath. Papadopoulos (2021) calls this #adversity_activated_development, insisting that clinicians look for what has been strengthened as well as what has been broken. Assuming universal damage is itself a form of disrespect, and can lead to services that offer therapy to people who need a lawyer. 2.3 From Trauma-Focused to Trauma-Informed Two related but distinct ideas are often confused. #Trauma_focused work is clinical treatment aimed directly at traumatic memories and their consequences. Examples include narrative exposure therapy, cognitive processing therapy, eye movement desensitisation and reprocessing, and trauma-focused cognitive behavioural therapy. These are specialist interventions, usually delivered by trained clinicians, usually to individuals or small groups, and usually over a defined number of sessions. They have a reasonable evidence base with refugee populations, though effect sizes vary and dropout can be high (Uphoff et al., 2020; Turrini et al., 2021). #Trauma_informed work is different. It is not a treatment. It is an organisational and relational stance that assumes trauma is common among the people a service meets, and redesigns the service so that it does not add to the harm. The widely used formulation from the United States Substance Abuse and Mental Health Services Administration sets out six principles: safety; trustworthiness and transparency; peer support; collaboration and mutuality; empowerment, voice, and choice; and attention to cultural, historical, and gender issues (SAMHSA, 2014). A systematic review by Han and colleagues (2021) found that trauma-informed interventions across many sectors most often target service culture, staff training, and screening practices, and that evaluation of their impact remains uneven. The practical implication is large. A food distribution point can be trauma-informed. So can a school, a housing office, a legal aid clinic, or a border reception centre. None of these settings delivers therapy. All of them can either respect or undermine a person's sense of safety, dignity, predictability, and control. This is why trauma-informed practice belongs to the whole system and not only to mental health professionals. Im, Rodriguez, and Grumbine (2021) argue for a multi-tier model in refugee resettlement in which trauma-informed and culture-informed principles operate at every level of a system of care, from front-line reception to policy. Their argument is central to this article. Once trauma-informed practice is understood as a property of systems, the question of #systemic_advocacy stops being an optional add-on and becomes a logical requirement. A system cannot be trauma-informed if it is also the source of the trauma. 3. Literature Review 3.1 The Scale of the Mental Health Burden The prevalence literature is now substantial. Blackmore and colleagues (2020), in a meta-analysis of refugees and asylum seekers, reported pooled prevalence estimates for post-traumatic stress disorder and depression that were many times higher than in general host populations, with considerable heterogeneity between studies. Henkelmann and colleagues (2020), focusing on refugees resettling in high income countries, reached broadly similar conclusions and noted that elevated rates persist for years after arrival, contradicting the assumption that distress resolves automatically once safety is reached. Charlson and colleagues (2019), in the study that produced the World Health Organization's revised estimates for conflict-affected populations, found that around one in five people in settings affected by conflict has a mental disorder at any point in time, with a smaller proportion experiencing severe conditions. Mesa-Vieira and colleagues (2022) extended this picture to migrants with pre-migration exposure to armed conflict, again finding high burdens that persist after migration. Two cautions belong alongside these numbers. First, estimates vary widely between studies, and methodological differences in sampling, instruments, and translation explain much of the variation. Second, prevalence is not need. Some people who screen positive on a symptom checklist do not want or require clinical treatment, and some who screen negative are in serious difficulty for reasons the checklist does not capture, such as destitution or family separation. #Screening tells us something, but it cannot substitute for asking people what they want. 3.2 Daily Stressors and the Ecological Turn Perhaps the most important conceptual development of the last two decades is the recognition that war exposure is not the only, or even always the main, driver of distress in displaced populations. Miller and Rasmussen's ecological model argued that #daily_stressors such as poverty, unemployment, poor housing, discrimination, insecurity, family conflict, and lack of legal status mediate much of the relationship between war exposure and mental health outcomes. The empirical support for this position has strengthened. Hou and colleagues (2020), in a meta-analysis of everyday life experiences among conflict-affected forced migrants, found robust associations between daily living conditions and mental health, comparable in magnitude to associations with war exposure itself. Miller, Jordans, Tol, and Galappatti (2021) subsequently called for greater conceptual clarity in the MHPSS field, warning that the term is used so loosely that it obscures whether a programme is trying to treat disorder, prevent it, promote wellbeing, or change the conditions that produce distress. Their argument is a useful discipline. A programme that reduces #daily_stressors, for instance by providing legal representation or safe housing, may improve mental health more than a programme that offers counselling in a setting where the stressors continue unchanged. Nickerson and colleagues (2019) provided a particularly clear demonstration in a study of refugees in Australia, showing that visa insecurity was independently associated with poorer mental health, greater disability, and reduced social engagement. Legal status, in other words, is a mental health variable. This finding has direct implications for what counts as a #mental_health_intervention. Scharpf, Kaltenbach, Nickerson, and Hecker (2021) reviewed socio-ecological factors for refugee children and adolescents and found a similar pattern across levels: individual, family, community, and societal factors all contributed, with caregiver mental health and school environment emerging as especially influential. Bürgin and colleagues (2022), writing in response to the war in Ukraine, argued explicitly for multilevel, needs-oriented, and trauma-informed approaches for children affected by war and displacement. 3.3 What Interventions Have Been Tested The intervention literature can be organised roughly by the layers of the MHPSS pyramid, which distinguishes basic services and security, community and family supports, focused non-specialised supports, and specialised services. At the level of #focused_non_specialised_support, the most significant development has been the emergence of brief, manualised, transdiagnostic interventions that can be delivered by trained non-specialists. Problem Management Plus, developed by the World Health Organization, has been tested in several displacement-affected settings. Bryant and colleagues (2022) reported twelve-month follow-up data from a group version delivered to Syrian refugees in Jordan, finding benefits that were maintained beyond the immediate post-intervention period. Self-Help Plus, a guided self-help course delivered in large groups with audio recordings and a book, was tested by Purgato and colleagues (2021) among refugees and asylum seekers in Western Europe, with results supporting its role in preventing the onset of mental disorders. Tol and colleagues (2020) evaluated guided self-help among South Sudanese women in Uganda in a cluster randomised trial, again reporting reductions in psychological distress. The strength of these interventions is #scalability. They are short, they do not require a psychologist, and they can be adapted across languages and cultures with structured processes. The limits are equally important. Effects are typically modest, they attenuate over time in some studies, and trials rarely capture whether people's material circumstances improved. At the level of #specialised_services, Uphoff and colleagues (2020) conducted an overview of systematic reviews of mental health promotion, prevention, and treatment for refugees, asylum seekers, and internally displaced people. They found evidence for the effectiveness of psychological therapies for common mental disorders, but noted the low certainty of much of the evidence, small sample sizes, and limited data from low income settings. Tribe, Sendt, and Tracy (2019) similarly found supportive but heterogeneous evidence for psychosocial interventions with adult refugees and asylum seekers, and highlighted the near absence of long-term follow-up. At the level of #community_and_family_supports, family-based approaches have gained attention. Betancourt and colleagues (2020) reported feasibility and acceptability results for a family strengthening intervention with Somali Bantu and Bhutanese refugee families in the United States, developed through sustained collaboration with the communities concerned. Sangalang and Vang (2017) reviewed intergenerational trauma in refugee families and argued that individual-level treatment misses the transmission of distress across generations through parenting, silence, and disrupted attachment. At the level of #basic_services_and_security, the evidence is thinner in the mental health literature, largely because these interventions are rarely evaluated with mental health outcomes. An important exception is the work of Hussam, Kelley, Lane, and Zahra (2022), who found in a camp setting that the opportunity to work carried psychosocial value beyond its income, improving wellbeing even when the financial return was controlled. This finding is consistent with a broader literature on the psychological importance of role, purpose, and contribution, and it is difficult to reconcile with policies that prohibit refugees from working. 3.4 Help-Seeking, Access, and Stigma Availability is not access. Byrow, Pajak, Specker, and Nickerson (2020) systematically reviewed perceptions of mental health and barriers to help-seeking among refugees. Their findings included stigma, lack of knowledge about available services, language barriers, mistrust of authorities, fear of consequences for immigration claims, and cultural understandings of distress that do not map onto psychiatric categories. Practical barriers such as transport, cost, childcare, and working hours compounded these. Hajak, Sardana, Verdeli, and Grimm (2021) reviewed factors affecting the mental health and wellbeing of asylum seekers and refugees in Germany, finding that the asylum procedure itself, accommodation conditions, and restrictions on work and movement featured prominently in the accounts of both service users and providers. Kohrt and colleagues (2020) have argued more broadly that healing practices are shaped by social context and cultural expectation, and that global mental health must attend to why particular practices carry credibility in particular places. Taken together, this literature suggests that a trauma-informed service which is technically excellent but institutionally intimidating, linguistically inaccessible, or entangled with immigration enforcement will not be used by the people who need it most. 3.5 Critical Perspectives A vigorous critical literature challenges the assumptions of mainstream trauma work in humanitarian settings. The first line of critique is the risk of #medicalisation. Critics argue that framing collective political suffering as individual psychiatric disorder depoliticises it, turning survivors of state violence into patients and their claims for justice into symptoms. The concern is not that suffering is unreal, but that the diagnostic frame can obscure its causes. The second is #cultural_validity. Categories such as post-traumatic stress disorder were developed in specific settings and may not correspond to how distress is experienced or expressed elsewhere. Tay and colleagues (2019), reviewing the culture and mental health of Rohingya refugees, described idioms of distress and explanatory models that differ substantially from standard psychiatric constructs, with direct implications for assessment and intervention. Bemme and Kirmayer (2020) situate this within the wider challenges of global mental health as an interdisciplinary field in motion. The third is #evidence_hierarchy. Kienzler (2019) argues for practice-based evidence in humanitarian mental health reform, noting that the randomised controlled trial, while valuable, cannot capture much of what determines whether a programme works in a chaotic, resource-scarce, politically contested environment. The fourth is the critique of trauma-informed care itself. Sweeney and Taggart (2018) observe that the approach is frequently misunderstood, sometimes reduced to a training day, sometimes conflated with trauma-focused therapy, and rarely accompanied by the shifts in power that its own principles demand. Sweeney, Filson, Kennedy, Collinson, and Gillard (2018) argue that the core of the approach is a change in relationships within services, including relationships between staff and the people they serve and among staff themselves. These critiques do not invalidate trauma-informed practice. They discipline it. They insist that #power, #culture, and #politics are not background conditions but constitutive elements of the work. 3.6 The Advocacy Literature The literature on advocacy in displacement contexts is more scattered, spread across refugee studies, public health, social work, and law. Phillimore (2021) proposes the concept of refugee-integration-opportunity structures, shifting analytical attention from the characteristics of refugees to the characteristics of the contexts that receive them. Integration, on this account, is something a society does or fails to do, not merely something a refugee achieves. Strang and Quinn (2021) examine social connection and wellbeing, showing that relationships across social groups, including relationships with members of the receiving society and with institutions, are central to wellbeing and are strongly shaped by policy. Jordans and Kohrt (2020) set out a roadmap for scaling mental health care in low-resource settings, arguing that scale-up requires attention to health systems, financing, and workforce rather than intervention design alone. Their argument, though not framed as advocacy, points to the same conclusion: sustainable improvement depends on changing systems. What remains largely missing from the literature is empirical evaluation of #advocacy as an intervention. We have strong evidence that legal status affects mental health. We have very little evidence about what happens to mental health when an organisation successfully campaigns to change legal status rules. This is a serious gap, and it is discussed further in Section 10. 4. Theoretical Framework This article draws on four bodies of theory, which together provide the scaffolding for the dual-track model developed in Section 7. 4.1 The Socio-Ecological Model The socio-ecological model treats human development as occurring within nested systems: the individual, the immediate relationships of family and peers, the institutions of school, workplace, and clinic, the wider community, and the policies and cultural norms that surround them all. Applied to displacement, the model makes a simple but powerful claim. Distress produced at one level cannot be fully repaired at another. A child whose distress is generated by an unsafe camp, an absent parent, and an uncertain future cannot be restored to wellbeing by a weekly session that leaves those three conditions intact. Scharpf and colleagues (2021) demonstrate the empirical utility of this framing, showing that risk and protective factors for displaced children and adolescents operate at every level and interact with each other. The practical consequence is that an intervention should be designed at the level at which the main driver of harm is operating. This sounds obvious. It is routinely ignored, because funding and professional identity tend to anchor practitioners at the individual level. 4.2 The Social Determinants of Mental Health The social determinants perspective holds that the conditions in which people are born, grow, live, work, and age are the primary drivers of health inequality. In the displacement context the relevant determinants include legal status, the right to work, income, housing quality, food security, education access, family reunification, exposure to discrimination, and the experience of the asylum process itself. Nickerson and colleagues (2019) provide a direct example: visa insecurity functions as a determinant of mental health, independent of prior trauma exposure. Hussam and colleagues (2022) provide another: access to employment produces psychosocial benefit beyond income. The social determinants perspective supplies the empirical warrant for #advocacy. If the determinants are the drivers, then acting on the determinants is clinical work, not a diversion from it. 4.3 Structural Competency Structural competency extends cultural competency by asking practitioners to recognise how policies, institutions, and economic arrangements shape the problems that appear in the consulting room. Where cultural competency asks a clinician to understand a patient's beliefs, structural competency asks the clinician to understand the housing policy that put the patient in a damp flat, and to consider whether the appropriate response is a prescription, a letter to the housing authority, or both. For displacement work, structural competency reframes several familiar clinical dilemmas. A client who misses appointments may not be avoidant; they may be attending immigration reporting sessions that are scheduled without regard for anything else in their life. A client who cannot sleep may not need a sedative; they may need the noise, overcrowding, and fear in their accommodation to end. Structural competency does not replace clinical skill. It prevents clinical skill from being wasted on problems it cannot solve. 4.4 The Intervention Pyramid and Its Limits The MHPSS intervention pyramid organises support into four layers: basic services and security, community and family supports, focused non-specialised supports, and specialised services. It has been enormously useful. It reminds agencies that most people do not need therapy, that the base of the pyramid serves the largest number, and that a functioning response needs all layers rather than only the most clinical one. Its limits are also instructive. The pyramid describes what services should exist, but it says little about who has the power to decide, how the layers are financed, or what to do when the base of the pyramid is being actively undermined by government policy. Miller and colleagues (2021) note that the vagueness of the MHPSS label allows very different activities to be grouped together without clear theories of change. This article accepts the pyramid as a description of service architecture, while arguing that it needs a fifth dimension running vertically through all four layers: the deliberate effort to change the conditions that determine what each layer can achieve. That vertical dimension is #systemic_advocacy. 4.5 Synthesis: Two Obligations, One Practice Putting these theories together yields the core theoretical claim of this article. Trauma-informed practice with displaced populations carries two obligations that are conceptually distinct but practically inseparable. The first is the obligation of #proximity. It requires being present, competent, and reliable in the immediate encounter, offering safety, respect, information, and skilled help now, without waiting for structural conditions to improve. The second is the obligation of #structural_honesty. It requires naming, documenting, and acting on the conditions that produce and prolong distress, even when those conditions are created by the same authorities that fund or permit the service. Practitioners who honour only the first become highly skilled at helping people adapt to intolerable conditions. Practitioners who honour only the second become highly principled and unhelpful. The task is to hold both, and to build organisations in which holding both is normal rather than heroic. 5. Methodology 5.1 Design This article is an integrative narrative review with thematic synthesis. This design was selected because the research questions are conceptual as well as empirical, and because the relevant evidence is drawn from several disciplines that use different methods and vocabularies. A systematic review with a narrow clinical question would exclude much of the material that matters here, including qualitative studies, policy analysis, and critical scholarship. A narrative approach permits the integration of these sources, at the cost of reduced replicability, which is acknowledged in Section 9. 5.2 Literature Search Searches were conducted in electronic databases covering health, psychology, and social science literature, together with the publication repositories of major humanitarian and health agencies. Search terms combined three concept clusters: Population terms: refugee, asylum seeker, internally displaced, forced migration, displacement, stateless. Intervention terms: trauma-informed, mental health and psychosocial support, psychological first aid, psychosocial intervention, scalable intervention, task-sharing. Systemic terms: advocacy, social determinants, structural, policy, legal status, integration, rights. Reference lists of key papers were hand-searched, and forward citation tracking was used for highly cited recent reviews. 5.3 Inclusion and Exclusion Sources were included if they were published in English, addressed displaced populations or trauma-informed practice in a way relevant to the research questions, and were published from 2019 onwards. A small number of older sources were retained where they are foundational and have not been superseded, for example the original articulation of trauma-informed principles. Priority was given to systematic reviews, meta-analyses, randomised trials, and well-designed qualitative studies. Commentaries and conceptual papers were included where they contribute to the theoretical argument. Sources were excluded if they addressed non-displaced clinical populations only, if they were purely descriptive programme reports without analysis, or if their methods were too poorly described to permit appraisal. 5.4 Synthesis The synthesis followed a three-stage thematic process. First, included sources were read and coded for content relating to needs, interventions, barriers, and structural factors. Second, codes were grouped into descriptive themes. Third, descriptive themes were interpreted into the eight analytical themes presented in Section 6, and then into the model presented in Section 7. Discrepant findings were retained rather than smoothed over, and are reported where they occur. 5.5 Positionality and Reflexivity Reviews of this kind are shaped by the perspective of the person conducting them. This article is written from within a broadly critical, rights-based tradition that regards structural conditions as central to mental health. Readers who begin from a strictly clinical or strictly biomedical position will find some of the emphases here contestable, and the article tries to represent those positions fairly in Sections 3.5 and 7.4. Students using this article are encouraged to read the primary sources rather than relying on the summary given here, and to form their own view. 6. Findings: Eight Themes 6.1 Theme One: Distress Is Produced in the Present, Not Only in the Past The single most consistent finding across the recent literature is that #present_conditions matter enormously. Hou and colleagues (2020) show that everyday life experiences among conflict-affected forced migrants are strongly associated with mental health. Hajak and colleagues (2021) show that the asylum procedure, accommodation, and restrictions on work and movement are prominent in accounts of distress. Nickerson and colleagues (2019) show that visa insecurity independently predicts poorer outcomes. The implication for practice is uncomfortable for services organised around treating the past. If a person is distressed because they are destitute, unsafe, and facing deportation, then the most effective mental health intervention available may be a competent lawyer, a safe room, and a work permit. Offering therapy instead is not neutral. It communicates that the problem is inside the person. This does not mean therapy is useless. It means that the sequencing and framing of therapy matter. Where present conditions are the dominant driver, therapeutic work is best positioned as support for coping and functioning while the conditions are addressed, not as a treatment for a condition that the person is imagined to carry independently of their circumstances. A useful clinical habit follows: in every assessment, ask explicitly what is happening now, not only what happened then. Ask about money, housing, documents, safety, family separation, and hostility. Record the answers. Those records become the raw material of advocacy. 6.2 Theme Two: Brief Scalable Interventions Work, Within Limits The evidence for brief interventions delivered by trained non-specialists is now reasonably strong. Problem Management Plus and Self-Help Plus have both been tested in displacement-affected populations, with effects that are statistically significant and clinically meaningful for many participants (Bryant et al., 2022; Purgato et al., 2021; Tol et al., 2020). Overviews of the wider evidence base support the use of psychological interventions while noting the modest certainty of the evidence (Uphoff et al., 2020; Bangpan et al., 2019). Three limits deserve emphasis. First, effect sizes are typically moderate at best, and a substantial proportion of participants do not improve. These are not cures. They are useful supports. Second, #maintenance of effect is uncertain. Some trials show benefits persisting at follow-up; others show attenuation. Where the environment remains hostile, the loss of benefit over time is unsurprising and should not be attributed to a failure of the participant. Third, trials rarely measure what happens to material circumstances. A person may report fewer symptoms while remaining homeless. This is a real gain and should not be dismissed, but it should not be mistaken for resolution. The correct conclusion is not to abandon these interventions. It is to deploy them honestly, as the useful, cheap, scalable, and partial tools that they are, and to build the advocacy track alongside them rather than instead of them. 6.3 Theme Three: The Workforce Is Part of the Intervention Trauma-informed practice cannot survive in an organisation that treats its own staff badly. This claim is not rhetorical; it follows directly from the principles. A worker who is exhausted, unsupported, precariously employed, and afraid of their manager cannot reliably offer safety, transparency, collaboration, and choice to anyone else. Sweeney and colleagues (2018) place relationships within services at the centre of the trauma-informed paradigm, and this necessarily includes relationships among staff. Three features of the displacement workforce make this especially pressing. First, a large proportion of front-line workers are themselves from displaced or affected communities. They are often paid less, employed on shorter contracts, and given less supervision than international staff, while carrying the heaviest emotional load and the greatest exposure to the same conditions their clients face. #Vicarious_trauma is not the right term when the trauma is not vicarious at all. Second, task-sharing models, which are essential to scale, depend on non-specialist workers receiving continuing supervision, not just initial training. Where supervision collapses, fidelity collapses, and the intervention becomes a set of forms. Third, #moral_injury is common. Workers routinely see harm they are not permitted to prevent, and are asked to deliver services within systems whose rules they consider unjust. The distress this causes is not a personal weakness. It is an accurate response to an impossible position, and organisations that treat it as burnout to be managed with wellness sessions are misdiagnosing their own problem. Practical responses include structured group supervision, fair and equal contracts for national and displaced staff, genuine involvement of front-line staff in programme design, and formal channels through which staff can escalate structural concerns to the people with the power to raise them externally. 6.4 Theme Four: Culture Is Not a Decoration The literature is emphatic that intervention content must be adapted, and equally emphatic that adaptation is often done badly. Adaptation is frequently reduced to translation plus the substitution of local names and foods in vignettes. This is necessary and nowhere near sufficient. Tay and colleagues (2019) show that Rohingya idioms of distress and explanatory models differ substantially from psychiatric constructs, with implications for how symptoms are described, whom people consult, and what counts as recovery. Byrow and colleagues (2020) show that cultural understandings of mental health shape help-seeking directly. Kohrt and colleagues (2020) argue that the credibility of healing practices is itself culturally produced. Real #cultural_adaptation involves several steps: understanding local idioms of distress and using them in materials; identifying who is already trusted as a helper, including religious leaders, elders, traditional healers, and community mothers' groups; testing whether the theory of change makes sense to participants, not only whether the words translate; and being willing to change the intervention rather than only its packaging. There is a deeper point. The demand for adaptation is usually framed as a technical problem of fidelity versus fit. It is also a question of #authority. Who decides what a good outcome is? If the answer is always the agency, the instrument, and the donor, then the adaptation is cosmetic no matter how careful the translation. 6.5 Theme Five: Participation Is the Hardest Principle to Keep Every framework says that displaced people should participate in the design of the services intended for them. In practice, participation is often the first principle abandoned when timelines tighten. Consultation is not participation. A meeting in which a plan is presented and comments are invited, after the budget is fixed and the logframe approved, is a courtesy, not a transfer of power. Genuine #participation means that displaced people are involved when the questions are still open: what problem are we addressing, what would count as success, who should deliver the service, and what should we do if the main obstacle turns out to be policy? Betancourt and colleagues (2020) describe an intervention developed through long-term collaboration with the communities concerned, and show that this process changed the content of the intervention itself, not just its acceptability. This is the standard to aim at, and it takes years. Several structural obstacles stand in the way. Funding cycles are short. Agencies are accountable to donors rather than to service users. Participation is expensive in staff time and produces messy answers. And there is a genuine ethical tension: asking people who are exhausted, unpaid, and struggling to survive to donate their time to programme design can itself be extractive. #Payment for participation, childcare, transport, and interpretation are not luxuries. They are the minimum conditions of honest engagement. 6.6 Theme Six: Children, Families, and the Generational Question Children make up a large share of the displaced, and their needs cannot be met by treating them as small adults. Scharpf and colleagues (2021) show that caregiver mental health and school environment are among the strongest modifiable influences on displaced children's outcomes. Bürgin and colleagues (2022) argue for multilevel approaches that combine family support, school-based intervention, and policy attention. Sangalang and Vang (2017) document intergenerational transmission of distress through parenting, silence, and altered family roles. Three practical conclusions follow. First, supporting caregivers is a #child_intervention. A parent who is not sleeping, not eating, and terrified of deportation cannot provide the regulation and predictability a child needs. Programmes that treat children in isolation from their caregivers are working against the strongest lever available to them. Second, #school is often the single most powerful psychosocial intervention available. It provides routine, peer relationships, adult attention, language acquisition, and a future orientation. Barriers to school enrolment, including documentation requirements, fees, uniforms, transport, and hostility, are therefore mental health barriers, and removing them is mental health work. Third, role reversal is common and consequential. Children who interpret for their parents, negotiate with landlords, and mediate with authorities carry adult burdens without adult power. Services should notice this, avoid relying on children as interpreters, and provide professional interpretation as standard. 6.7 Theme Seven: Digital Delivery Offers Reach and New Risks Remote and digital delivery expanded rapidly during the coronavirus pandemic and has not retreated. Phone-based counselling, messaging support, self-help applications, and video sessions now feature in many programmes. The advantages are real. Reach extends to people who cannot travel. Privacy can be greater for those who fear being seen entering a mental health service. Costs may fall. Scale becomes possible in settings with few specialists. The risks are equally real and are often underestimated. #Digital_exclusion is patterned. Access to a phone, a private space, data, and electricity is not evenly distributed, and the people with least access are frequently those with the greatest need, including women in controlling households, older people, and those in the most precarious accommodation. #Data_protection is a protection issue, not merely a compliance issue. Information about a person's asylum claim, sexual orientation, political activity, or family location can endanger them or their relatives. A leaked case record is not an administrative embarrassment. It can be lethal. Finally, digital tools can become a way to appear responsive without changing anything material. Sending a wellbeing application to a person living in an unheated shelter is not care. It is a substitution of interface for substance. 6.8 Theme Eight: The Structural Track Is Under-Evaluated and Under-Funded The final theme is the gap that motivates this article. The evidence that structural conditions drive mental health outcomes is strong. The evidence about what happens when organisations try to change those conditions is very weak. There are structural reasons for this. Advocacy outcomes are slow, diffuse, and hard to attribute. Donors fund countable outputs. Governments that host displaced populations may permit service delivery while restricting political activity, and organisations that campaign too loudly can lose their registration, their access, or their staff visas. #Operational_risk is not imaginary, and dismissing it as cowardice is unfair to people working in genuinely dangerous environments. Nevertheless, the current balance is indefensible. Phillimore (2021) argues that opportunity structures in receiving societies determine integration outcomes. Strang and Quinn (2021) show that social connection, itself shaped by policy, is central to wellbeing. Jordans and Kohrt (2020) show that scale-up depends on systems and financing rather than intervention design. If all of this is true, then a sector that spends almost all of its resources on service delivery and almost none on changing the systems that determine the value of that delivery is optimising the wrong variable. The rest of this article is an attempt to describe what a better balance would look like. 7. Discussion: A Dual-Track Model of Trauma-Informed Practice 7.1 The Core Proposal The findings above point toward a single organising idea. Relief and advocacy should not be arranged as a sequence, in which agencies deliver services now and hope to influence policy later. They should be arranged as two parallel tracks, planned together, resourced together, staffed together, and evaluated together, from the first day of a programme. Track One is #relief. It answers the question: what does this person need today, and can we provide it competently, safely, and respectfully? Its activities include safety and shelter, food and cash, psychological first aid, brief scalable psychological interventions, family support, school access, and specialist clinical care for those who need it. Track Two is #advocacy. It answers the question: what conditions are producing and prolonging this suffering, and what are we doing about them? Its activities include documentation, evidence generation, legal support and strategic litigation, coalition building with affected communities, policy engagement, media work where safe, and internal reform of the organisation's own practices. The two tracks are joined by a connective tissue that is easy to overlook: the ordinary clinical and casework record. Every intake form that records destitution, every note that records a client unable to attend because of an immigration reporting requirement, every observation of a child excluded from school for lack of documents, is a data point. Aggregated and anonymised, these data points become the evidence base for Track Two. This is why #documentation is described here as a bridge rather than as bureaucracy. It converts individual suffering into structural argument without requiring any individual to expose themselves. 7.2 Five Propositions The model can be stated more precisely as five propositions, each of which is empirically testable. Proposition 1: Relief without advocacy produces diminishing returns. Where structural conditions continue to generate distress, the effects of individual-level intervention will attenuate over time, and services will encounter the same people repeatedly with the same problems. Predicted observation: high rates of re-presentation and poor maintenance of treatment gains in settings with hostile policy environments, compared with similar interventions in permissive environments. Proposition 2: Advocacy without relief loses legitimacy and access. Organisations that campaign but do not deliver lose the trust of affected communities, whose immediate needs are urgent, and lose the operational access that generates credible evidence. Predicted observation: advocacy organisations with no service arm produce less granular, less persuasive evidence and have weaker mandates from affected communities. Proposition 3: Documentation is the cheapest available advocacy input. Systematically recorded, anonymised, aggregated service data has high evidential value at low marginal cost. Predicted observation: organisations that invest modestly in data infrastructure produce disproportionately more usable advocacy evidence than those that invest in advocacy staff alone. Proposition 4: Trauma-informed principles applied internally predict trauma-informed practice externally. Organisations that offer their own staff safety, transparency, voice, and fair treatment will deliver more consistent trauma-informed care. Predicted observation: staff-reported organisational climate correlates with service-user-reported experience of care. Proposition 5: Participation changes what gets advocated for. Where displaced people set the agenda, the priorities that emerge will differ from those set by agencies, and will more often concern legal status, work, and family reunification than clinical services. Predicted observation: participatory needs assessments rank structural issues higher than professionally led ones. 7.3 What the Model Demands of Organisations Adopting the dual-track model has concrete organisational implications. #Budgeting must change. If advocacy is a track rather than an afterthought, it needs a line, a percentage, and a person. A useful discipline is to set a minimum proportion of programme budget devoted to structural work and to defend it when cuts come, on the grounds that cutting it is equivalent to cutting the effectiveness of everything else. #Measurement must change. Symptom scales alone cannot capture the outcomes that matter. A dual-track programme should measure functioning, social connection, and material circumstances alongside symptoms, and should track structural indicators such as the proportion of clients with resolved legal status, in secure housing, in school, or in lawful work. #Governance must change. Displaced people should sit on the bodies that set priorities, with real votes, paid time, and interpretation, not on advisory panels whose recommendations are noted and ignored. #Risk_management must change. Advocacy carries risk, including to registration, funding, and staff. The response is not to avoid advocacy but to manage it deliberately: choosing which battles to fight publicly and which to fight through coalitions, protecting field staff by locating public statements at headquarters level, and being honest with communities about what the organisation can and cannot say. 7.4 Objections and Responses Several objections deserve a fair hearing. The first is the objection from #neutrality. Humanitarian actors often argue that political engagement compromises the neutrality that grants them access to populations in need. This is a serious argument and cannot be waved away. The response is twofold. First, neutrality about parties to a conflict is not the same as silence about conditions in a reception centre; the two are frequently conflated to justify silence that serves institutional convenience. Second, neutrality has an evidential cost, and organisations should at minimum be honest with themselves about the trade they are making rather than describing it as a principle when it is a calculation. The second is the objection from #competence. Clinicians may reasonably say that they are trained to treat, not to campaign, and that amateur advocacy can do harm. This is correct, and the model does not require every clinician to become a lobbyist. It requires the organisation to have the capacity, and it requires the clinician to record what they see and pass it on. Division of labour is legitimate; ignorance is not. The third is the objection from #evidence. Advocacy outcomes are hard to measure, and scarce resources should go to interventions with demonstrated effect. This objection has real force, and it is precisely why Section 10 calls for advocacy to be evaluated rather than exempted. But note the asymmetry: we routinely fund clinical interventions whose effects are modest and short-lived, while demanding proof of effect from structural work before it is funded at all. The standard should be the same for both. The fourth is the objection from #urgency. When people are hungry today, structural work can seem like a luxury. The response is that the dual-track model does not propose delaying relief. It proposes not stopping there. The child fed today will be hungry tomorrow if the ration is cut, and the ration is a policy. 7.5 Ethics Three ethical commitments run through the model. #Do_no_harm requires attention to the harm that services themselves can cause: repeated retelling of traumatic accounts for assessments that produce no benefit, data practices that endanger, raised expectations that are not met, and the recruitment of people into research from which they gain nothing. #Informed_consent in displacement settings is complicated by dependency. A person whose food assistance comes from an agency cannot easily refuse that agency's research request. Separating service provision from research participation, and making the separation explicit and credible, is a minimum requirement. #Reciprocity requires that communities receive something back. Findings should be returned to communities in accessible form and in their own languages, and community members should be paid for their labour. 8. Practical Implications 8.1 For Front-Line Practitioners Ask about the present. Every assessment should include current housing, income, documents, safety, and family separation, recorded systematically. Explain everything. Uncertainty is a wound in displacement contexts. Say what will happen, when, who will see the information, and what the limits are. If you do not know, say that you do not know. Offer choice wherever it exists, including small choices about where to sit, when to meet, whether to use an interpreter of a particular gender, and what to talk about first. Small choices restore a sense of agency that has been systematically removed. Use professional interpreters. Never use children. Brief the interpreter beforehand and debrief afterwards; interpreters carry the material too. Do not extract stories you do not need. Asking a person to narrate torture in order to complete a form is harm without benefit. Write things down in a way that can be used. Aggregated case notes are the raw material of change. Know the referral map, including legal aid, housing, and education, and treat referral as clinical work. 8.2 For Programme Managers Design both tracks at proposal stage. Retrofitting advocacy is much harder than building it in. Fund supervision, not just training. A trained non-specialist without supervision drifts within months. Pay national and displaced staff fairly and give them contracts of the same length and security as international staff wherever legally possible. Build data systems that can produce anonymised, aggregate evidence of structural harm, and agree in advance who is permitted to use it and how. Plan for #protracted timelines. Most displacement lasts years. Programme cycles of twelve months guarantee that the intervention will end before the problem does. 8.3 For Policy Makers and Donors Recognise that legal status, the right to work, family reunification, and housing quality are mental health policies, whether or not they are labelled as such. Fund longer. Three-year and five-year cycles are not a generosity; they are a condition of effectiveness. Fund the base of the pyramid and the vertical track. Cash, shelter, school access, and legal aid are among the most cost-effective mental health interventions available, and are chronically underfunded relative to their impact. Require but do not fetishise evidence. Demand a theory of change, honest measurement, and learning from failure, rather than only randomised trials that few displacement programmes can afford or ethically conduct. Protect the space for civil society to speak. Restrictions on advocacy are, in effect, restrictions on health. 8.4 For Students and Educators Students entering this field should take three lessons from the literature reviewed here. First, resist the temptation to see suffering only as pathology. Many of the people you meet will be reacting normally to abnormal conditions, and will need documents, money, and safety more than they need a diagnosis. Second, resist the opposite temptation to dismiss clinical work as a distraction. People with severe conditions exist in every population, and they are frequently abandoned in humanitarian settings because everyone is busy debating structures. #Severe_mental_illness in displacement contexts is one of the most neglected areas in global health. Third, learn to write. A great deal of structural change is achieved by people who can turn a pile of case notes into a clear, accurate, and undeniable document. 9. Limitations Several limitations should be borne in mind. The review is narrative rather than systematic. Sources were selected by judgement, and a different reviewer could reasonably have selected differently. Publication bias, language bias toward English-language sources, and the dominance of research conducted in high income settings all shape the picture presented here. The evidence base itself is uneven. Trials are concentrated in a small number of settings and populations. Long-term follow-up is rare. Outcome measures are dominated by symptom scales that may not capture what participants value. The dual-track model is a conceptual proposal, not a validated framework. Its propositions are stated so that they can be tested, but they have not yet been tested here. Finally, the article is written from a particular normative position, described in Section 5.5. Readers should treat its conclusions as arguments to be examined rather than findings to be accepted. 10. Directions for Future Research Five priorities emerge. Evaluate advocacy. The most striking gap in the literature is the absence of evaluation of structural interventions using mental health outcomes. Natural experiments offer a feasible route: when a jurisdiction grants work rights, extends visa duration, or ends a form of detention, mental health outcomes in affected populations can be compared before and after, and against comparison groups. #Natural_experiments of this kind would be among the most policy-relevant studies the field could produce. Measure what matters. Instruments are needed that capture functioning, social connection, dignity, and material security, developed with displaced people rather than imposed on them, and validated across languages and settings. Study the long term. Follow-up beyond twelve months is rare. Given that displacement typically lasts years, the field needs cohort studies that follow people through changes in legal status, housing, and employment. Study the workforce. Systematic research is needed on the wellbeing, retention, and effectiveness of displaced and national staff, who deliver most of the world's frontline psychosocial support and are studied far less than the interventions they deliver. Test participation. The claim that participatory design improves outcomes is widely asserted and rarely tested. Comparative studies of participatory and conventional design in comparable settings would be valuable, including honest reporting of cases where participation slowed things down or produced conflict. 11. Conclusion Trauma-informed practice with displaced populations has come a long way in a short time. It has changed how services speak to people, how they design their spaces, how they train their staff, and how they understand behaviour that once looked like non-compliance. These are real gains and they should be defended. But the approach is incomplete, and the incompleteness is not accidental. It is easier for an organisation to soften its tone than to challenge the policies that shape its clients' lives. It is easier to fund a counselling post than a legal one. It is easier to measure a symptom score than a change in the law. The path of least resistance leads to services that are kind, well-meaning, and structurally silent. The argument of this article is that such silence is a departure from trauma-informed principles rather than a neutral position. If safety is a principle, then a service that returns people to unsafe conditions without protest has not honoured it. If empowerment, voice, and choice are principles, then a system that denies people the right to work, to reunite with their families, or to know when their case will be decided is a system that a trauma-informed practitioner is obliged to name. The dual-track model proposed here is an attempt to make that obligation practical rather than rhetorical. It insists on competent, immediate, respectful help, delivered now, without waiting for justice. It insists equally on documenting, aggregating, and acting on the conditions that make that help necessary. It treats the case note and the policy submission as parts of the same task. None of this is easy, and none of it is safe. Organisations that speak lose access. Practitioners who challenge lose jobs. Communities that organise face reprisal. These risks are real and must be managed with care rather than denied with bravado. But the alternative is worse. A field that becomes very good at helping people endure conditions it declines to challenge has not solved a problem. It has only made the problem more comfortable to look at. #Displaced_populations do not need the world's sympathy administered in fifty-minute sessions. They need safety, status, work, family, school, and a future, and they need practitioners willing to say so, in writing, with evidence, to the people who can change it. Hashtags #trauma_informed_care #displaced_populations #refugee_mental_health #forced_displacement #psychosocial_support #systemic_advocacy #social_determinants #humanitarian_practice #MHPSS #asylum_seekers #internally_displaced_persons #structural_competency #community_participation #mental_health_policy #protracted_displacement References Bangpan, M., Felix, L., & Dickson, K. (2019). Mental health and psychosocial support programmes for adults in humanitarian emergencies: A systematic review and meta-analysis in low and middle-income countries. BMJ Global Health, 4(5), e001484. Bemme, D., & Kirmayer, L. J. (2020). Global mental health: Interdisciplinary challenges for a field in motion. Transcultural Psychiatry, 57(1), 3-18. Betancourt, T. S., Berent, J. M., Freeman, J., Frounfelker, R. L., Brennan, R. T., Abdi, S., Maalim, A., Abdi, A., Mishra, T., Gautam, B., Creswell, J. W., & Beardslee, W. R. (2020). Family-based mental health promotion for Somali Bantu and Bhutanese refugees: Feasibility and acceptability trial. Journal of Adolescent Health, 66(3), 336-344. Blackmore, R., Boyle, J. A., Fazel, M., Ranasinha, S., Gray, K. M., Fitzgerald, G., Misso, M., & Gibson-Helm, M. (2020). The prevalence of mental illness in refugees and asylum seekers: A systematic review and meta-analysis. PLOS Medicine, 17(9), e1003337. Bryant, R. A., Bawaneh, A., Awwad, M., Al-Hayek, H., Giardinelli, L., Whitney, C., Jordans, M. J. D., Cuijpers, P., Sijbrandij, M., Ventevogel, P., Dawson, K., & Akhtar, A. (2022). Twelve-month follow-up of a randomised clinical trial of a brief group psychological intervention for common mental disorders in Syrian refugees in Jordan. Epidemiology and Psychiatric Sciences, 31, e81. Bürgin, D., Anagnostopoulos, D., Board and Policy Division of ESCAP, Vitiello, B., Sukale, T., Schmid, M., & Fegert, J. M. (2022). Impact of war and forced displacement on children's mental health: Multilevel, needs-oriented, and trauma-informed approaches. European Child and Adolescent Psychiatry, 31(6), 845-853. Byrow, Y., Pajak, R., Specker, P., & Nickerson, A. (2020). Perceptions of mental health and perceived barriers to mental health help-seeking amongst refugees: A systematic review. Clinical Psychology Review, 75, 101812. Charlson, F., van Ommeren, M., Flaxman, A., Cornett, J., Whiteford, H., & Saxena, S. (2019). New WHO prevalence estimates of mental disorders in conflict settings: A systematic review and meta-analysis. The Lancet, 394(10194), 240-248. Hajak, V. L., Sardana, S., Verdeli, H., & Grimm, S. (2021). A systematic review of factors affecting mental health and well-being of asylum seekers and refugees in Germany. Frontiers in Psychiatry, 12, 643704. Han, H. R., Miller, H. N., Nkimbeng, M., Budhathoki, C., Mikhael, T., Rivers, E., Gray, J., Trimble, K., Chow, S., & Wilson, P. (2021). Trauma informed interventions: A systematic review. PLOS ONE, 16(6), e0252747. Henkelmann, J. R., de Best, S., Deckers, C., Jensen, K., Shahab, M., Elzinga, B., & Molendijk, M. (2020). Anxiety, depression and post-traumatic stress disorder in refugees resettling in high-income countries: Systematic review and meta-analysis. BJPsych Open, 6(4), e68. Hou, W. K., Liu, H., Liang, L., Ho, J., Kim, H., Seong, E., Bonanno, G. A., Hobfoll, S. E., & Hall, B. J. (2020). Everyday life experiences and mental health among conflict-affected forced migrants: A meta-analysis. Journal of Affective Disorders, 264, 50-68. Hussam, R., Kelley, E. M., Lane, G., & Zahra, F. (2022). The psychosocial value of employment: Evidence from a refugee camp. American Economic Review, 112(11), 3694-3724. Im, H., Rodriguez, C., & Grumbine, J. M. (2021). A multitier model of refugee mental health and psychosocial support in resettlement: Toward trauma-informed and culture-informed systems of care. Psychological Services, 18(3), 345-364. Jordans, M. J. D., & Kohrt, B. A. (2020). Scaling up mental health care and psychosocial support in low-resource settings: A roadmap to impact. Epidemiology and Psychiatric Sciences, 29, e189. Kienzler, H. (2019). Mental health system reform in contexts of humanitarian emergencies: Toward a theory of practice-based evidence. Culture, Medicine, and Psychiatry, 43(4), 636-662. Kohrt, B. A., Ottman, K., Panter-Brick, C., Konner, M., & Patel, V. (2020). Why we heal: The evolution of psychological healing and implications for global mental health. Clinical Psychology Review, 82, 101920. Mesa-Vieira, C., Haas, A. D., Buitrago-Garcia, D., Roa-Diaz, Z. M., Minder, B., Gamba, M., Salvador, D., Gomez, D., Lewis, M., Gonzalez-Jaramillo, W. C., Pahud de Mortanges, A., Buttia, C., Muka, T., Trujillo, N., & Franco, O. H. (2022). Mental health of migrants with pre-migration exposure to armed conflict: A systematic review and meta-analysis. The Lancet Public Health, 7(5), e469-e481. Miller, K. E., Jordans, M. J. D., Tol, W. A., & Galappatti, A. (2021). A call for greater conceptual clarity in the field of mental health and psychosocial support in humanitarian settings. Epidemiology and Psychiatric Sciences, 30, e5. Nickerson, A., Byrow, Y., O'Donnell, M., Mau, V., McMahon, T., Pajak, R., Li, S., Hamilton, A., Minihan, S., Liu, C., Bryant, R. A., Berle, D., & Liddell, B. J. (2019). The association between visa insecurity and mental health, disability and social engagement in refugees living in Australia. European Journal of Psychotraumatology, 10(1), 1688129. Papadopoulos, R. K. (2021). Involuntary dislocation: Home, trauma, resilience, and adversity-activated development. Routledge. Phillimore, J. (2021). Refugee-integration-opportunity structures: Shifting the focus from refugees to context. Journal of Refugee Studies, 34(2), 1946-1966. Purgato, M., Carswell, K., Tedeschi, F., Acarturk, C., Anttila, M., Au, T., Bajbouj, M., Baumgartner, J., Biondi, M., Churchill, R., Cuijpers, P., Koesters, M., Gastaldon, C., Ilkkursun, Z., Lantta, T., Nose, M., Ostuzzi, G., Papola, D., Popa, M., ... Barbui, C. (2021). Effectiveness of Self-Help Plus in preventing mental disorders in refugees and asylum seekers in Western Europe: A multinational randomized controlled trial. Psychotherapy and Psychosomatics, 90(6), 403-414. Sangalang, C. C., & Vang, C. (2017). Intergenerational trauma in refugee families: A systematic review. Journal of Immigrant and Minority Health, 19(3), 745-754. Scharpf, F., Kaltenbach, E., Nickerson, A., & Hecker, T. (2021). A systematic review of socio-ecological factors contributing to risk and protection of the mental health of refugee children and adolescents. Clinical Psychology Review, 83, 101930. Strang, A. B., & Quinn, N. (2021). Integration or isolation? Refugees' social connections and wellbeing. Journal of Refugee Studies, 34(1), 328-353. Substance Abuse and Mental Health Services Administration. (2014). SAMHSA's concept of trauma and guidance for a trauma-informed approach. SAMHSA. Sweeney, A., Filson, B., Kennedy, A., Collinson, L., & Gillard, S. (2018). A paradigm shift: Relationships in trauma-informed mental health services. BJPsych Advances, 24(5), 319-333. Sweeney, A., & Taggart, D. (2018). (Mis)understanding trauma-informed approaches in mental health. Journal of Mental Health, 27(5), 383-387. Tay, A. K., Riley, A., Islam, R., Welton-Mitchell, C., Duchesne, B., Waters, V., Varner, A., Silove, D., & Ventevogel, P. (2019). The culture, mental health and psychosocial wellbeing of Rohingya refugees: A systematic review. Epidemiology and Psychiatric Sciences, 28(5), 489-494. Tol, W. A., Leku, M. R., Lakin, D. P., Carswell, K., Augustinavicius, J., Adaku, A., Au, T. M., Brown, F. L., Bryant, R. A., Garcia-Moreno, C., Musci, R. J., Ventevogel, P., White, R. G., & van Ommeren, M. (2020). Guided self-help to reduce psychological distress in South Sudanese female refugees in Uganda: A cluster randomised trial. The Lancet Global Health, 8(2), e254-e263. Tribe, R. H., Sendt, K. V., & Tracy, D. K. (2019). A systematic review of psychosocial interventions for adult refugees and asylum seekers. Journal of Mental Health, 28(6), 662-676. Turrini, G., Purgato, M., Tedeschi, F., Acarturk, C., Anttila, M., Augustinavicious, J., Carswell, K., Churchill, R., Cuijpers, P., Koesters, M., Lantta, T., Nose, M., Ostuzzi, G., Popa, M., Sijbrandij, M., Tol, W. A., Valimaki, M., Wancata, J., White, R., & Barbui, C. (2021). Psychosocial interventions for adult refugees and asylum seekers: A systematic review and meta-analysis. Epidemiology and Psychiatric Sciences, 30, e57. United Nations High Commissioner for Refugees. (2024). Global trends: Forced displacement in 2023. UNHCR. Uphoff, E., Robertson, L., Cabieses, B., Villalon, F. J., Purgato, M., Churchill, R., & Barbui, C. (2020). An overview of systematic reviews on mental health promotion, prevention, and treatment of common mental disorders for refugees, asylum seekers, and internally displaced persons. Cochrane Database of Systematic Reviews, 9, CD013458. World Health Organization. (2022). World mental health report: Transforming mental health for all. World Health Organization. World Health Organization. (2023). Mental health of refugees and migrants: Risk and protective factors and access to care. World Health Organization.

This article offers a full academic review of Joseph Campbell's 1949 book The Hero with a Thousand Faces for a student readership. It has four aims. First, it explains the intellectual background of the book, showing that Campbell built his model on earlier work by Otto Rank, Lord Raglan, Vladimir Propp, Sir James Frazer, and above all Carl Jung. Second, it sets out the architecture of the #monomyth in plain language, describing the three movements of Departure, Initiation, and Return and the seventeen stages that Campbell placed inside them. Third, it traces how the model travelled out of comparative mythology and into screenwriting, video games, theme park design, marketing, psychotherapy, counselling, engineering education, and data communication. Fourth, it presents the serious scholarly objections that have accumulated against the model, including the charge of source selection bias raised by folklorists, the charge of cultural flattening captured in the phrase "Hollywood McMyth," the gender critique advanced most fully by Maria Tatar, and the ideological critique that reads the lone hero as a figure of individualism. The article argues that the correct student posture toward Campbell is neither devotion nor dismissal. The monomyth is best treated as a #heuristic, a set of questions to bring to a story, rather than as a law that stories must obey. Used that way, it remains one of the most productive teaching tools in narrative studies. Used as a template, it narrows the range of stories that get told and the range of readings that get made. The article closes with a practical protocol students can apply when they use Campbell in coursework, and with an agenda for future research. Keywords: monomyth, hero's journey, Joseph Campbell, comparative mythology, archetype, narrative structure, screenwriting, narrative identity 1. Introduction Few books written by an academic have escaped the academy as completely as The Hero with a Thousand Faces. Published in 1949 by a professor of literature at Sarah Lawrence College, it was a work of comparative mythology aimed at other scholars. Within four decades it had become a standard reference in Hollywood story departments, a fixture in creative writing syllabi, a talking point in advertising agencies, and eventually a #measurable_variable in peer reviewed psychology (Rogers et al., 2023). Campbell died in 1987. His model has kept working without him. The reason for that afterlife is simple to state. Campbell claimed that beneath the enormous variety of the world's #hero_myths there is one recurring shape. A person leaves the world they know, crosses into a strange region, faces trials there, wins something of value, and comes back changed, carrying that value to the community they left. He called this pattern the #monomyth, borrowing the word from James Joyce's Finnegans Wake, and he argued that the pattern recurs because it corresponds to something in the structure of the human mind (Campbell, 2008). That claim is either very large or very small, depending on how it is read. Read strongly, it says that all hero myths everywhere reduce to one story, and that the story is a map of psychological development shared by our species. Read weakly, it says that a great many hero myths, including many of the ones that Western readers already know, share a family resemblance that can be usefully named. Most of the trouble that has followed Campbell's book comes from the gap between these two readings. Campbell himself moved between them. His most enthusiastic readers took the strong version. His sharpest critics attacked the strong version. And a large amount of the useful scholarly work of the last five years has been an attempt to salvage the weak version while abandoning the strong one (Hanney, 2024). This article is written for students who are meeting Campbell for the first time, or who have met him through a screenwriting handbook and now need to place him properly in an academic argument. It does three things that a summary alone cannot do. It shows where the model came from, which explains a great deal about its shape. It shows what the model has been used for since 1949, which explains why it has survived criticism. And it takes the criticism seriously, because a student who cites Campbell without engaging the objections will produce weak work, and a student who dismisses him without engaging the model will produce work that is just as weak. The central argument is stated once here and defended through the rest of the article. The #hero_journey is valuable as an analytical instrument and dangerous as a #narrative_template. As an instrument it asks good questions of a story: what world does this character start in, what pulls them out of it, who helps, what is the worst moment, what do they carry back. As a template it answers those questions in advance, and a template that answers questions in advance stops being research and starts being formula. 1.1 Research questions The article is organised around four questions. What are the intellectual sources of Campbell's model, and how do those sources shape and constrain what the model can see? What exactly does the model claim, stage by stage, and what work does each stage do? Where has the model been applied outside comparative mythology, and what does the recent evidence say about whether those applications succeed? What are the strongest objections to the model, and what remains of the model after those objections are granted? 1.2 Why this matters for students now Three developments make this an unusually good moment to revisit Campbell. The first is that his model has finally been tested empirically rather than only argued about. A large study published in the Journal of Personality and Social Psychology found that people who see their own life story as a hero's journey report a stronger sense of #meaning_in_life, and that a short writing exercise which reframes a life story along those lines can raise that sense causally (Rogers et al., 2023). Whatever one thinks of Campbell as a mythographer, this is a real finding about a real effect, and students should know it exists. The second is that the critical scholarship has become organised rather than scattered. In 2024 the journal Media Practice and Education published an entire special issue devoted to unravelling the hero's journey, gathering critiques of its dominance in screen education, its treatment of gender, its fit with immersive media, and its ideological load (Hanney and Norrington, 2024). Before that issue, a student wanting to criticise Campbell had to assemble the case from fragments. Now the case is assembled. The third is that the model is being ported into new domains at speed. It has been used to structure #data_videos so that audiences follow a statistical argument (Wei, Qu and Xu, 2025). It has been used as a metaphor to explain rational emotive behaviour therapy to clients (Turner, 2025). It has been used to organise narrative interviews with engineering students. Each of these ports raises the same question in a new setting: is the structure being discovered in the material, or imposed on it? 2. Background: Where the Monomyth Came From Campbell did not invent the idea that hero stories repeat. He inherited a research tradition that was already about eighty years old when he began writing, and understanding that inheritance is the fastest way to understand both the strengths and the blind spots of his book. 2.1 The nineteenth century comparative impulse By the 1870s, anthropologists working in the comparative tradition had noticed that #myths from widely separated cultures kept producing similar motifs: floods, dying and reviving gods, miraculous births, journeys to the land of the dead. The dominant explanation was that these similarities reflected a shared stage of human mental development. Sir James Frazer's The Golden Bough, in its many editions, gathered an enormous quantity of ritual and mythic material and organised it around cycles of #death_and_rebirth. Frazer's method was to strip a story from its setting, place it beside a story from another continent, and let the resemblance speak. That method is exactly what later folklorists would object to in Campbell, and Campbell learned it from this tradition. 2.2 Otto Rank and the psychoanalytic hero In 1909 the psychoanalyst Otto Rank published The Myth of the Birth of the Hero, which took roughly thirty hero legends, from Sargon and Moses to Oedipus and Romulus, and extracted a common #birth_pattern. The hero is born to distinguished parents, there is a prophecy of danger, the infant is exposed or set adrift, is rescued and raised by humble foster parents or animals, grows up, discovers the truth of his origins, takes revenge on his father, and achieves honour (Rank, 2004). Rank read the pattern as a family romance, a dramatisation of the child's relationship to the parents. Two features of Rank's work carry straight into Campbell. The first is the idea that a #recurring_structure can be abstracted from many stories and stated as a numbered list. The second is the idea that the structure means something psychological rather than historical. Campbell kept both. 2.3 Lord Raglan and the ritual pattern Lord Raglan, writing in 1936, produced a twenty two point profile of the traditional hero, covering birth, accession, reign, downfall, and death, and scored figures such as Oedipus, Theseus, Moses, and King Arthur against it. Raglan's aim was partly deflationary. He wanted to show that figures assumed to be historical actually fit a #ritual_pattern so closely that their historicity was doubtful. Where Rank stopped at the hero's youth, Raglan carried the pattern to the grave. Campbell's model, which runs from ordinary life through trial to return, sits between the two. 2.4 Vladimir Propp and formal structure Working in a different tradition, the Russian formalist Vladimir Propp analysed a corpus of Russian wonder tales and argued that although characters and settings vary endlessly, the #narrative_functions they perform do not. He identified thirty one functions, in a fixed order, from the initial absence through the villainy, the departure, the donor's test, the struggle, the return, and the wedding (Propp, 1968). Propp's discipline is important. He worked on a defined corpus, from a single tradition, and made no claim that his functions applied to human storytelling as such. Campbell worked on an undefined corpus, from every tradition he could reach, and claimed exactly that. The contrast is the origin of much later criticism. 2.5 Carl Jung and the archetypes The decisive influence on Campbell was Carl Jung. Jung proposed that beneath the personal unconscious lies a #collective_unconscious, a layer of inherited psychic structure common to all humans, whose contents are #archetypes: the shadow, the anima and animus, the wise old man, the mother, the self (Jung, 1968). Archetypes are not fixed images but tendencies to produce images. They surface in dreams, in art, in symptoms, and in myth. Jung gave Campbell the answer to the question that Rank, Raglan, Frazer, and Propp had all left hanging. Why would the same shapes recur across cultures with no contact between them? Jung's answer was that they recur because they come from inside, not from cultural transmission. Campbell adopted this answer wholeheartedly. It is the load bearing wall of his entire argument, and it is also the part of the argument that most contemporary anthropologists and folklorists refuse to accept, because it explains cross cultural similarity by positing an inherited mental structure for which the evidence has never been produced. 2.6 The term itself The word #monomyth is not Campbell's coinage. He lifted it from James Joyce's Finnegans Wake, a novel he had studied closely and co-written a reader's guide to. The borrowing is telling. Campbell was, by training and instinct, a literary comparatist rather than a field anthropologist. He read myths as texts, looked for patterns as a literary critic looks for patterns, and interpreted symbols as a critic interprets symbols. He never conducted #fieldwork. This is not a moral failing, but it is a methodological fact with consequences that run through the whole book. 3. Method and Approach This article is a #conceptual_review, sometimes called a critical narrative review. It does not report new experimental data. It synthesises Campbell's primary text, the intellectual sources that fed it, and the scholarly literature that has evaluated it, with deliberate weighting toward material published in the last five years. Sources were gathered in three groups. The first group is the primary material: the third edition of The Hero with a Thousand Faces, together with Campbell's later interviews. The second group is the foundational secondary material that no serious discussion can omit, even though it is older than five years: Jung, Propp, Rank, Raglan, and the folkloristic critique associated with Alan Dundes and Barre Toelken. The third group, which carries most of the argumentative weight in sections 5 through 7, is recent peer reviewed work published between 2020 and 2025, including the 2024 special issue of Media Practice and Education, the 2023 Journal of Personality and Social Psychology study on narrative identity, the 2025 Frontiers in Psychology paper on therapeutic metaphor, and the 2025 IEEE study on data storytelling. The analytic strategy is straightforward. For each claim Campbell makes, the article asks three questions: what is the claim, what evidence supports it, and what would count as evidence against it. This last question matters more than it may appear. A theory that cannot be falsified by any story, because every story can be made to fit with enough interpretive effort, is not a strong theory. It is a #vocabulary. Part of the work below is deciding which parts of Campbell are theory and which are vocabulary. A limitation should be admitted at the outset. This is a review conducted in English, and the secondary literature it draws on is overwhelmingly Anglophone. That is itself an instance of one of the problems the article identifies. A genuinely #cross_cultural evaluation of a claim about universality cannot be conducted from inside one language. 4. The Architecture of the Monomyth Campbell summarised his model in a single sentence that students should memorise, because almost everything else is commentary on it: the hero ventures out from the world of common day into a region of supernatural wonder, encounters fabulous forces and wins a decisive victory there, and comes back from this mysterious adventure with the power to give gifts to others (Campbell, 2008). Three movements, then: #Departure, #Initiation, and #Return. Campbell divided these into seventeen stages. Not every myth contains all seventeen. Campbell was clear about this, and readers who accuse him of demanding a full set are being unfair to him. But he did claim that the sequence describes the deep shape even when stages are missing, merged, or reordered. What follows is a stage by stage account in plain language, with a note on the function each stage performs. 4.1 Departure 1. The Call to Adventure. Something disturbs the ordinary world. A messenger arrives, a hunt goes wrong, a stranger appears, a loss occurs. The function of the #call is to make the familiar world insufficient. Until the call, the hero has no reason to move. Campbell reads the call as a signal that a phase of life has been outgrown, and that the psyche is ready, whether the person is willing or not. 2. Refusal of the Call. The hero says no. This is not weakness in Campbell's reading but a real and dangerous possibility. The refusal converts the adventure into its negative: the person becomes stuck, and the world becomes a wasteland. Dramatically, the #refusal is what makes the eventual acceptance meaningful, because it establishes the cost. 3. Supernatural Aid. Once the call is accepted, a protective figure appears with something the hero needs: advice, a weapon, an amulet, a rule. This is the #mentor in later screenwriting terminology. Campbell's point is that the hero who commits receives help, and that the help arrives from a source outside ordinary competence. 4. The Crossing of the First Threshold. The hero passes the boundary of the known world. At that boundary stands a #threshold_guardian, a figure who tests whether the hero has the right to pass. The threshold is a real dividing line, not a decorative one. Beyond it, the rules the hero has lived by no longer hold. 5. The Belly of the Whale. The hero is swallowed. Campbell gives the stage a deliberately physical name because he wants the reader to feel the finality of it. This is the point of no return. The old identity is not merely left behind but consumed. Anyone who has seen the hero disappear into a cave, a ship, a labyrinth, a prison, or a coma has seen this stage. 4.2 Initiation 6. The Road of Trials. A series of tests, usually escalating. The #trials do more than fill time. They strip the hero of the assumptions and habits carried from home, and they are typically failed at least once. Campbell insists that the road of trials is not a demonstration of strength already possessed but a process by which strength is produced. 7. The Meeting with the Goddess. The hero encounters a figure of total, unconditional love, often maternal, often mysterious. Campbell reads this as the moment when the hero, having survived the trials, is offered union with the world rather than mastery over it. 8. Woman as Temptress. The hero is tempted to abandon the quest for physical or emotional satisfaction. The stage title is one of the most heavily criticised elements of the entire book, and section 7.3 returns to it in detail. The structural function Campbell intended is a #temptation to settle, to stop short of the goal, to take the reward that is available instead of the reward that is right. 9. Atonement with the Father. A confrontation with the figure who holds ultimate power in the hero's world. Campbell's spelling of atonement as at one ment is not a joke. The point is not that the hero defeats the father but that the hero comes to occupy the same position, seeing the world as the father sees it, and thereby stops being a child. This is the psychological centre of the whole model. 10. Apotheosis. The hero attains a state of expanded understanding, often after a symbolic or literal death. The #apotheosis is a moment of rest at the deepest point of the journey, and it is where Campbell's mystical commitments show most clearly. 11. The Ultimate Boon. The goal is achieved. The elixir, the fire, the treasure, the knowledge, the child, the cure. Campbell is careful to say that the #boon is usually less important in itself than in what its winning has made of the hero. 4.3 Return 12. Refusal of the Return. Having reached the goal, the hero may not want to go back. Why leave a place of illumination for a world that will not understand? Campbell treats this refusal as a moral failure, and the treatment reveals his priorities. For Campbell the journey is not complete until the community receives something. 13. The Magic Flight. The return is often a chase. The powers of the other world do not always release their treasures willingly. 14. Rescue from Without. The hero, exhausted, must sometimes be brought back by the help of others. This stage matters more than it usually receives credit for, because it undercuts the image of the wholly self sufficient hero. 15. The Crossing of the Return Threshold. The hero re-enters the ordinary world and faces the hardest problem in the entire structure: how to translate what was learned in a place where ordinary language does not apply into a form the community can use. 16. Master of Two Worlds. The hero achieves the capacity to move between the ordinary and the extraordinary without being destroyed by either. This is #integration, not conquest. 17. Freedom to Live. The hero lives without anxiety about the future or regret about the past. Campbell's final stage is not a plot event at all. It is a psychological condition, and it makes explicit what the whole model was aiming at. 4.4 What holds the stages together The seventeen stages are not seventeen separate insights. They are variations on a single movement: #separation, #transformation, and #reintegration. Read that way, the model is recognisably a description of a #rite_of_passage, and Campbell drew openly on the anthropology of initiation ritual to make the connection. The hero leaves the social body, is symbolically killed, is remade, and is returned to the social body with a new status. This is why the model feels so intuitive to readers. It is not only the shape of many myths. It is the shape of adolescence, of migration, of illness and recovery, of education, of grief. And that generality is simultaneously the model's greatest strength and the source of its greatest weakness, as section 7 will show. 4.5 The Vogler compression Most students today do not meet the seventeen stages. They meet twelve. Christopher Vogler, working as a story analyst at Disney, wrote a seven page memo summarising Campbell for studio executives, and later expanded it into The Writer's Journey, which compresses the monomyth into twelve stages with clearer dramatic labels: Ordinary World, Call to Adventure, Refusal of the Call, Meeting with the Mentor, Crossing the Threshold, Tests Allies and Enemies, Approach to the Inmost Cave, The Ordeal, Reward, The Road Back, The Resurrection, Return with the Elixir (Vogler, 2020). Vogler's version is not Campbell's. It is a #screenwriting_tool derived from Campbell, and the differences matter. Vogler adds an explicit Ordinary World at the start, which Campbell assumed rather than staged. He replaces Campbell's mystical vocabulary with dramatic vocabulary. He recasts the mythological figures as #character_archetypes that a writer can cast: hero, mentor, threshold guardian, herald, shapeshifter, shadow, trickster. And he presents the sequence as advice about what to write rather than as a description of what has been found. The recent literature has emphasised that a great deal of what circulates as "Campbell" in film schools, in advertising, and in self help is in fact Vogler, or a further simplification of Vogler (Turner, 2025). Students should not treat the twelve stage version as a shorthand for the seventeen stage version. They are different objects, with different purposes, and criticism aimed at one does not automatically land on the other. 5. Reception and Application The influence of The Hero with a Thousand Faces is easier to demonstrate than to measure. This section surveys the main domains of application and, where possible, reports what the recent evidence says about whether the application works. 5.1 Cinema and screenwriting The best known transmission route runs from Campbell through George Lucas to the American film industry, and then from Vogler's memo into studio development practice. By the 1990s the #hero_journey had become an unofficial grammar of the mainstream feature film, and by the 2010s it had become close to a requirement in pitch documents. Recent scholarship has treated this dominance as a problem to be explained rather than an achievement to be celebrated. Hanney (2024), introducing the Media Practice and Education special issue, argues that the model spread not because it was proved but because it was useful to an industry that needed a reliable, teachable, defensible way of assessing scripts at speed. A structure that lets a reader say "the ordeal comes too late" is administratively valuable whether or not it is true. Once the structure is embedded in commissioning, it becomes self confirming: films are written to fit it, and are then cited as evidence that it fits. Aston (2024) makes the corresponding argument about teaching. If film and media departments teach the monomyth as the structure of story, students will not merely learn it, they will fail to learn the alternatives, and will then reproduce the monomyth professionally. Her call is for #multiperspectival teaching that presents Campbell alongside non-Western structures such as the four part kishotenketsu, ensemble and network narratives, and non-linear and essayistic forms. 5.2 Games and interactive media Story driven video games adopted the monomyth early, and for an obvious reason: a game is already a structure of thresholds, trials, guardians, and rewards. The fit is so close that it can be circular. A designer who builds a level as a road of trials and then observes that the level follows the road of trials has demonstrated nothing. The harder problem in games is that the monomyth is a #linear_structure and interactive media are not. Campbell's hero has one path. A player has many. Recent work on #immersive_media has therefore asked whether the model survives the move at all. Groppel-Wegener (2024) examines theme parks, where the visitor is cast as a protagonist inside a designed environment, and finds that what the guest experiences is not a journey with a transformation but a sequence of staged encounters with the appearance of one. The hero's journey supplies the aesthetic of transformation without its substance, because nothing about the guest actually changes. 5.3 Manga, anime, and non-Western popular narrative The monomyth has been applied extensively to Japanese popular culture, and the results are instructive. Scally (2022) reads the films of Hayao Miyazaki through Campbell's framework and finds real correspondences, particularly in the departure and initiation movements. But the same literature keeps encountering a structural mismatch. Much Japanese narrative organises itself around #kishotenketsu, a four part form of introduction, development, twist, and reconciliation, which contains no obligatory conflict and no obligatory victory. A story can be complete without anyone defeating anything. This matters for the universality claim in a way that students should notice. If a widely used narrative structure exists that does not have conflict at its centre, then the monomyth is not the shape of story as such. It is the shape of one very successful family of stories. 5.4 Marketing and organisational storytelling Campbell entered marketing through the idea that a brand should not be the hero but the mentor, giving the customer, who is the hero, the tool needed to cross their own threshold. The framework spread quickly because it is easy to teach and produces usable copy. The scholarly reception is more sceptical. Preece, Daskalopoulou and Schau (2023), editing a collection on heroic marketing, note that the discipline is saturated with heroic undertones, in its theories, in its choice of exemplars, and in the way it celebrates its own practitioners. Their concern is that #heroic_framing is not a neutral communication device. It assigns roles, it manufactures antagonists, and it converts ordinary commercial transactions into moral quests. A model that can make anything feel meaningful is a powerful instrument, and instruments that powerful deserve scrutiny about who is holding them. 5.5 Psychology, counselling, and narrative identity This is where the most interesting recent evidence lies, and where students will find the strongest defence of Campbell available today. Rogers and colleagues (2023) approached the hero's journey as a hypothesis about #narrative_identity rather than about mythology. Drawing on the tradition of life story research associated with Dan McAdams, they reduced the model to seven elements: a protagonist, a shift, a quest, allies, a challenge, a transformation, and a legacy. They then built a measurement instrument, the Hero's Journey Scale, to detect how strongly those elements are present in the way a person narrates their own life. Across eight studies, with online samples and with older adults in a community sample, the presence of these elements predicted #meaning_in_life. More importantly, they developed a #restorying intervention that guided people to re-describe the events of their life using the seven elements, and found that it increased meaning in life causally, apparently by helping people connect scattered experiences into a coherent and directional account (Rogers et al., 2023). Read carefully, this is not a vindication of the strong Campbellian claim. It says nothing about whether all myths share a structure, and nothing about a collective unconscious. What it says is that this particular narrative shape, when applied to a life, produces a measurable psychological benefit. That is a claim about #storytelling_and_wellbeing, not about comparative mythology. It is a serious finding, and it deserves to be reported precisely. Turner (2025) extends the therapeutic use in a different direction, mapping the stages of the hero's journey onto the stages of rational emotive behaviour therapy, and arguing that the resulting metaphor, which he calls the rational emotive behavioural monomyth, helps clients understand a difficult clinical process by attaching it to a narrative structure they already know from books, films, and games. Note the logic of the argument. The monomyth is useful here precisely because it is #culturally_familiar, not because it is universal. Familiarity is a fact about a particular audience in a particular media environment. It is exactly what universality is not. 5.6 Education and technical communication The model has been used to structure narrative interviews with engineering students, giving researchers a coding scheme for turning long, messy interview transcripts into readable trajectories, and giving educators a way to make the experiences of underrepresented students legible and memorable to colleagues. Wei, Qu and Xu (2025) take the model into #data_communication, developing design guidance for creating data videos structured as a hero's journey, on the reasoning that audiences follow an argument better when it is delivered as a trajectory with stakes than when it is delivered as a sequence of findings. The same caution applies as in therapy. What is being exploited is the audience's prior familiarity with the shape. The shape works because they have seen it a thousand times, not because it is inscribed in the species. 6. Analysis: What the Model Actually Does Before turning to criticism, it is worth stating clearly what the monomyth accomplishes, because a critique that cannot explain a theory's success is a weak critique. 6.1 It names things that were previously unnamed Much of Campbell's practical value lies in the #vocabulary he supplied. Before Campbell, a reader could feel that a story had a moment where the hero passed a point of no return. After Campbell, that reader could say "belly of the whale" and be understood. A shared vocabulary lets people compare, argue, and teach. This is a genuine contribution, and it is independent of whether the universality thesis holds. 6.2 It converts description into diagnosis The model gives a writer or analyst a way to say what is missing. If a story feels weightless, the model suggests checking whether the threshold was ever really crossed, whether anything was actually lost, whether the return costs anything. Whether or not myths obey the pattern, the pattern is a good #diagnostic_checklist, and the fact that so many working writers keep it on the desk is evidence of that. 6.3 It links the outer story to the inner one Campbell's most durable insight may be the least controversial one. The events of the plot exist to produce a change in the character, and the change in the character is the point. #Transformation is what distinguishes a journey from a trip. This proposition survives the loss of every mystical claim in the book, and it is the proposition that the empirical psychology literature has picked up on. 6.4 Where the model becomes unfalsifiable Set against these strengths is a structural problem. The seventeen stages are defined loosely enough, and Campbell permits enough compression, substitution, reordering, and symbolic reading, that a determined analyst can find the pattern in almost any narrative. The mentor may be a person, an object, a memory, or a dream. The goddess may be male. The father may be an institution. The death may be metaphorical. Each of these permissions is individually reasonable. Together they mean that no story can #disconfirm the theory. A theory that cannot be disconfirmed can still be useful, but it cannot claim to be an empirical discovery about the world's myths. It has become an #interpretive_lens. Students should be precise about which of these two things they are using when they cite Campbell, because reviewers and examiners will be. 7. The Critical Case Against the Monomyth The objections fall into five families. They are cumulative rather than alternative: a student can hold all five at once. 7.1 Source selection bias and method The oldest and most damaging objection comes from #folklorists, whose discipline is built on exactly the material Campbell used. Alan Dundes, in his address to the American Folklore Society, was blunt about the problem: Campbell was accepted as an authority on folklore by the public without having done the work that the field requires, and he repeated errors that folklore scholarship had already corrected (Dundes, 2005). The methodological charge is #source_selection_bias. Campbell selected the myths that fitted his pattern, and the enormous quantity of myth that does not fit was simply not discussed. There is no defined corpus, no sampling procedure, no statement of what would count as a counterexample, and no accounting of how many traditions were examined and rejected. Compare Propp, who stated his corpus, or Raglan, who published his scoring scheme so that others could apply it and disagree. Campbell published conclusions. Barre Toelken pressed the related point about #cultural_context: the same plot does not mean the same thing in different cultures, and a pattern extracted from a story without its context of performance, belief, and social function may be a coherent shape whose meaning has been entirely misread (Toelken, 1996). His example of a Northwest Coast narrative in which a powerful male seeker's egotistical actions destroy rather than renew his culture is not a marginal exception. It is a story that inverts the moral logic of the monomyth while retaining its plot skeleton. 7.2 The universality claim and the colonial inheritance The strongest recent statement of the second objection is the phrase "Hollywood McMyth," which Hanney (2024), drawing on the critique developed by C. P. Nield, uses to summarise what the monomyth has become. The charge is severe and should be stated in full. The monomyth, on this reading, is a cut and paste quest myth assembled from decontextualised fragments of the world's mythologies, mixed with Jungian archetypes and psychoanalytic readings of dreams. The resulting product reflects a #colonial_inheritance: it takes the stories of others, strips their particularity, and reassembles them into a shape that centres European narrative and then declares that shape universal. The consequence, Hanney argues, is that a framework which advertised itself as a celebration of #cross_cultural_plurality has in practice become a monolithic dogma that drowns out other voices and other ways of telling. That is a specific, testable claim about the media industries, not a vague accusation, and the evidence for it is the observable narrowing of commissioned screen narrative around a single template. Students should notice the irony that gives the critique its force. Campbell's aim was ecumenical. He wanted to show that the world's cultures were saying the same thing in different languages, and he meant this as an argument against provincialism. The effect, transmitted through an industry with commercial reasons to standardise, was to make one culture's story shape into the price of admission for everyone else's. 7.3 Gender The gender critique is the one that most students will find easiest to verify from the primary text, because it is visible in the stage names. In Campbell's seventeen stages, the hero is grammatically and conceptually male. Women appear as the #goddess to be met, the #temptress to be resisted, or the prize to be won. There is no stage in which a woman departs, is initiated, and returns. Campbell's own gloss, that woman represents the totality of what can be known while the hero is the one who comes to know it, does not soften the problem. It restates it. The most thorough answer is Maria Tatar's The Heroine with 1,001 Faces (2021), whose title is a direct reply. Tatar's argument is not that women are missing from myth and folklore but that a model built around the #warrior_hero cannot see what they are doing. Heroines rarely wield swords, and are frequently denied even a pen. They act instead through the tools available to them: weaving, storytelling, listening, care, and #curiosity, which the tradition codes as a female vice and Tatar reframes as a virtue. Philomela, silenced by mutilation, weaves the story of her assault into a tapestry, and the tapestry is the act of heroism. Scheherazade survives by narrating. Tatar traces this line of #heroinism forward through Jane Eyre, Nancy Drew, Lisbeth Salander, Hermione Granger, and Katniss Everdeen, and argues that its recurring aim is repair and social justice rather than conquest and personal glory. The critical point is structural, not decorative. Tatar is not asking for a female hero to be inserted into Campbell's slots. She is arguing that the slots themselves encode a theory of what counts as significant action, and that the theory is wrong. 7.4 Individualism and ideology The fourth objection reads the monomyth politically. The Campbellian hero is singular. The community exists to be left at the start and served at the end, but the transformation happens to one person, in isolation, through personal ordeal. Gough-Brady (2024), writing in the same special issue, connects this figure to the #neoliberal productive citizen: the individual who is responsible for their own transformation, who converts suffering into personal growth, and whose success is understood as earned. Once this connection is made, a great deal of the model's contemporary popularity becomes legible. The monomyth flourishes in #self_help, in corporate leadership training, and in personal branding because it tells a story that those settings need to be true, namely that the individual can and should transform themselves through trial, and that the reward is deserved. Structural conditions, collective action, luck, and inherited advantage have no stage in the seventeen. The contributors to the 2024 special issue therefore argue for alternative paradigms built on community, companionship, and commonality rather than solitary ordeal (Hanney and Norrington, 2024). This is not a call to abolish the hero's journey. It is a call for #narrative_plurality, so that the collective story, the ensemble story, and the story of ordinary maintenance rather than exceptional rupture are also available. 7.5 Ahistoricism The final objection is philosophical and follows from Campbell's Jungian commitments. If the meaning of a myth lies in its archetypal content, then the historical circumstances of its telling are, at best, local colour. Campbell's method systematically converts the historical, the particular, and the event into the psychological and the universal. Critics have argued that this move has serious consequences when it is applied to real suffering, because a framework that reads catastrophe as an inner phase of a universal process has no way to hold the particular fact of what happened to particular people. The general form of the objection is that #psychologising_history dissolves responsibility. Whatever a student concludes about the specific controversies attached to Campbell's own remarks, the methodological point stands on its own and should be engaged. 8. Discussion: How to Use Campbell Responsibly If the objections in section 7 are granted, what is left? A good deal, provided the model is repositioned. This section proposes a working protocol. 8.1 Treat it as a question set, not an answer set The single most useful move a student can make is to convert each stage into a #question rather than a checkpoint. Instead of asking "where is the refusal of the call," ask "does this character resist the change, and if not, why does the story not need them to?" A story that has no refusal is not a defective story. It is a story that is doing something else, and the absence is data. Used this way, the model earns its keep even when the answer is negative. The stages become a systematic way of noticing what a narrative has chosen not to do, and the choices a narrative declines are often more revealing than the ones it makes. 8.2 Separate description from prescription Campbell was describing. Vogler was prescribing. These are different speech acts and they carry different burdens of proof. A description can be wrong about the world. A #prescription can only be more or less useful for a purpose. Confusing them is the most common error in undergraduate work on this topic, and it produces sentences like "all stories must have a mentor," which is neither a finding nor good advice. 8.3 Always name the corpus Any claim of the form "this pattern is universal" requires an answer to the question "in which sample?" Campbell never provided one. A student who says "the monomyth appears in the following twelve narratives, drawn from the following traditions, selected for the following reasons, and it fails to appear in these three" has already done more methodologically respectable work than the book under discussion. Defining the #corpus is the cheapest available upgrade to the quality of an essay. 8.4 Read the counterexamples on purpose Because the model is easy to confirm, confirmation is worth very little. The productive research move is to go looking for the stories that break it: narratives without conflict, narratives whose hero fails and whose failure is the point, narratives of collective rather than individual change, narratives that end without return, narratives from traditions where transformation is understood as loss rather than gain. Each #counterexample is worth ten confirmations. 8.5 Distinguish the four different claims Campbell's book actually contains four claims of very different strength, and they are usually collapsed. Separating them is the analytical heart of this article. The descriptive claim: many hero narratives share a departure, initiation, return structure. This is well supported and relatively uncontroversial for a large family of stories. The universality claim: this structure appears in all cultures because it is the structure of hero narrative as such. This is not supported. The corpus was never defined, the counterexamples were never engaged, and the existence of major narrative traditions organised on other principles tells against it. The psychological claim: the structure recurs because it corresponds to an inherited layer of the human mind. This rests entirely on Jung, and Jung's collective unconscious has not been established by any independent evidence that would satisfy contemporary psychology. The normative claim: living well means undertaking such a journey. This is a piece of #philosophy, and it should be argued about as philosophy, not smuggled in as an empirical result. Interestingly, it is the claim that the recent psychology comes closest to supporting, though only in the narrow sense that people who narrate their lives this way report more meaning (Rogers et al., 2023). A student who keeps these four claims apart, and states which one they are using, will write better than a student who does not. This is the practical payoff of the whole article. 8.6 A worked illustration of the method Consider a narrative that appears in almost every discussion of the monomyth: the Odyssey. On the standard reading, Odysseus leaves, endures trials, receives supernatural help, descends to the land of the dead, achieves the boon of homecoming, and returns transformed. Now apply the protocol. Does the poem have a refusal of the call? Not in the shape Campbell describes, because Odysseus is already in transit before the poem begins, and what he wants throughout is not adventure but its end. Does it have a transformation? This is genuinely arguable, and the argument is more interesting than the label. Odysseus is cunning at the start and cunning at the finish, and what changes may be the world he returns to rather than the man who returns. Is the return a triumph? It requires a massacre, and the poem does not entirely relax afterwards. Every one of these observations comes from taking Campbell's stages seriously as questions and then reporting honestly that the answers are complicated. A student who writes that the Odyssey follows the hero's journey has told the reader nothing. A student who writes that the Odyssey inverts the departure, since the hero's quest is to end the adventure rather than to begin it, has begun to think. 9. Limitations of This Review Three limitations should be recorded. First, the recent critical literature drawn on most heavily here is concentrated in media and screen studies. Anthropology and folklore have largely moved on from Campbell rather than continuing to engage him, which means that the contemporary critique available in English is stronger on the cultural politics of the model than on its ethnographic accuracy. A student with access to folklore archives could do original work here. Second, this review has treated the seventeen stage and twelve stage versions as distinct but has not systematically compared other derivatives now in circulation, including the seven element reduction used in the psychological literature and the simplified versions used in marketing. Whether these reductions preserve anything essential, or whether they simply produce a generic story arc that would work equally well without Campbell's name attached, is an open and answerable question. Third, the review is monolingual, and the central claim under examination is a claim about all human cultures. That is a real constraint and it should be stated rather than hidden. 10. Directions for Future Research Several questions follow directly from the discussion above and are within reach of a determined student. Corpus based testing. Take a defined, pre-registered sample of narratives from at least three traditions, code them against Campbell's stages using a published scheme and multiple coders, and report the fit rate honestly, including the misses. This has been talked about far more often than it has been done. Isolating the active ingredient. The finding that a hero's journey framing increases meaning in life (Rogers et al., 2023) invites a follow up: is the effect produced by the specific mythic structure, or by any coherent, directional, causally connected life narrative? A study comparing a hero's journey restorying condition against a non-heroic but equally coherent restorying condition would tell us a great deal. Non-conflict structures. Systematic comparison of #kishotenketsu, ensemble narrative, and cyclical or seasonal narrative forms against the monomyth, with attention to what each form can express that the others cannot. The template effect in industry. An empirical study of whether commissioning practices in screen industries measurably narrow narrative diversity, using script coverage documents and development notes rather than finished films. Heroinism as a structure. Tatar (2021) has provided the material for an alternative model but has not formalised it into stages, and may not have wished to. Whether a non-reductive formalisation of #heroinism is possible, or whether formalising it would repeat Campbell's mistake, is a genuinely interesting theoretical problem. 11. Conclusion The Hero with a Thousand Faces is a flawed book that has been extraordinarily productive. Both halves of that sentence are true and neither cancels the other. The flaws are real and they are not minor. Campbell selected his evidence, defined no corpus, engaged no counterexamples, conducted no fieldwork, and rested his explanation on a Jungian architecture that has not been independently established. He built a model of heroism in which women are landscape, and a model of transformation in which community is scenery. Transmitted through an industry with strong incentives to standardise, his model has helped narrow the range of stories that reach large audiences, and it has done so under a banner of #universality that its own method could not support. The productivity is equally real. Campbell gave the study of narrative a vocabulary that people actually use. He insisted that plot exists for the sake of #transformation, which remains the single most useful thing anyone can tell a beginning writer. He built a bridge between narrative and psychology that has now been crossed in the other direction, by researchers who have shown that people who tell their own lives in this shape find more meaning in them (Rogers et al., 2023), and by clinicians who use the shape to make a difficult therapeutic process comprehensible (Turner, 2025). He supplied a frame that has proved robust enough to organise engineering interviews, data videos, and theme parks, even where the fit is imperfect and instructive precisely because it is imperfect. The resolution this article proposes is not a compromise but a reclassification. The monomyth is not a discovery about the world's myths. It is an instrument for reading them, and like every instrument it shows some things by hiding others. A microscope is not wrong for failing to show galaxies. It becomes wrong only when its user announces that galaxies do not exist. Students should therefore learn Campbell thoroughly and then hold him lightly. Learn the stages, because the vocabulary is genuinely useful and because you cannot criticise what you cannot describe. Learn the sources, because they explain the shape of the thing. Learn the objections, because they are strong and because a citation of Campbell that ignores them will not survive a reader who knows the field. And then, when you sit down with an actual story, ask Campbell's questions rather than accepting his answers, and pay closest attention at the moments where the story refuses to cooperate. Those refusals are where the story is telling you what it is really about, and where Campbell, in spite of himself, remains the most useful guide available. Hashtags #JosephCampbell #TheHeroWithAThousandFaces #Monomyth #HerosJourney #ComparativeMythology #NarrativeStructure #Archetypes #Jung #Storytelling #LiteraryTheory #Screenwriting #MythStudies #NarrativeIdentity #CampbellCritique #StudentResearch References Aston, J. (2024). There is more to life than the monomyth: multiperspectival approaches to teaching narrative and story in university film and media departments. Media Practice and Education, 25(2), 123-136. https://doi.org/10.1080/25741136.2024.2331342 Campbell, J. (2008). The hero with a thousand faces (3rd ed.). New World Library. (Original work published 1949) Campbell, J., and Moyers, B. (1988). The power of myth. Doubleday. Dundes, A. (2005). Folkloristics in the twenty-first century. Journal of American Folklore, 118(470), 385-408. Gough-Brady, C. (2024). The heroic character, the neo-liberal productive citizen, and the feminist filmmaker. Media Practice and Education, 25(2). Groppel-Wegener, A. (2024). Immersive protagonists: exploring the notion of the hero in theme parks. Media Practice and Education, 25(2), 137-148. https://doi.org/10.1080/25741136.2024.2324089 Hanney, R. (2024). One myth to rule them all and in the darkness bind them: a critical examination of Joseph Campbell's The Hero's Journey. Media Practice and Education, 25(2), 113-122. https://doi.org/10.1080/25741136.2024.2333649 Hanney, R., and Norrington, A. (Eds.). (2024). Unravelling the hero's journey: new narrative paradigms for story, semantics and the psychology of experience [Special issue]. Media Practice and Education, 25(2). Jung, C. G. (1968). The archetypes and the collective unconscious (2nd ed.). Princeton University Press. Preece, C., Daskalopoulou, A., and Schau, H. J. (2023). Troubled times demand heroes: heroic marketing and marketing heroes. Marketing Theory, 23(1), 3-13. https://doi.org/10.1177/14705931221148513 Propp, V. (1968). Morphology of the folktale (2nd ed.). University of Texas Press. Raglan, Lord. (1936). The hero: a study in tradition, myth and drama. Methuen. Rank, O. (2004). The myth of the birth of the hero: a psychological exploration of myth (expanded ed.). Johns Hopkins University Press. (Original work published 1909) Rogers, B. A., Chicas, H., Kelly, J. M., Kubin, E., Christian, M. S., Kachanoff, F. J., Berger, J., Puryear, C., McAdams, D. P., and Gray, K. (2023). Seeing your life story as a Hero's Journey increases meaning in life. Journal of Personality and Social Psychology, 125(4), 752-778. https://doi.org/10.1037/pspa0000341 Scally, D. (2022). Miyazaki and the hero's journey. McFarland. Segal, R. A. (1990). Joseph Campbell: an introduction (rev. ed.). Penguin. Tatar, M. (2021). The heroine with 1,001 faces. Liveright. Toelken, B. (1996). The dynamics of folklore (rev. ed.). Utah State University Press. Turner, M. J. (2025). Tales of rationality: the rational emotive behavioral monomyth as a metaphorical alignment of rational emotive behavior therapy and the Hero's Journey narrative structure. Frontiers in Psychology, 16, 1572636. https://doi.org/10.3389/fpsyg.2025.1572636 Vogler, C. (2020). The writer's journey: mythic structure for writers (25th anniversary ed.). Michael Wiese Productions. Wei, Z., Qu, H., and Xu, X. (2025). Telling data stories with the hero's journey: design guidance for creating data videos. IEEE Transactions on Visualization and Computer Graphics, 31(1), 962-972. https://doi.org/10.1109/TVCG.2024.3456330

Cities that sit at the crossroads of global movement have become a defining feature of the present century. Places such as Dubai, Singapore, Toronto, Doha, Berlin, Auckland and Johannesburg now host populations in which a very large share of residents were born somewhere else, and in which a significant number expect to leave again within a few years. This article examines what community integration means in these settings, and asks what #social_work can realistically contribute to the building of cohesion when the population itself is constantly turning over. Drawing on an integrative review of scholarship published mainly between 2019 and 2024 across migration studies, urban sociology, community development and social work, the article develops the concept of the high-mobility hub as a distinct type of urban environment. It argues that mainstream integration models, which were built for settlement societies with a permanent-residence horizon, do not travel well to cities where a large part of the population is legally or practically temporary. The article proposes an alternative practice framework, described here as anchored mobility, which is organised around five pillars: rapid relational anchoring, shared infrastructure, structural advocacy, portable capability, and reciprocal recognition. It then reviews intervention approaches at micro, mezzo and macro levels, discusses measurement and evaluation, sets out ethical tensions, and outlines implications for social work education. The central claim is that #cohesion in a high-mobility hub cannot be produced by asking newcomers to assimilate into a stable host community, because in many such cities no stable host community exists in the classical sense. Cohesion must instead be built through the deliberate design of shared spaces, shared institutions and fair rules, with social workers acting as designers, brokers and advocates rather than only as caseworkers. Keywords: community integration, high-mobility cities, social cohesion, migration, social work practice, urban belonging, superdiversity, temporary migration, community development 1. Introduction 1.1 The Problem in Plain Terms Something unusual has happened to a group of cities over the past three decades. They have grown very fast, they have grown mainly through the arrival of people born elsewhere, and a large share of those people arrive knowing they may not stay. The result is a kind of city that does not fit neatly into the categories that social scientists inherited from the twentieth century. It is not the classic immigrant gateway, where newcomers arrive expecting to become citizens over a generation or two. It is not the ordinary global city with a diverse but broadly settled population. It is something more restless: a hub through which enormous numbers of people pass, and in which a meaningful minority stay long enough to build lives, raise children and grow old, but under conditions that never quite promise permanence. Consider what this means for #community_integration in practice. Most of the frameworks that guide integration policy assume a direction of travel. Newcomers arrive, they learn the language, they find work, they form friendships across group lines, they eventually acquire secure status, and over time the distinction between newcomer and established resident fades. That is the story embedded in the standard indicator sets used by governments and international organisations. It is a reasonable story for some places. It is close to fiction in a city where a third of the workforce holds a two-year contract, where residence is tied to employment, and where the average length of stay is shorter than the time it takes most integration programmes to show measurable results. This article takes that mismatch seriously. It asks a practical question with a difficult answer: if people keep moving, what holds a city together, and what can #social_workers do about it? 1.2 Why This Question Matters Now Three developments make this question urgent. The first is scale. International migration has grown steadily in absolute numbers, and the great majority of migrants end up in cities rather than rural areas. Urban authorities, not national governments, are the level of the state that people actually encounter when they look for housing, register children in school, or seek help in a crisis. Yet cities frequently have limited legal power over the immigration rules that shape their populations. The second is the rise of #temporary_migration as a policy preference. Many receiving states have deliberately shifted away from permanent settlement pathways toward circular, contract-based and time-limited schemes. Some of these systems are relatively generous; others tie legal residence directly to a single employer, which concentrates enormous power in the hands of that employer. Either way, the effect on community life is the same: a large group of residents lives with a horizon that is measured in years rather than decades. The third is the visible strain on #social_trust in diverse urban settings. The research literature on whether diversity itself reduces trust has become considerably more careful than it was fifteen years ago, and the current weight of evidence suggests that the effect, where it exists at all, is small, highly context-dependent, and often better explained by economic disadvantage, residential instability and unequal treatment than by cultural difference (Dinesen, Schaeffer and Sonderskov, 2020). That is an important corrective. But it does not mean that cohesion looks after itself. It means that cohesion is produced or destroyed by conditions, and conditions can be designed. 1.3 Aims and Contribution This article has four aims. First, it defines the high-mobility hub as an analytical category, distinguishing it from the settlement city and from the ordinary diverse city. Second, it reviews and criticises the dominant models of #integration and #social_cohesion, showing where they break down under conditions of high turnover. Third, it proposes the anchored mobility framework as a practice orientation for social work in these settings, and maps concrete interventions to each of its pillars. Fourth, it addresses the awkward questions: how do we measure success when the people we serve leave, how do we practise ethically when the state limits what we can advocate for, and how should professional education change. The contribution is intended to be practical rather than only conceptual. The article is written for students and practitioners as much as for researchers, and it tries to be honest about where the evidence is thin. 1.4 Structure of the Article Section 2 sets out the methodology of the review. Section 3 defines the high-mobility hub. Section 4 examines the conceptual foundations of integration and cohesion, including their critiques. Section 5 presents the theoretical framework. Section 6 introduces the anchored mobility model. Sections 7, 8 and 9 discuss micro, mezzo and macro interventions respectively. Section 10 addresses digital and infrastructural dimensions. Section 11 discusses measurement. Section 12 examines ethical tensions. Section 13 covers implications for education and training. Section 14 sets out limitations and a research agenda. Section 15 concludes. 2. Methodology 2.1 Design This is an integrative conceptual review rather than a systematic review in the strict clinical sense. The aim was not to estimate an effect size but to synthesise scattered literatures into a usable framework for practice. Integrative reviews are appropriate when a field is fragmented across disciplines, when the object of study is contested, and when the goal is theory-building rather than effect measurement. All three conditions apply here. 2.2 Search and Selection Literature was drawn from four bodies of work: migration and #superdiversity studies, urban sociology and geography, community development, and social work with migrant and displaced populations. Priority was given to material published from 2019 onward, with a small number of older foundational works retained because they remain the conceptual reference points that later scholarship argues with. Where an older framework is cited, it is cited as a starting point that subsequent work has revised. Sources were included when they addressed at least one of the following: the meaning and measurement of integration or cohesion, the experience of migrants in high-turnover urban settings, the design or evaluation of community-level interventions, or the professional role of #social_work in migration contexts. Sources were excluded when they focused narrowly on border enforcement, on national-level admission policy without a community dimension, or on individual clinical treatment without a social component. 2.3 Analysis Material was read and coded thematically. Codes were grouped first descriptively, then organised into the five pillars presented in Section 6. The framework was developed inductively from the literature and then checked back against it. Where the literature offered contradictory findings, the contradiction is reported rather than resolved artificially. 2.4 A Note on Limits This review does not present new empirical data. It cannot therefore claim that the proposed framework works; it can only claim that the framework is consistent with existing evidence and addresses gaps that the existing models leave open. Section 14 discusses how the framework could be tested. 3. Defining the High-Mobility Hub 3.1 Four Defining Features A high-mobility hub can be identified by four characteristics that appear together. High foreign-born share. A substantial proportion of residents, in extreme cases a clear majority, were born outside the country. This is not unusual in itself; several established immigrant cities meet this criterion. High turnover. What distinguishes the hub is not only who is there but how long they stay. In a settlement city, the foreign-born population is largely a stock: people who came and remained. In a high-mobility hub, a large part of the foreign-born population is a flow: people who came and will leave. #Population_churn is the defining variable, and it is the one most often left out of integration statistics. Structural temporariness. The turnover is not accidental. It is produced by law and by labour market design. Residence permits tied to employment, contract-limited visas, restrictions on family reunification, and the absence or near-absence of naturalisation pathways all create a population that is legally prevented from settling regardless of how long it stays or how much it wants to. Some individuals in such cities have lived there for twenty or thirty years and remain, in legal terms, temporary. Layered stratification. Mobility is stratified by class, nationality and occupation. The senior professional on an expatriate package and the construction worker on a labour contract may both be legally temporary, but they experience the city in almost entirely different ways: different neighbourhoods, different mobility, different access to services, different capacity to bring family. Any analysis that treats migrants as a single category will be useless here (Vertovec, 2023). 3.2 What Makes This Different The consequence of these four features is that the classical integration question is malformed. Integration into what? If the host society is itself composed largely of people who arrived recently and expect to leave, then asking newcomers to integrate into it is close to a category error. There may be a national citizen population that is culturally and legally distinct, but in some hubs that population is a numerical minority and is separated from most residents by an unbridgeable legal boundary. This creates a specific and under-theorised condition: a city that functions, that is often prosperous and orderly, and that nevertheless has weak formal mechanisms for producing common membership. Cohesion in such a city, where it exists, tends to be produced informally, through workplaces, neighbourhoods, religious institutions, sports clubs, national associations, schools and online groups. It is real, but it is fragile, and it is unevenly distributed. Those with money and mobility find it easily. Those without find it far harder. 3.3 A Typology of Residents For practice purposes, it helps to distinguish several positions within a high-mobility hub, since the intervention needs of each differ sharply. The transient professional stays two to five years, has good income and mobility, and integrates mainly with other transients. The risk here is not exclusion but shallow attachment: a comfortable life with no stake in the city. The long-term temporary resident has lived in the city for a decade or more without secure status. This group is often deeply attached to the city and deeply insecure within it. It includes many who were born there and have never lived anywhere else. The contract worker has restricted mobility, limited language access, employer-controlled housing, and little practical ability to participate in civic life. This is the group with the greatest structural vulnerability and the least visibility in integration policy. The citizen or permanent resident may be a numerical minority and may experience the rapid change of the city as loss. Ignoring this group's concerns is both analytically and politically foolish; it is also, in many hubs, the group whose sense of #belonging is least examined. The displaced person or asylum seeker may occupy an ambiguous position, sometimes without any legal status at all, and is frequently absent from official data altogether. Effective #community_work in a hub means holding all of these positions in view simultaneously, which is difficult, because they rarely meet. 4. Conceptual Foundations and Their Discontents 4.1 The Standard Integration Model The most influential practical model of integration in the English-language literature organises the concept around markers of achievement (employment, housing, education, health), social connections (bonds within groups, bridges between groups, links to institutions), facilitators (language, cultural knowledge, safety and stability), and a foundation of rights and citizenship. The model has been enormously useful. It gave practitioners a shared vocabulary, and it made the point that integration is not simply about jobs or language but about relationships and rights together. Yet the model rests on assumptions that a high-mobility hub violates. It assumes that #citizenship is available as a foundation. It assumes that time works in the migrant's favour, so that longer residence produces deeper integration. It assumes that there is a receiving society into which one integrates. It assumes that the relevant unit of analysis is the individual migrant and the relevant outcome is that migrant's position relative to the majority. 4.2 The Critiques Recent scholarship has pressed hard on these assumptions, and the criticisms are worth setting out clearly because they shape everything that follows. Integration is normatively loaded. The word implies a deficit on one side. It positions the migrant as the person who must change and the receiving society as the standard against which change is measured. Spencer and Charsley (2021) identify this as one of several core critiques, alongside the tendency of the concept to be used politically rather than analytically, its unclear relationship to equality, and its frequent conflation of description with prescription. The two-way process is often rhetorical. Policy documents routinely state that #integration is a two-way process in which both migrants and receiving societies adapt. In practice, almost all indicators measure migrants. Klarenbeek (2021) argues that this is not merely sloppy implementation but a conceptual problem: as long as integration is defined as the movement of a minority toward a majority, adding a sentence about mutual adaptation does not change the underlying logic. A genuinely relational conception would treat non-integration as something produced by both parties, and would therefore treat exclusionary behaviour by the majority as a form of failure to integrate. Context does more work than individual effort. Phillimore (2021) develops the idea of integration opportunity structures, shifting the analytic focus from what refugees do to what the environment allows them to do. Two people with identical skills, motivation and language ability will achieve very different outcomes depending on the housing market, the recognition of their qualifications, the behaviour of employers, the density of local organisations and the attitudes they encounter. If that is true, then measuring the individual and ignoring the context is measuring the wrong thing. Settlement is not the only valid outcome. The literature on transnationalism established long ago that people can live meaningful lives across two or more places without this representing a failure. The high-mobility hub makes this point unavoidable. A worker who sends remittances home, maintains close family ties abroad, and plans to return is not failing to integrate. She may be doing exactly what she came to do. What she needs is fair treatment, safety, dignity and connection while she is here, not a programme designed to attach her permanently to a place she does not intend to stay. 4.3 Social Cohesion: A Second Contested Term #Social_cohesion is at least as slippery. It is used to mean shared values, mutual trust, absence of conflict, willingness to cooperate, sense of belonging, and participation in common institutions, sometimes all at once. Fonseca, Lukosch and Brazier (2019) note that the concept is frequently invoked without definition and that its measurement is correspondingly inconsistent. For the purposes of this article, cohesion is treated as having three distinguishable components, because they can move independently. Relational cohesion is the density and quality of ties between people, including ties across group boundaries. It is measurable through network methods and through survey items on friendship, contact and mutual help. Institutional cohesion is the extent to which people share and trust common institutions: schools, clinics, transport, courts, local government. It can be high even where relational cohesion is low, and vice versa. Perceived fairness is the belief that the rules of the city apply to everyone and that one's group is not systematically disadvantaged. This is the component most often neglected and, arguably, the one that matters most. A city can have friendly neighbours and functioning buses and still lack cohesion if a large part of the population believes the system is rigged against it. The separation matters for practice. A great deal of #cohesion_work concentrates on relational cohesion, because contact events are easy to organise and photograph. Very little concentrates on perceived fairness, because that requires confronting structures. 4.4 Contact, Trust and What Actually Changes Attitudes The intergroup contact tradition provides the theoretical basis for a large share of community cohesion programming. The core finding, that contact under favourable conditions reduces prejudice, has held up reasonably well. But the more recent and more sober reviews are important reading for anyone about to design a #community_event. Paluck, Porat, Clark and Green (2021) surveyed the prejudice reduction literature and found that the field's evidence base is weaker than its confidence. Many studies use unrepresentative samples, measure attitudes immediately after the intervention rather than months later, and rarely measure behaviour at all. Effects that appear robust in the laboratory frequently shrink or vanish in the field. Light-touch interventions, the single workshop or the one-off festival, tend to produce small and short-lived changes. This should not lead to despair, but it should lead to humility. It suggests that #intergroup_contact works best when it is sustained rather than episodic, when it involves cooperation toward a shared goal rather than mere exposure, when participants have roughly equal standing in the activity, and when it is supported rather than undermined by the surrounding institutional environment. It also suggests that the one-day multicultural food festival, whatever its charms, is not an integration strategy. 4.5 Social Capital: Bonds, Bridges and Links The bonding, bridging and linking distinction remains useful, and recent empirical work has refined it. Strang and Quinn (2021), studying refugees' social connections, found that the relationship between social ties and wellbeing is not simply that more ties are better. Ties can carry obligation, surveillance and shame as well as support. Bonding ties within one's own national or ethnic group can be a resource and a constraint at the same time. Relationships with people who are also struggling may offer emotional support but few material resources. Ryan (2023) makes a related point about the temporal quality of migrant networks. Ties are not fixed assets; they are made, maintained and lost over time, and they take work. A newcomer who arrives with no local network is not merely lacking a resource; she is facing a demanding project of relationship-building at precisely the moment when she has least time, least language and least energy. In a high-mobility hub, this project must be repeated by each new cohort, and it must be repeated by people whose existing ties keep leaving. This last point deserves emphasis, because it is the specific #network_problem of the hub. In a stable city, a newcomer's network grows over time. In a hub, it grows and is depleted simultaneously, because the friends she makes are themselves on time-limited contracts. Many long-term residents of high-mobility cities describe a distinctive fatigue: the exhaustion of repeatedly forming attachments that are then broken by other people's departures. This is not a marginal emotional detail. It is a structural feature that shapes willingness to invest in local relationships at all, and it directly undermines relational cohesion. 5. Theoretical Framework Four theoretical resources are combined here. 5.1 Ecological Systems Thinking The ecological perspective, long standard in social work, locates the individual within nested systems: the immediate settings of daily life, the connections between those settings, the wider institutional structures that shape them, and the cultural and legal environment that surrounds everything. Its value in a high-mobility hub is that it forces attention upward. If a worker's isolation is produced by an employer-controlled accommodation compound located far from any public space, then counselling that worker about loneliness is treating a symptom while ignoring its cause. Ecological thinking makes the cause visible and therefore actionable. The framework also adds a time dimension. Life events, transitions and historical change interact with the systems around a person. Migration is precisely such a transition, and it typically involves simultaneous disruption across every level of the system at once. This is why migration is so demanding even for people who chose it freely and are doing well. 5.2 Superdiversity The superdiversity perspective insists that #diversity in contemporary cities cannot be captured by ethnicity alone (Vertovec, 2023). Residents differ by nationality, language, religion, migration channel, legal status, labour market position, gender, age and length of stay, and these variables interact in ways that produce highly specific situations. Two people from the same country may have nothing in common if one arrived as a professional on a renewable permit and the other on a construction contract with a confiscated passport. For practice, superdiversity implies that ethnic-group-based service design will systematically misfire. The relevant unit of intervention is often not the ethnic community but the situation: workers in a particular sector, mothers with children in a particular school, residents of a particular building. 5.3 Opportunity Structures and Emplacement Combining Phillimore's (2021) opportunity structure argument with the concept of emplacement gives a second theoretical pillar. Emplacement refers to the process by which people become part of a specific place, not a nation. It happens through the everyday: the walk to work, the shop where one is recognised, the park bench, the mosque or church, the school gate, the barber. Wessendorf and Phillimore (2019) show that newcomers in superdiverse areas often build their sense of place through such ordinary encounters rather than through any programme. This has a direct practice implication that is easy to miss. If emplacement happens in ordinary places, then the condition of ordinary places is a #social_work concern. Whether there are benches. Whether the library is open in the evening. Whether the park is lit. Whether there is anywhere at all to sit down without buying something. These are not peripheral to integration; in a high-mobility hub they may be the main mechanism available. 5.4 Anti-Oppressive and Structural Practice The fourth pillar is the tradition of anti-oppressive and structural social work, which insists that individual difficulties frequently have structural causes and that practice which ignores those causes becomes a way of managing injustice rather than addressing it (Ledwith, 2020). In migration contexts this tradition is uncomfortable, because the structures in question are often legal ones that the social worker cannot change and may be employed to administer. The tradition does not require the practitioner to solve the structural problem single-handedly. It requires the practitioner to name it accurately, to avoid pretending that its effects are personal failings, to document its consequences, and to work with others toward change over the long term. In some jurisdictions this can be done openly. In others it must be done quietly, through professional bodies, through evidence, and through incremental institutional argument. The choice of method is contextual. The refusal to pretend is not. 6. The Anchored Mobility Framework 6.1 The Core Idea The proposal here is to stop asking how to make mobile people settle and start asking how to give mobile people a stake. The framework is called anchored mobility because it accepts movement as a permanent condition and seeks to provide anchors within it, rather than treating movement as the problem to be solved. An anchor, in this sense, is anything that gives a person a reason and a means to invest in the place while they are in it: a relationship, a role, a shared space, a right, a recognised contribution. Anchors do not require permanence. They require reliability, access and reciprocity. 6.2 The Five Pillars Pillar One: Rapid relational anchoring. Because time is short, the first weeks matter far more than they do in a settlement city. A newcomer who has not formed a single local relationship within three months is at high risk of remaining isolated for the whole of their stay. Interventions must therefore be front-loaded, proactive and low-threshold rather than waiting for people to present with problems. Pillar Two: Shared infrastructure. Cohesion in a mobile city is carried disproportionately by places that everyone uses, because those are the only sites where people who would never seek each other out actually meet. Libraries, clinics, parks, transport, schools and markets are the load-bearing walls of #urban_cohesion. Protecting and improving them is integration work, even though it is rarely funded as such. Pillar Three: Structural advocacy. Where the conditions of temporariness themselves produce harm, through employer-tied residence, wage theft, restrictions on family life, or exclusion from health coverage, then advocacy is not an optional extra. It is the intervention. The other four pillars will not compensate for a structure that makes ordinary life impossible. Pillar Four: Portable capability. Since many people will leave, services should be designed so that what people gain is not lost when they move: recognised skills, documented work histories, transferable qualifications, savings that can be moved, health records, and networks that survive relocation. This reframes success. A worker who leaves with certified skills, savings intact and health preserved has been well served, even though every settlement indicator would record her as a failure. Pillar Five: Reciprocal recognition. Cohesion requires that both newcomers and longer-established residents see themselves as recognised, and it collapses when either group feels that the other is being favoured. This means that #cohesion_programmes must have something in them for established residents too, and must be honest about the real costs that rapid change can impose on existing communities, particularly around housing costs and pressure on services. 6.3 Relationship to Existing Models Anchored mobility does not discard the standard integration domains. Employment, housing, health, education, language and safety all remain central. What it changes is the horizon against which success is judged and the weight given to context over individual adaptation. It converts the settlement question, will this person become one of us, into a fairness question, is this person able to live a decent and connected life here under rules that apply equally to everyone, for as long as they are here. 7. Micro-Level Interventions 7.1 Arrival Support and the First Ninety Days The evidence on early intervention is consistent across contexts: the first months set trajectories. In high-mobility hubs the case is even stronger, because a delay of six months may consume a fifth of the person's entire stay. Effective #arrival_support in these settings shares several features. It is proactive rather than referral-based, reaching people through workplaces, schools and housing providers rather than waiting for self-referral. It combines practical orientation with relational introduction, because information alone does not reduce isolation. It uses peer navigators, that is, people who arrived recently themselves and can explain the city in the newcomer's language and from the newcomer's position. And it front-loads the things that unlock everything else: a bank account, a phone number, a transport card, an understanding of one's rights, and one named person who can be contacted again. Peer navigation deserves particular attention because it addresses two problems at once. It gives newcomers accessible help from someone who understands their position, and it gives the navigator a valued role, which is itself an anchor. The role must be paid or otherwise properly recognised, however. Unpaid peer work extracted from people who are themselves precarious is exploitation, however warmly it is described. 7.2 Culturally Responsive and Trauma-Informed Casework A significant minority of arrivals in any hub, and a large majority in some displaced populations, carry histories of violence, loss or exploitation. Systematic evidence indicates elevated rates of depression, anxiety and post-traumatic stress among refugees and asylum seekers relative to host populations, though estimates vary widely across studies and settings (Blackmore et al., 2020). Trauma-informed practice in a high-mobility hub must contend with a specific complication: the trauma is frequently ongoing rather than past. A worker whose passport is held by an employer, who is owed six months of wages and who cannot leave, is not recovering from adversity. She is in it. Clinical frameworks designed for post-trauma recovery do not fit a situation of continuing structural danger, and applying them can be actively harmful if they imply that the problem is the person's response rather than the situation. In such cases the intervention that helps is the one that changes the situation, and #trauma_informed_practice must therefore include practical and legal action, not only therapeutic conversation. Cultural responsiveness is best understood as cultural humility rather than cultural competence. The competence model implies that a practitioner can learn the relevant facts about a culture and then act on them. In a superdiverse hub, the number of relevant cultures, sub-cultures and situations is effectively unbounded, and generalisations about national groups will be wrong more often than they are right. Humility means asking rather than assuming, treating the person as the authority on their own life, and staying alert to the ways in which the practitioner's own position shapes the encounter. 7.3 Language, Literacy and Communication Language is the single most commonly cited barrier in every study of migrant experience, and it is frequently the one that services address worst. Classes are often scheduled at times when working people cannot attend, held in places that require expensive travel, and pitched at a level that assumes prior literacy. Practical improvements are usually mundane. Classes at the workplace or in the accommodation compound. Classes at hours that fit shift patterns. Separate provision for people who are not literate in any language, since teaching a second language to someone who cannot read the first is a fundamentally different task. Childcare during classes, without which mothers cannot attend. And a realistic acceptance that a functional survival vocabulary delivered quickly may serve a two-year resident better than a grammar-first curriculum aimed at fluency in five. Interpretation is the other half of the problem. The routine use of children as interpreters for their parents in medical and legal settings remains widespread and is indefensible. It burdens the child, distorts the information and violates the parent's privacy. Building a properly trained, properly paid interpreter pool is unglamorous #service_design work with very large returns. 7.4 Addressing Departure and Loss One micro-level need is almost entirely unaddressed in the literature and in practice: helping people manage the repeated loss of friends who leave. In high-mobility hubs this is a normal, recurring experience, and it has recognisable emotional consequences, including a defensive reluctance to form new attachments. Naming this experience is itself useful. So is normalising it, since people frequently interpret their own withdrawal as a personal failing rather than as a reasonable response to repeated loss. Practical responses include supporting continuity of relationships across distance, helping groups to mark departures rather than letting people simply disappear, and designing community structures around roles rather than around individuals, so that when a person leaves, the role can be filled and the group survives. This last point is a small but genuinely important piece of #community_organising craft in a mobile city. 8. Mezzo-Level Interventions: Groups, Organisations and Neighbourhoods 8.1 Groupwork That Does Something Groupwork remains one of the most powerful tools available, but its effectiveness in cohesion terms depends heavily on design. Drawing on the contact literature discussed in Section 4.4, the groups most likely to build durable cross-group ties are those in which people work together on something that matters to them, over a sustained period, on roughly equal terms. In practice this means that a community garden, a repair workshop, a residents' committee that actually has a budget, a football league, a parents' group organising around a school issue, or a joint campaign about street lighting will usually outperform a diversity awareness session. The task supplies the reason to be in the room, the cooperation supplies the contact, and the sustained duration supplies the depth. Cohesion is a by-product rather than the stated aim, which is precisely why it works. The design principles are worth stating plainly. Give the group a real task with a real outcome. Give it a budget or a decision, however small, so that participation has consequences. Ensure that the different groups present have equal standing within the activity, which usually requires deliberate attention to who chairs, who speaks and in what language. Keep it going long enough for relationships to form, which usually means months. And build in redundancy so that the group survives departures. 8.2 Arrival Infrastructures and the Value of Ordinary Places One of the more useful recent contributions is the attention paid to what has been called arrival infrastructure: the ecosystem of places and organisations that newcomers actually use, whether or not those places were designed for them. Wessendorf (2022) describes the public library functioning as a central resource for newcomers in a diverse London district, offering not only information and internet access but warmth, safety, a legitimate reason to be somewhere without spending money, and informal encounters with staff and other users that lead to further help. The lesson generalises. In most hubs, the institutions doing the heaviest integration work are not the ones labelled as integration services. They are libraries, health clinics, schools, places of worship, barbershops, corner shops, laundries, transport hubs, and the informal advice networks that run through all of them. These places share features that formal services often lack: they are open, they do not require eligibility, they do not require disclosure of status, they are used by everyone, and they are already there. For social work, this suggests a shift in strategy. Rather than building parallel services that newcomers must find, invest in the places they already go. Put a worker in the library. Train the clinic reception staff. Support the school as a community hub. Fund the community centre's opening hours rather than its programmes. This approach has the additional advantage of serving established residents at the same time, which addresses the reciprocal recognition pillar and reduces the resentment that migrant-only services can generate. 8.3 Migrant-Led and Faith-Based Organisations In most high-mobility hubs, the densest support networks are run by migrants themselves: national associations, hometown groups, religious congregations, professional networks and informal savings groups. These bodies do enormous work and are frequently invisible to official service maps. The professional relationship to them is delicate. They can be a route to communities that formal services cannot reach, and their trust is a genuine asset. But they can also reproduce internal hierarchies, exclude women or minorities within the group, and, in some cases, be tied to employers or to home-country politics in ways that compromise their independence. Bonding capital, as noted earlier, is not automatically benign. The practical stance that seems most defensible is partnership with clear eyes: work with such organisations, resource them properly rather than extracting free labour from them, be honest about the limits of what they can be asked to do, and do not treat any single organisation as the authorised voice of a national group that in reality contains many voices. Support the emergence of newer, more inclusive groups alongside established ones. And be especially careful when the organisation depends financially on an employer whose workers are the intended beneficiaries. 8.4 Working with Established Residents Cohesion work that focuses only on newcomers will fail politically and often practically. In many hubs, longer-established residents, whether citizens or long-settled migrants, experience rapid demographic change alongside rising housing costs, congested services and a sense that the place they knew is disappearing. Some of the resulting resentment is prejudice. A good deal of it is a rational response to real material pressure, and treating it as pure bigotry both misdiagnoses it and guarantees that it will worsen. Serious #community_cohesion practice therefore engages this group directly. It listens to grievances about housing and services without either dismissing them or endorsing the conclusion that migrants are the cause. It ensures that community facilities benefit everyone. It creates roles for long-established residents in welcoming work, which turns a threatened identity into a valued one. And it makes the case, patiently and repeatedly, that the scarcity of housing and school places is a question of investment and planning rather than of who is allowed to live in the city. 8.5 Schools as the Central Institution If one institution deserves priority in a high-mobility hub, it is the school. Schools are compulsory, universal within their catchment, and they bring together adults who would never otherwise meet. The school gate is one of the few reliable sites of repeated cross-group contact between adults in an entire city. The interventions that follow from this are straightforward and underused. Support parent groups that are genuinely accessible, meaning interpretation, timing that fits working hours, and childcare. Use the school as a venue for adult language classes and community activities in the evening. Train teachers in the specific needs of children with interrupted education and of children who will leave within two years. Address the particular position of children who arrive mid-year, repeatedly, and who may attend five schools in eight years. And recognise the emotional labour carried by children who interpret for their parents and who often manage the family's relationship with institutions from a young age (Chase and Allsopp, 2020). 9. Macro-Level Interventions: Policy, Planning and Advocacy 9.1 Urban Planning as Integration Policy The physical form of a city determines who meets whom. Where housing for low-wage migrant workers is located far from the city, connected by inadequate transport and surrounded by nothing, isolation is not an unfortunate outcome but a designed one. No amount of community programming will overcome a two-hour commute and a curfew. Social work has traditionally kept out of planning, treating it as a technical field belonging to others. In high-mobility hubs this is a serious mistake, because planning decisions are integration decisions. The location of worker accommodation, the provision of public space, the routing of buses, the presence of shade and seating in hot climates, the availability of places to gather that do not require spending money, and the mixing or separation of housing types all determine the possibility of contact before any programme begins (UN-Habitat, 2022). The professional contribution here is evidentiary and advocative. Social workers see the consequences of planning decisions before anyone else does, and they are well placed to document them and to put them in front of the people who make those decisions. 9.2 Advocacy on Legal Status and Labour Rights Certain structural arrangements are so damaging to community life that no community intervention can compensate for them. Residence permits tied to a single employer create a relationship in which the worker cannot leave, cannot complain and cannot bargain, because the cost of doing so is deportation. Recruitment debt, incurred to obtain the job in the first place, deepens the trap. Wage theft, passport confiscation, and unsafe conditions all follow from the imbalance of power, and all corrode any possibility of the worker participating in the life of the city. Where such conditions exist, the highest-value intervention available to the profession is #structural_advocacy: reform of employer-tied residence, effective and accessible complaint mechanisms that do not risk the complainant's status, enforcement of wage protection, and regulation of recruitment fees. Reform in this area has been slow and uneven, and progress in law has often outpaced progress in enforcement. But it is the terrain on which the substantive question is decided. Advocacy of this kind is politically constrained in many hubs, and it would be dishonest to pretend otherwise. Practitioners who work in such settings must make careful judgements about what can be said, by whom, and through what channel. Documentation, professional associations, quiet institutional argument, service data, and international professional bodies all offer routes that individual public confrontation may not. What is not defensible is the pretence that the structural problem is a personal one. 9.3 Access to Services Regardless of Status A recurring and solvable problem is the exclusion of people with irregular or dependent status from basic services, and the fear of accessing services even where access is technically permitted. When people believe that attending a clinic or reporting a crime will bring immigration consequences, they stop doing both. The result is worse public health, more unreported crime, and a population that has learned to avoid the state. The intervention is a firewall: a clear, credible and publicised separation between service provision and immigration enforcement. Health, education, child protection and policing should be reachable without status consequences, and people must believe this to be true, which requires consistent behaviour over years, not a policy statement. Where such firewalls have been implemented at the municipal level, they have improved service uptake. They are also, notably, in the interest of the whole city rather than of migrants alone, which makes them politically arguable on grounds other than compassion. 9.4 Participation and Voice Without Citizenship If a large share of residents cannot vote, the city has a democratic problem regardless of how well its services perform. The people most affected by decisions have no formal say in them. Several partial remedies exist. Municipal voting rights for long-term residents, which some cities have adopted. Migrant advisory councils with genuine access to decision-makers, rather than consultative bodies with no power, which quickly become tokenistic and breed cynicism. Participatory budgeting, which allows residents to make binding decisions about small amounts of public money and, crucially, does not require citizenship to participate. Workplace representation. And the inclusion of non-citizens on the boards of the organisations that serve them, which is the least that #co_production should mean (Beresford, 2021). The test of any such mechanism is simple: does it change a decision that would otherwise have been made differently? If it does not, it is decoration. 9.5 Coordination and the Fragmentation Problem A final macro issue is that services for mobile populations are typically fragmented across employers, municipalities, national ministries, charities, faith organisations and consulates, with no one holding the whole picture. Newcomers are left to navigate this maze themselves, at exactly the point when they are least equipped to do so. Practical remedies are unexciting and effective: a single accessible entry point, whether physical or digital, shared referral protocols, common data standards that allow a person's situation to be understood without repeated retelling, and a designated body with the mandate and budget to coordinate. The absence of coordination is not usually a mystery. It is a consequence of the fact that no one is responsible for it. 10. Digital, Infrastructural and Spatial Dimensions 10.1 The Smartphone as Infrastructure For most migrants, the mobile phone is not a convenience but the primary infrastructure of their life: the means of contacting family, sending money, finding work, learning language, navigating the city, and accessing information about rights. The literature on migrants and digital media has documented the extent to which digital connection sustains both transnational family life and local coping (Alencar, 2020; Udwan, Leurs and Alencar, 2020). The practice implications are direct. #Digital_inclusion is a core integration issue, not a technical afterthought. Data costs, device access, digital literacy and language of interface are all barriers with material consequences. Services delivered exclusively through apps that require high literacy, a local bank account and a permanent address will exclude precisely the people who most need them. At the same time, digital tools should not be romanticised. Online groups can be a lifeline and a source of misinformation. Employers and states can use digital systems for surveillance. Recruitment scams operate through the same channels as community support. And digital connection to home, valuable as it is, does not substitute for local relationships; there is some evidence that it can reduce the pressure to build them, though this varies greatly between individuals. 10.2 Public Space and Everyday Encounter Section 5.3 argued that emplacement happens through ordinary places. This deserves practical elaboration, because it is one of the areas where small interventions have disproportionate effects. The conditions that support informal encounter are well established: places to sit, shade or shelter, no requirement to spend money, safety at the times when people are actually free, accessible transport, and proximity to routes people already travel. In hot climates, shade and water are not amenities but preconditions for any outdoor public life at all. In cities where low-wage workers have one day off a week, the availability of somewhere to go on that day, other than a shopping mall where they may be unwelcome, determines whether that day is restorative or merely empty. Where public space is scarce, hostile or effectively privatised, informal encounter collapses, and with it the main mechanism by which strangers become familiar. Arguing for benches sounds trivial. It is not. 10.3 Housing and Residential Patterns Housing shapes cohesion in two ways: through location and through security. On location, the debate about residential concentration has moved on from the simple assumption that clustering is bad. Concentrations of co-nationals provide support, shops, language, religious institutions and job information, and for newcomers they can be a resource rather than a trap. What is harmful is not clustering itself but enforced clustering combined with poor services, poor transport and no exit. The policy question is therefore not how to disperse people but how to ensure that neighbourhoods are well served and well connected, and that people who wish to move elsewhere are able to. On security, the tie between housing and employment is the critical variable. Where accommodation is provided by the employer, the worker who loses the job loses the home on the same day, which makes any complaint effectively impossible. Separating housing from employment, or at minimum providing a period of grace and an independent route to alternative accommodation, is among the highest-leverage reforms available in many hubs. 11. Measuring What Matters 11.1 The Problem With Existing Indicators Standard integration indicator sets measure employment rates, income, education, housing, health and, in more developed frameworks, social connections and civic participation, usually comparing migrants to the native-born (OECD and European Commission, 2023; Ndofor-Tah et al., 2019). These are useful and should not be abandoned. But they carry two assumptions that fail in a high-mobility hub. They assume that longer residence should produce convergence with the native-born population, which is impossible where legal status prevents it regardless of effort. And they treat departure as an absence rather than an outcome, so that a person who leaves simply drops out of the data. This means that the standard indicators are structurally incapable of recording the most common trajectory in the city they are being used to assess. 11.2 An Alternative Indicator Set An anchored mobility approach would supplement the standard measures with indicators of five further kinds. Speed of anchoring. How long after arrival does a person acquire a first local relationship outside their own national group, a first regular place they go, a first role in any organisation. Time-to-first-tie is a measurable and highly informative quantity that almost no one currently collects. Perceived fairness. Do residents believe the rules apply equally to them. This is measurable through straightforward survey items, and it should be disaggregated by status, nationality and occupation, since the aggregate figure will conceal exactly the disparities that matter. Institutional trust and use. Which institutions do people trust enough to actually approach, particularly police, health services and complaint mechanisms. Under-use by a particular group is a warning sign, not a sign that the group has no problems. Exit quality. What happens to people when they leave. Do they leave with wages paid, savings intact, skills certified, health preserved, and documentation of their work history. A city that treats departure as a success condition rather than as an administrative disappearance is a city that takes its temporary residents seriously. Cross-group tie density. Not merely whether people report having friends from other groups, which is subject to social desirability bias, but network measures of actual contact, help exchanged and cooperation. 11.3 Evaluating Interventions Evaluation in this field suffers from a familiar set of weaknesses: small samples, no comparison group, outcomes measured immediately after the activity, attitude measures rather than behaviour, and publication only of positive results. The prejudice reduction literature reviewed by Paluck and colleagues (2021) makes these problems explicit. Improvements are available even for small organisations with small budgets. Measure behaviour where possible, meaning who actually turned up, who came back, who helped whom, rather than only what people said on a feedback form. Follow up after months rather than minutes. Use comparison groups where ethically possible, including simple waiting-list designs. Report failures, since a field in which every reported intervention works is a field that is not learning. And accept that in a high-turnover population, attrition is not merely a methodological nuisance; it is the phenomenon under study, and it should be analysed rather than treated as missing data. 12. Ethical Tensions in Practice 12.1 Between Employment and Advocacy The practitioner who works in a system that produces harm is in a position that no ethical code resolves cleanly. Refusing to work within it abandons the people inside it. Working within it risks legitimising it. Most practitioners live somewhere between the two, doing what good they can while trying not to lie about what they see. This tension should be taught openly rather than smoothed over, because students who enter the field expecting clean answers will be unprepared for what they find. 12.2 Between Confidentiality and Reporting Where service data can be accessed by immigration authorities, confidentiality is not a professional nicety but a matter of a person's safety. Practitioners must understand precisely what happens to the information they record, and must be honest with service users about the limits of confidentiality rather than offering assurances they cannot keep. Where the system compels disclosure that harms clients, that fact should be documented and contested through professional channels. 12.3 Between Respect for Communities and Protection of Individuals Working through community organisations and community leaders is efficient and often necessary. It also risks deferring to hierarchies that disadvantage women, young people, sexual minorities and lower-caste or lower-status members within a group. Cultural respect does not extend to accepting harm to individuals, and a practitioner who allows a community leader to speak for a woman who has not been asked has not been respectful; she has been convenient. The resolution is to maintain independent routes for individuals alongside community partnerships, so that no one is reachable only through a gatekeeper. 12.4 Between the Present and the Future Finally, there is a tension between meeting immediate need and changing the conditions that produce it. Both are necessary; neither substitutes for the other. The risk in high-pressure services is that immediate need consumes everything, and the structural work is postponed indefinitely. Building even a small, protected allocation of time for documentation and advocacy is a way of ensuring that the profession does not become purely a repair service for damage that it is uniquely well placed to explain. 13. Implications for Social Work Education and Training The competencies required in a high-mobility hub differ enough from generalist practice to justify explicit teaching. Six areas stand out. Migration law literacy. Practitioners do not need to be lawyers, but they must understand visa categories, the relationship between employment and residence, family reunification rules and access entitlements. Without this, they will give wrong advice with serious consequences. Structural analysis. Students must be able to trace an individual's presenting problem back to the policy or market condition that produced it, and to distinguish reliably between a personal difficulty and a structural one. Community development skills. Group facilitation, asset mapping, participatory methods, coalition-building and negotiation with institutions are as important as casework and are frequently taught less well. Working with interpreters. This is a specific and teachable skill, and doing it badly damages both the assessment and the relationship. Cultural humility rather than cultural competence. The stance of not knowing, and of asking, is more useful and more honest than a catalogue of national characteristics. Reflexivity about position. Practitioners in these settings frequently occupy positions of relative privilege in relation to the people they serve, and may themselves be migrants of a different class. Understanding how one's own status, nationality and language shape the encounter is not an academic exercise. To these should be added a general disposition: a tolerance for working in situations where the presenting problem cannot be solved by the practitioner, and where the honest answer to a service user's question is that the rule is unjust and cannot currently be changed. Sustaining ethical practice under that condition, without either burning out or becoming cynical, is perhaps the hardest thing the profession asks of anyone working in this field. It should be prepared for, supervised for, and supported. 14. Limitations and a Research Agenda 14.1 Limitations This article is conceptual, and its framework is untested. The literature it draws on is heavily weighted toward Europe, North America and Australasia, and the settings where the high-mobility hub is most extreme, particularly in the Gulf and parts of East and Southeast Asia, are comparatively under-researched, partly because research access is restricted. The article therefore risks importing assumptions from settlement societies into contexts where they do not apply, which is precisely the error it criticises. Readers should treat its claims about non-Western hubs as hypotheses rather than findings. The five-pillar framework is also, at present, an organising device rather than a validated model. It has not been operationalised, and its pillars may not be independent. 14.2 A Research Agenda Six priorities follow. First, longitudinal research that follows people through and out of high-mobility hubs, rather than sampling only those who remain. The current evidence base is systematically biased toward stayers, which is a serious problem when leavers are the majority. Second, direct measurement of time-to-first-tie and its determinants, which would give practice a usable early indicator. Third, rigorous evaluation of arrival infrastructure interventions, particularly library-based and school-based models, using designs strong enough to support causal claims. Fourth, research on the emotional consequences of repeated relational loss in high-turnover populations, an experience that is widely reported anecdotally and almost entirely absent from the formal literature. Fifth, comparative work across hub types, examining whether the same intervention performs differently in a legally open hub and a legally closed one. The suspicion advanced here, that structural advocacy dominates all other interventions where status is insecure, is testable. Sixth, research on established residents in rapidly changing cities, whose experience of #belonging is frequently discussed politically and rarely studied carefully. 15. Conclusion The high-mobility hub is not a temporary anomaly. It is a stable feature of the global economy, and the number of cities that resemble it is increasing rather than decreasing. Social work has largely responded to these cities with tools designed for somewhere else, and the results have been predictably modest. The argument of this article has been that the mismatch is conceptual before it is practical. As long as #integration is understood as the movement of newcomers toward a settled majority, the concept will misdescribe a city in which the majority is not settled and the newcomers are not permitted to arrive fully. What is needed is not a better version of the settlement model but a different question. The different question proposed here is one of fair conditions rather than eventual sameness. Can a person who will be here for three years live decently, safely and connectedly while they are here. Can they use the same institutions as everyone else. Can they complain without losing everything. Can they form a relationship, hold a role, contribute something that is recognised. Can they leave with what they came for, and with what they earned. A city that can answer yes to those questions is cohesive in the only sense that is achievable under conditions of permanent movement, and it is probably more cohesive than a city that speaks a great deal about shared values while quietly maintaining a category of residents who have no way in. For social work, this implies a shift of centre of gravity: from casework toward community development, from services toward infrastructure, from adaptation toward advocacy, and from a professional identity built around helping individuals cope toward one built around changing the conditions that make coping necessary. None of this is new to the profession. It is, in fact, where the profession began. The high-mobility hub simply makes it unavoidable again. References Alencar, A. (2020). Mobile communication and refugees: An analytical review of academic literature. Sociology Compass, 14(8), 1-13. Ager, A. and Strang, A. (2008). Understanding integration: A conceptual framework. Journal of Refugee Studies, 21(2), 166-191. Beresford, P. (2021). Participatory Ideology: From Exclusion to Involvement. Bristol: Policy Press. Bhatia, M. and Canning, V. (eds.) (2021). Stealing Time: Migration, Temporalities and State Violence. Cham: Palgrave Macmillan. Blackmore, R., Boyle, J. A., Fazel, M., Ranasinha, S., Gray, K. M., Fitzgerald, G., Misso, M. and Gibson-Helm, M. (2020). The prevalence of mental illness in refugees and asylum seekers: A systematic review and meta-analysis. PLOS Medicine, 17(9), e1003337. Bronfenbrenner, U. (1979). The Ecology of Human Development: Experiments by Nature and Design. Cambridge, MA: Harvard University Press. Caponio, T., Scholten, P. and Zapata-Barrero, R. (eds.) (2019). The Routledge Handbook of the Governance of Migration and Diversity in Cities. London: Routledge. Chase, E. and Allsopp, J. (2020). Youth Migration and the Politics of Wellbeing: Stories of Life in Transition. Bristol: Bristol University Press. de Haas, H. (2023). How Migration Really Works: A Factful Guide to the Most Divisive Issue in Politics. London: Viking. Dinesen, P. T., Schaeffer, M. and Sonderskov, K. M. (2020). Ethnic diversity and social trust: A narrative and meta-analytical review. Annual Review of Political Science, 23, 441-465. Erel, U. and Ryan, L. (2019). Migrant capitals: Proposing a multi-level spatio-temporal analytical framework. Sociology, 53(2), 246-263. Fonseca, X., Lukosch, S. and Brazier, F. (2019). Social cohesion revisited: A new definition and how to characterize it. Innovation: The European Journal of Social Science Research, 32(2), 231-253. Hall, S. M. (2021). The Migrant's Paradox: Street Livelihoods and Marginal Citizenship in Britain. Minneapolis: University of Minnesota Press. Healy, L. M. and Thomas, R. L. (2021). International Social Work: Professional Action in an Interdependent World. 3rd edn. New York: Oxford University Press. International Labour Organization (2021). ILO Global Estimates on International Migrant Workers: Results and Methodology. Geneva: International Labour Office. International Organization for Migration (2024). World Migration Report 2024. Geneva: IOM. Klarenbeek, L. M. (2021). Reconceptualising integration as a two-way process. Migration Studies, 9(3), 902-921. Ledwith, M. (2020). Community Development: A Critical and Radical Approach. 3rd edn. Bristol: Policy Press. Ndofor-Tah, C., Strang, A., Phillimore, J., Morrice, L., Michael, L., Wood, P. and Simmons, J. (2019). Home Office Indicators of Integration Framework 2019. London: Home Office. OECD and European Commission (2023). Indicators of Immigrant Integration 2023: Settling In. Paris: OECD Publishing. Paluck, E. L., Porat, R., Clark, C. S. and Green, D. P. (2021). Prejudice reduction: Progress and challenges. Annual Review of Psychology, 72, 533-560. Phillimore, J. (2021). Refugee-integration-opportunity structures: Shifting the focus from refugees to context. Journal of Refugee Studies, 34(2), 1946-1966. Ryan, L. (2023). Social Networks and Migration: Relocations, Relationships and Resources. Bristol: Bristol University Press. Scholten, P. (ed.) (2022). Introduction to Migration Studies: An Interactive Guide to the Literatures on Migration and Diversity. Cham: Springer. Scholten, P., Crul, M. and van de Laar, P. (eds.) (2019). Coming to Terms with Superdiversity: The Case of Rotterdam. Cham: Springer. Spencer, S. and Charsley, K. (2021). Reframing integration: Acknowledging and addressing five core critiques. Comparative Migration Studies, 9(1), 18. Strang, A. B. and Quinn, N. (2021). Integration or isolation? Refugees' social connections and wellbeing. Journal of Refugee Studies, 34(1), 328-353. Udwan, G., Leurs, K. and Alencar, A. (2020). Digital resilience tactics of Syrian refugees in the Netherlands: Social media for social support, health, and identity. Social Media and Society, 6(2), 1-11. UN-Habitat (2022). World Cities Report 2022: Envisaging the Future of Cities. Nairobi: United Nations Human Settlements Programme. Vertovec, S. (2023). Superdiversity: Migration and Social Complexity. London: Routledge. Wessendorf, S. (2022). The library is like a mother: Arrival infrastructures and migrant newcomers in East London. Migration Studies, 10(2), 172-189. Wessendorf, S. and Phillimore, J. (2019). New migrants' social integration, embedding and emplacement in superdiverse contexts. Sociology, 53(1), 123-138. Yuval-Davis, N., Wemyss, G. and Cassidy, K. (2019). Bordering. Cambridge: Polity Press. Zapata-Barrero, R. (2019). Intercultural Citizenship in the Post-Multicultural Era. London: Sage. #Community_Integration #High_Mobility_Cities #Social_Cohesion_Research #Social_Work_Practice #Urban_Belonging #Migration_Studies #Globalizing_Cities #Superdiversity_Studies #Temporary_Migration_Policy #Community_Development #Migrant_Wellbeing #Anchored_Mobility #Inclusive_Cities #Social_Work_Education #Cohesion_Interventions

Adult learners who move from vocational training into university degree programmes represent one of the fastest growing and least understood groups in contemporary higher education. They arrive with practical skill, workplace maturity and a strong sense of purpose, yet they often struggle with academic writing conventions, abstract theory, isolation, money worries and the quiet feeling that they do not belong. This article examines that transition through the lens of #social_work, a profession whose core commitments to the person in the environment, to advocacy and to structural change make it unusually well suited to the task. Using an integrative and conceptual review of scholarship published mainly between 2020 and 2026, the article maps six interlocking domains of difficulty: academic literacy, identity and belonging, material hardship, role conflict, institutional navigation and emotional wellbeing. It then argues that most university support systems respond to these domains separately, through fragmented services that assume a young, full time, school leaving student. Drawing on transition theory, ecological systems thinking and the recognition of prior learning literature, the article proposes an integrated practice framework, described here as the BRIDGE model, that positions the social worker as a coordinating figure across admissions, teaching, welfare and community services. The article closes with implications for #university_policy, for social work education and for future research, arguing that the transition from workshop to seminar room is not merely an academic adjustment but a social and biographical one that deserves professional support. Keywords: adult learners; vocational education; higher education transition; social work; widening participation; recognition of prior learning; sense of belonging; student retention 1. Introduction The idea that a person completes their education in one continuous block, between the ages of five and twenty one, and then works for forty years without returning to formal study, is no longer accurate anywhere in the world. Labour markets shift, technologies replace tasks, and industries that once promised lifelong employment shrink within a decade. In response, more and more people who first trained in a workshop, a salon, a clinic, a construction site or a kitchen now find themselves applying to a university. They may be twenty six or fifty six. They may be a qualified electrician who wants to become an engineer, a healthcare assistant who wants to become a nurse, an apprentice chef who wants to study hospitality management, or a childcare worker who wants a degree in early years education. What unites them is that they are #adult_learners entering a system that was not designed with them in mind. This article is about the specific and often invisible difficulty of that journey, and about the contribution that #social_work can make to it. The move from #vocational_training to #university_education is usually described in administrative language, as a matter of entry requirements, credit transfer and course prerequisites. That language captures very little of what actually happens. A person who has spent eight years being valued for what their hands can do walks into a lecture theatre and is suddenly valued for what they can argue in a three thousand word essay. A person who has been the confident senior technician in a team becomes the confused newcomer who does not know what a "literature review" is or why the reading list is so long. The technical word for this is #academic_transition. The lived word for it, as many students describe it, is disorientation. There is a substantial body of research on students moving from school to university. There is far less on those moving from #vocational_education into higher education, and less still that treats this shift as a social and psychological event rather than an administrative one. Where the topic is discussed, it is usually held within education studies. This article deliberately places it inside social work, for three reasons. First, the difficulties that adult learners face during this transition are rarely purely academic. They involve childcare, debt, employer flexibility, housing, family expectations, health, immigration status and self worth. These are the classic territory of the social work profession, not of the study skills tutor. Second, social work already possesses the theoretical vocabulary needed. The concept of the person in the environment, the ecological systems perspective, the strengths approach and the commitment to #advocacy give practitioners a way of seeing the student as a whole person embedded in overlapping systems rather than as a learner with a skills deficit. Third, social workers are increasingly employed within educational institutions, in student support teams, in disability services, in wellbeing units and in outreach and #widening_participation offices, yet their role in that setting remains poorly defined and undertheorised. The article therefore asks three questions. What actually happens to adult learners during the transition from vocational to university study? Why do current institutional responses so often fail them? And what would a social work informed model of support look like in practice? The article proceeds as follows. Section 2 sets out the global context of vocational to higher education pathways. Section 3 develops the theoretical framework. Section 4 explains the review method. Sections 5 to 10 present the six domains of transition difficulty. Section 11 introduces the BRIDGE practice framework. Sections 12 and 13 discuss implications and limitations, and Section 14 concludes. 2. Background: The Global Growth of Vocational to University Pathways 2.1 The scale of the phenomenon Vocational education is not a marginal part of the world's education systems. It is a mainstream route. Across member countries of the Organisation for Economic Co-operation and Development, roughly one in three adults aged between twenty five and thirty four hold a vocational qualification as their highest level of attainment, whether at upper secondary, post secondary non tertiary or short cycle tertiary level (OECD, 2023a). That is an enormous population, and a growing proportion of it eventually seeks a degree. Several forces push in this direction. Credential inflation means that jobs once open to a diploma holder now require a bachelor's degree. Professional regulation in fields such as nursing, social care, teaching and allied health has steadily raised the minimum qualification. Automation is hollowing out routine technical work while expanding demand for the analytical and supervisory roles that degrees are thought to signal. And there is a quieter, more personal motive: many people who left school early carry an unfinished sense of their own intellectual capacity, and want to test it. Policy has responded. Governments in Europe, Africa, Asia, Australia and the Americas have built qualification frameworks intended to make movement between vocational and academic tracks possible. Double degree and integrated pathway programmes now exist in several systems, allowing a student to combine a higher vocational qualification with a university degree in less time than it would take to complete them separately. National frameworks for #micro_credentials and for the #recognition_of_prior_learning have proliferated, at least on paper (OECD, 2023b). 2.2 The gap between policy and experience Yet the existence of a pathway does not mean the pathway is comfortable to walk. Research consistently finds that students entering university with vocational qualifications face constrained choices, cluster in lower tariff institutions, and achieve poorer degree outcomes on average than peers who arrived through academic routes (Katartzi and Hayward, 2020; Hastings and Noyes, 2023). Studies comparing degree outcomes by entry qualification have found meaningful differences that persist even after prior attainment is taken into account, which suggests that something about the transition itself, and not only the student's earlier grades, is doing damage (Hastings and Noyes, 2023). The reasons are structural rather than personal. Vocational curricula are built around competence, demonstration and workplace assessment. University curricula are built around argument, critique, citation and independent reading. These are different intellectual cultures, and the mismatch between them is not the student's fault. Yet it is very often experienced as a personal failing, which is precisely why it produces shame, and why it belongs in the field of #social_work rather than only in the field of pedagogy. 2.3 Who the adult learner actually is The category "adult learner" hides enormous variation. It includes the twenty three year old apprentice who never really left education, and the forty eight year old care worker returning after two decades. It includes single parents, informal carers, refugees whose earlier qualifications are not recognised, people with disabilities, veterans, and people who left school with a fractured or hostile relationship to formal learning. Common threads nevertheless appear. Adult learners typically hold multiple simultaneous roles: worker, parent, carer, partner, community member. They are usually paying for study out of constrained income. They tend to be highly motivated but time poor. They often have long gaps since their last formal study, and therefore rusty habits of writing, note taking and examination. They are more likely to study part time, at a distance, or in the evening, which reduces their contact with campus life. And they are more likely to leave before completing (Hayman et al., 2024). Two further features are worth naming. Many adult learners are #first_generation_students, meaning nobody in their immediate family has attended university, so they have no informal guide to how the system works. And many carry what might be called an occupational identity that is genuinely valuable but institutionally unrecognised. The theatre nurse knows things the second year student does not. The system rarely tells her so. 3. Theoretical Framework 3.1 Transition theory The most directly useful framework for this topic is Schlossberg's transition theory, which has been applied in recent empirical work on mature students entering English universities (Hayman et al., 2024). The theory proposes that how a person copes with a life transition depends on four resources, often called the four S factors: situation, self, support and strategies. Situation refers to the nature of the change: whether it was chosen or forced, whether it is happening alongside other stressors, whether the timing is good. Self refers to the person's own attributes, including confidence, prior experience of change, and #self_efficacy. Support refers to the networks available, both formal and informal. Strategies refers to the coping methods the person can actually deploy. The value of this model for our purposes is that it refuses to reduce transition to academic preparedness. A student may be entirely capable of the coursework and still fail the transition because their situation is unstable, their support has collapsed or their strategies are exhausted. Recent work using this lens has found that mature students are frequently academically engaged but socially and emotionally disengaged, a pattern that standard academic monitoring will never detect (Hayman et al., 2024). 3.2 The person in the environment and ecological systems thinking Social work's foundational orientation, the #person_in_environment perspective, insists that behaviour and distress cannot be understood apart from the systems in which a person lives. Applied here, this means the struggling adult learner is not simply an individual with a writing problem. She is a node in a web that includes a partner who resents the lost income, a manager who will not adjust the shift rota, a child with additional needs, a landlord raising rent, a lecturer who assumes everyone can attend a nine o'clock seminar, and an admissions system that discounted her prior experience. The ecological systems perspective, familiar throughout social work education, gives us a way to name these levels: the immediate setting, the connections between settings, the wider institutional and policy context, and the cultural beliefs that surround all of them. Intervention can occur at any level. This is the crucial difference between a social work reading of the problem and a purely educational one. The educational reading tends to intervene only at the level of the individual student, through study skills workshops. The social work reading permits intervention at the level of the timetable, the assessment policy, the employer and the funding regime. The same perspective animates school based social work, where practitioners are trained to see academic and behavioural difficulty as embedded in family, community and structural conditions rather than located inside the child. That professional habit of mind transfers directly to the university setting. 3.3 Habitus, capital and the hidden curriculum Bourdieusian concepts remain indispensable here, and continue to be used in current studies of vocational entrants. #Cultural_capital refers to the knowledge, dispositions and codes that a person carries and that a given field rewards. #Habitus refers to the deeply internalised sense of how to act, what is normal and what is possible. The vocational student's habitus was formed in a field that rewards punctuality, manual precision, deference to a supervisor, and visible productivity. The university field rewards a different set of things: scepticism towards authority, tolerance of ambiguity, willingness to speak in abstractions, and comfort with unstructured time. Neither set is superior. But only one is legitimated by the institution. What follows is that the difficulty is a mismatch of field and habitus, not a deficit in the person. Over the first year, students who persist typically acquire the new capital and adapt their habitus, which is why outcomes tend to improve in later years for those who survive year one. The task of support is therefore to keep people alive through the mismatch period, and to make the implicit rules explicit. This is often called the hidden curriculum: everything the university expects but never teaches. 3.4 Andragogy and self directed learning Adult learning theory adds a further layer. Adults typically want learning that connects to their existing experience, that is problem centred rather than subject centred, and that respects their autonomy. #Andragogy predicts that adult learners will resent being treated as empty vessels and will disengage when their prior knowledge is ignored. This has a sharp practical consequence. The single most reliable way to demoralise a returning vocational student is to teach them, in week one, something they have been doing competently for a decade, while simultaneously failing to teach them the academic conventions they genuinely lack. #Self_directed_learning is often assumed by universities to be a natural adult capacity. In fact it is a learned academic skill, and vocational training, with its structured tasks and immediate supervisory feedback, does not usually develop it. 3.5 A synthesis Putting these together, we can state the theoretical position of this article. The vocational to university transition is a socially structured process in which a person with valuable but unrecognised capital enters a field whose rules are unstated, at a point in life when their other obligations are heaviest, and with support systems that were designed for someone else. Failure in this process is therefore best read as a systems failure, and the appropriate professional response is one that works simultaneously on the person, the institution and the surrounding social conditions. That is a description of social work. 4. Method This article is an integrative and conceptual review rather than a report of new empirical data. Its purpose is synthesis: to bring together scattered literatures from education studies, social work, adult learning and student wellbeing research, and to build from them a practice framework. Literature was identified through structured searching of academic databases and reference chaining from key recent articles. Priority was given to peer reviewed journal articles and scholarly books published between 2020 and 2026, in English, that addressed at least one of the following: transitions from vocational or further education into higher education; the experiences of mature or non traditional students; the recognition of prior learning; social work practice in educational settings; and interventions such as #peer_mentoring or peer support in higher education. A small number of older theoretical works are cited where they are foundational and continue to be used in current scholarship. Sources were read thematically. Themes were generated inductively and then organised into six domains, which form Sections 5 to 10. Because the aim is conceptual rather than statistical, no meta analysis or effect size pooling was attempted, and no claim is made to exhaustive coverage of the field. Three limitations of the method should be acknowledged at the outset. The literature is dominated by studies from the United Kingdom, Australia, Northern Europe, South Africa and North America, so the picture is skewed towards wealthy and Anglophone systems. Much of the qualitative work involves small samples in single institutions. And there is very little research that directly evaluates social work led interventions in this specific transition, which is itself one of the article's central findings. 5. Domain One: Academic Literacy and the Shock of Written Argument 5.1 What changes Ask an adult learner in their first semester what has surprised them most and the answer is usually about writing. Vocational assessment tends to be practical, portfolio based, competency referenced and immediate. You either wired the circuit correctly or you did not. University assessment is discursive. You must construct an argument, support it with published evidence, acknowledge counter positions, and reference every claim in a format that appears arbitrary and unforgiving. #Academic_literacy is not one skill but a cluster. It includes reading long and dense texts strategically rather than word by word. It includes understanding what a "critical" discussion means, as opposed to a descriptive one. It includes #referencing conventions and the rules on #plagiarism, which are often explained once, quickly, in a document nobody reads. It includes the ability to write in a register that is formal but not pompous, and to structure an essay around a claim rather than a chronology. None of this is intellectually beyond a capable adult. All of it is unfamiliar. The gap between capacity and familiarity is where the damage occurs, because students interpret unfamiliarity as incapacity. 5.2 Why current support often fails Universities do provide #study_skills support. The problem is threefold. The support is usually generic, delivered as a one hour workshop on "essay writing" that treats all students as identical and none of them as adults with fifteen years of professional writing behind them in other genres. The support is usually voluntary and separately timetabled, which means it is inaccessible to the very students who need it most, because those students are working shifts, collecting children or commuting. And the support is usually framed as remedial. It is offered as help for those who are struggling. Adults with strong occupational identities are often reluctant to identify themselves as struggling, particularly in front of younger peers. The framing of the support therefore repels its intended audience. 5.3 What better practice looks like The evidence points towards embedding rather than bolting on. Academic literacy taught inside the discipline module, by the subject lecturer, using the discipline's own texts, is more effective than a detached workshop, because it makes visible the specific conventions of that field. It also points towards making the hidden curriculum explicit. Publishing an annotated example of a strong essay with a commentary explaining why each move works costs very little and demystifies a great deal. Explaining, plainly, what markers actually want when they write "be more critical" is more valuable than another session on paragraph structure. Finally, it points towards using rather than ignoring prior knowledge. The care worker's understanding of safeguarding is real knowledge. Asking her to write about it academically, and showing her how to translate what she knows into the required form, is both good pedagogy and a defence of her dignity. This is where #andragogy and #academic_writing meet. 6. Domain Two: Identity, Belonging and the Feeling of Being an Impostor 6.1 Belonging as a predictor #Sense_of_belonging has become one of the most examined constructs in higher education research, and for good reason. It is consistently associated with persistence, satisfaction and attainment, and it is consistently lower among non traditional learners. Belonging appears to be produced through two partly independent channels, academic engagement and social engagement, and both matter for #retention (Ahn and Davis, 2023). For adult learners the social channel is frequently blocked. Campus social life is built around a set of assumptions: that students are young, unattached, financially supported, living near campus, and free in the evening. The adult learner is often none of these things. She arrives ten minutes before the lecture and leaves the moment it ends because the nursery closes at six. Studies of mature students describe exactly this pattern: academic engagement without social or emotional engagement, and a first year marked by isolation (Hayman et al., 2024). Some mature students report that they do not particularly want campus friendship, because they already have established networks elsewhere. That deserves respect rather than correction. But the distinction matters: choosing solitude is different from being excluded from a community that was never designed to include you. 6.2 Impostor feelings #Imposter_syndrome is frequently reported by mature entrants, and is strongly linked to long breaks from study and to unfamiliarity with the academic environment. The internal script is recognisable. They made a mistake letting me in. Everyone else understands the reading. If I ask a question they will discover I do not belong here. Two features make these feelings especially corrosive in this population. First, adults have more to lose. They have told their family, their employer and themselves that this time it will work. Failure would be public. Second, many carry an older wound: an experience of school that told them, explicitly or implicitly, that they were not academic. University confirms that verdict at the first low grade. Social work has a well developed literature on shame, stigma and identity, and practitioners are trained to work with exactly this kind of internalised negative narrative. That expertise is currently underused in higher education settings. 6.3 Occupational identity as a resource, not a liability There is an important reframing available here. The adult learner is often described in terms of what they lack: academic preparation, digital fluency, free time. A strengths based reading inverts this. What the ex apprentice actually possesses is a set of capacities that most eighteen year olds do not have: reliability, time discipline, tolerance of difficulty, the ability to take criticism from a supervisor without collapsing, an understanding of how organisations really work, and a clear reason for being there. #Motivation among adult learners is typically high and instrumental. They are not there to find themselves. They are there for a specific reason, and they will endure a great deal to reach it. The institution that names these strengths out loud, in induction, in tutorials and in feedback, does something powerful. It tells the student that their previous life was not a detour but a foundation. This is not flattery. It is accurate, and it is protective of #confidence. 7. Domain Three: Money, Work and Material Precarity 7.1 The arithmetic of returning to study For most adult learners the decision to study is an economic gamble. They reduce their working hours, or stop working, and simultaneously incur tuition costs. Many are supporting dependants. Many are doing this in a period of high housing costs and inflation. #Financial_hardship is not a background condition in this population; it is often the decisive one. Studies of student attrition repeatedly find that external factors such as money and family obligation carry heavy weight in decisions to leave, alongside academic and institutional factors. Research on adult learners who left university programmes points to work commitments, financial concerns and health as major personal drivers of attrition, operating alongside institutional factors such as course design and the quality of the relationship with teaching staff. The relationship between money and #mental_health is bidirectional and vicious. Financial insecurity is strongly associated with anxiety, depression and emotional exhaustion, and financially insecure students are more likely to forgo healthcare, including psychological consultation, precisely because of cost. The student who most needs the counselling service is the one least able to afford the bus fare to reach it. 7.2 Employment as both enabler and threat Most adult learners work while studying. Employment funds the degree, so it cannot be given up. But employment also consumes exactly the resource the degree demands, which is uninterrupted thinking time. Shift work is particularly hostile to university study. A rota published two weeks in advance cannot be reconciled with a fixed seminar timetable. An unexpected overtime request cannot be refused by someone whose income depends on the manager's goodwill. Employers vary enormously in their willingness to accommodate study, and there is often no mechanism through which a university can negotiate with an employer on a student's behalf. This is a striking absence, because negotiating with employers, landlords and agencies on a client's behalf is elementary social work practice. 7.3 Practical responses Several responses are available and are within institutional control. Hardship funds should be simple, fast and non humiliating to access. Long application forms that require proof of destitution deter precisely the proud, self reliant adults who need them. Timetabling should be treated as a welfare intervention, not an administrative afterthought. Blocking teaching into fewer, longer days, publishing timetables far in advance, and recording lectures reduce material barriers more effectively than any wellbeing campaign. Assessment deadlines should be predictable and, where possible, spaced. A cluster of three deadlines in one week is survivable for a nineteen year old with no dependants. For a working parent it can end the degree. And #financial_advice should be proactive rather than reactive. Many adult learners do not know what they are entitled to. Benefit interactions, tax credits, childcare subsidies and employer training levies are genuinely complex. Explaining entitlements is a recognised social work function. 8. Domain Four: Role Conflict, Caring and Time 8.1 The multiplicity of roles The adult learner is rarely only a learner. She is also, simultaneously, a parent, a daughter caring for an ageing father, an employee, a partner and often a member of a community or faith organisation with its own obligations. #Caring_responsibilities have a measurable effect on study. National survey evidence from higher education systems shows that students with regular informal care duties report poorer study progress and greater loneliness, with worse outcomes for those carrying two or more hours of care per day. Caring is also gendered, which means that role conflict falls disproportionately on women, and disproportionately on the very cohorts most likely to be entering university from vocational and care sector backgrounds. 8.2 Guilt as an academic variable Something interesting appears repeatedly in qualitative studies of mature students: guilt. Time spent studying is time not spent with children. Time spent with children is time not spent studying. Neither role can be performed to the standard the person sets for herself, and the result is chronic guilt that expresses itself as #anxiety about academic performance. This is not a soft or peripheral finding. Guilt reduces concentration, delays help seeking, and makes students more likely to withdraw quietly rather than ask for an extension. It is also invisible to any system that measures only attendance and grades. 8.3 Family and the ecology of study Families can be the greatest asset an adult learner has, or the greatest obstacle. A partner who takes over the evening routine can make a degree possible. A partner who experiences the degree as a threat, as evidence that his wife is outgrowing him, can make it impossible. Adult children may be proud. Parents may consider the whole project foolish at her age. Working with families and with the ecology around a person is core social work territory. Yet universities almost never engage with the student's household at all. Induction events are for students. Communications are addressed to students. The system behaves as though the student were an isolated unit, when in fact her success is a household project. A modest but potentially powerful innovation would be to include partners and family members in a short, welcoming induction session that explains what the degree will demand, how long it will take, and what the household should expect. This costs almost nothing and treats the family as an ally rather than an irrelevance. 9. Domain Five: Institutional Navigation and the Recognition of Prior Learning 9.1 The university as a bureaucracy Universities are old, complicated organisations with their own vocabulary, their own hierarchies and their own unwritten rules. To an eighteen year old whose parents attended university, this is navigable. To a #first_generation adult learner, it can be baffling. Research on mature students consistently finds that they are unaware of the support services that exist for them. This is often reported as though it were the student's oversight. It is more accurately a failure of communication design. Services are advertised in places adult learners do not go, at times they are not on campus, in language that assumes prior familiarity, and under names that mean nothing. A student does not know she needs "academic mitigation" because she does not know the phrase exists. The result is that the students with the least #institutional_capital receive the least support, which is the exact inverse of what an equitable system would produce. 9.2 Recognition of prior learning The #recognition_of_prior_learning is the formal process by which a university assesses learning acquired outside a degree programme and grants credit or entry on that basis. It is the single most important structural mechanism in the vocational to university transition, and it is chronically underused. The literature identifies a set of recurrent problems. Processes are bureaucratic, involving special committees and unclear criteria, with inconsistent decisions between departments. Students are frequently unaware that the option exists, and where they are aware, they often lack the vocabulary to articulate and evidence what they already know. Staff time is not funded. And the whole apparatus is treated as an exception to normal admissions rather than as a designed feature of it, so that applicants must actively fight for access to it (OECD, 2023b). Well designed recognition, by contrast, has documented benefits. It shortens the route to qualification, reduces cost, improves #social_mobility by improving the socioeconomic position of successful applicants, and functions as a mechanism of redress by valuing forms of learning that the credential system has historically ignored (Conrad, 2022). There is also a psychological dimension that the policy literature tends to miss. When a university formally recognises what a person already knows, it says something about that person's worth. When it refuses to, or makes the process so tortuous that they abandon it, it says something too. #Recognition is not merely administrative. It is a statement about whose knowledge counts. 9.3 Advocacy as the missing function What is missing from most institutions is anybody whose job it is to walk beside the student through the bureaucracy. Someone who knows that the mitigation deadline is next Tuesday, that the disability service requires a specific form of evidence, that the hardship fund exists, that the module leader can be asked for an extension, and that the recognition of prior learning claim needs to be framed against the learning outcomes rather than the job description. That function has a name in social work. It is #advocacy, supported by #case_management. It is what social workers do for clients in health, housing and welfare systems every day. Its absence in higher education is not inevitable; it is a design choice. 10. Domain Six: Wellbeing, Mental Health and Digital Readiness 10.1 The wellbeing picture Transition into higher education is a documented period of elevated psychological distress, and students frequently conceal their struggles from both family and friends (Worsley, Harrison and Corcoran, 2021). For adult learners the concealment is often stronger, because they have positioned themselves in their family as the capable one, the one who was going to do this. The distress is not usually a clinical disorder. It is more often a normal response to sustained overload: too many demands, too little sleep, no margin for error. But sustained overload becomes clinical if it lasts long enough, and university terms last a long time. Counselling services in most systems are stretched, with more than a decade of rising demand and waiting lists that can extend beyond the length of the crisis. This has driven interest in alternative and lower intensity forms of support. 10.2 Peer support and peer mentoring The evidence on #peer_mentoring is encouraging without being miraculous. Systematic reviews find that peer mentoring in the study entry phase is associated with improved social and academic integration, stronger sense of belonging, reduced dropout intention and, in a number of studies, improved retention (Gehreke et al., 2024). Reviews of the wider benefits identify four consistent domains of gain: academic performance, retention, emotional and psychological wellbeing, and social integration (Le, Sok and Heng, 2024). Peer support interventions for student mental health show promise, though the review evidence also notes considerable variation in design and quality and the need for clearer implementation guidance (Pointon-Haas et al., 2024). Two design points matter for adult learners specifically. Mentors should be matched on life stage where possible, since mature students consistently report preferring to learn alongside peers with comparable life and work experience. And mentoring must be scheduled realistically, which usually means brief, online, and asynchronous rather than a weekly coffee at four in the afternoon. 10.3 Digital readiness There is a persistent assumption that digital exclusion is a problem of the elderly. In practice it affects adult learners across the age range, and it is more subtle than an absence of devices. Vocational training often uses technology narrowly: an attendance app, a portfolio system, a supplier database. University study requires a different digital repertoire: searching bibliographic databases with Boolean logic, managing reference software, navigating a virtual learning environment, submitting through plagiarism detection systems, joining synchronous video seminars, and now negotiating the contested terrain of generative artificial intelligence tools. Research with non traditional students in distance and blended programmes has found low levels of online learning readiness, compounded by the competing socioeconomic roles these students occupy. #Digital_readiness is therefore not a technical footnote. It is a genuine barrier, and one that is easily mistaken for lack of ability. 10.4 Technology, judgement and caution There is growing interest in algorithmic tools that flag students at risk of dropping out, and in decision support systems more broadly. The social work literature is appropriately cautious here. Scoping work on decision support algorithms in social work identifies real utilities alongside significant limitations, including opacity and the displacement of professional judgement (Jacobi and Christensen, 2023). Discussions in social work education have framed the challenge as a competency and critique dilemma: practitioners must be able to use these systems competently while retaining the capacity to criticise them (James et al., 2023). Applied to #early_warning_systems in universities, the caution translates cleanly. A dashboard that flags a student because she has not logged into the virtual learning environment for ten days is useful only if a human being then makes contact, with warmth and without accusation, and is empowered to do something about what they find. Without that, the flag is surveillance without service. 11. Towards a Social Work Model of Transition Support: The BRIDGE Framework 11.1 Why a new framework is needed The six domains above are not independent. Money worries produce anxiety, which impairs concentration, which produces a low grade, which confirms impostor feelings, which reduce help seeking, which allows a fixable problem to become a withdrawal. The chain is causal and it is fast. Universities, however, are organised as though the domains were separate. The writing centre handles literacy. The counselling service handles distress. The finance office handles money. The admissions office handles credit. Nobody handles the person. Each service is competent within its silo and none of them sees the chain. What is needed is a coordinating function with a whole person mandate. That is what social work offers, and what the framework below attempts to specify. It is presented as six components under the acronym BRIDGE: Before, Recognise, Integrate, Do, Guard, Evaluate. 11.2 B: Before arrival Transition support that begins in week one begins too late. Decisions that determine success are made months earlier: what to study, whether to reduce working hours, how childcare will be covered, whether to claim prior learning credit. Pre entry work should therefore include: honest information about workload expressed in hours per week, not in vague reassurance; a structured conversation about the household's capacity to absorb the change; early identification of #recognition_of_prior_learning opportunities with hands on help to prepare the claim; a benefits and finance check; and a short bridging or #foundation_year option for those who want to rebuild academic confidence before full immersion. Bridging provision is often criticised as an extra year of cost. It should instead be evaluated against the cost of withdrawal in year one, which is higher for everyone concerned. 11.3 R: Recognise the person The second component is a deliberate act of recognition, both formal and symbolic. Formally, this means functioning credit recognition, applied consistently and explained clearly, with a named person responsible for supporting the claim. Symbolically, it means an induction that names the strengths adult learners bring, that includes their families, and that explicitly tells them that confusion in the first term is universal and is not evidence of unsuitability. It means teaching staff who ask, in the first seminar, what people already know, and who then use it. #Recognition is cheap. Its absence is expensive. 11.4 I: Integrate the support The third component is the abolition of the silo. In practice, this means a single named contact for each adult learner in the first year, who holds an overview across academic, financial, personal and administrative domains, and who has the authority to make referrals and to negotiate with other parts of the institution. This is not a personal tutor as conventionally understood, because personal tutors are academics with heavy teaching loads and no training in welfare work. Nor is it a counsellor, because the role is not primarily therapeutic. It is closest to a case coordinator, and the profession that trains people to do exactly this is #social_work. Integration also means that #interprofessional_collaboration becomes routine rather than exceptional: the social worker, the module leader, the finance officer and the disability adviser sharing information appropriately and acting together, rather than each responding to the fragment of the person that reaches their desk. 11.5 D: Do the structural work The fourth component is the one most often omitted. It is intervention in the environment rather than the individual. This includes: negotiating with employers over shift flexibility for students on part time routes; challenging timetabling decisions that make attendance impossible for carers; pushing for assessment deadlines to be spaced; ensuring lecture capture is standard; arguing within the institution for hardship funds that can be released within days rather than months; and working with local authorities, childcare providers and community organisations to build the external scaffolding a student needs. It also includes political and policy #advocacy at the sector level, on funding for part time study, on the portability of vocational credit, and on the recognition of overseas qualifications. Social work has always understood that casework without structural change is an endless bailing out of a leaking boat. The same logic applies here. 11.6 G: Guard wellbeing The fifth component is proactive attention to wellbeing rather than reactive crisis response. Practically: brief, regular, low intensity check ins during the first year, timed to coincide with known pressure points such as the first assessment and the post reading week slump. Life stage matched #peer_mentoring. Clear, quick routes into counselling for those who need it, with recognition that adult learners often cannot attend during working hours. Normalising language throughout, so that seeking help is framed as competence rather than weakness. And attention to the specific emotional signature of this group: guilt, shame, fear of exposure, and the exhaustion of people who are doing three jobs at once. 11.7 E: Evaluate honestly The final component is evaluation, and it must be honest. Institutions have a strong incentive to report that their support services are excellent. The relevant question is not how many workshops were run but whether adult learners with vocational backgrounds completed at rates comparable to their peers, and whether their attainment was comparable. Data should be disaggregated by entry route, age, working status and caring status. #Student_voice should be gathered systematically, including from those who left, since the people who withdraw hold the most important information and are the least likely to be asked for it. Where interventions do not work, they should be discontinued rather than defended. This is #reflective_practice applied at the institutional level. 12. Implications 12.1 Implications for social work practice The first implication is that higher education should be understood as a legitimate and important field of social work practice, not a marginal one. Social workers are already present in universities in counselling teams, disability services, care experienced student programmes, safeguarding roles and basic needs initiatives. What is largely missing is a coherent professional identity for that work and a body of scholarship to support it. The second implication concerns skills. The work described in this article is recognisable social work: assessment, #case_management, advocacy, crisis response, group work, community liaison, and structural analysis. But it also requires specific institutional knowledge, including how credit systems work, what mitigation processes require, and how assessment regulations can be used in a student's favour. Practitioners entering this field need that knowledge explicitly, and it is not currently taught. The third implication concerns boundaries. There is a risk that a social worker embedded in a university becomes an agent of institutional retention targets rather than an advocate for the student. Retention and the student's interests usually align, but not always. Sometimes the right advice is to leave, or to pause, or to transfer. The profession's ethical commitment to self determination requires practitioners to be able to give that advice without institutional penalty, which means the role needs a degree of structural independence. 12.2 Implications for social work education Social work programmes themselves recruit heavily from vocational and practice backgrounds. Many social work students are former care workers, support workers and health assistants. This means the transition described in this article is happening inside social work classrooms constantly, and social work educators are unusually well placed to model good practice. Recent scholarship on social work education has examined how programmes build #professional_resilience, and how students understand it (Hitchcock et al., 2021; Hitchcock et al., 2024). This literature is relevant here in two ways. It is directly applicable to the resilience of adult learners in transition. And it is a reminder that resilience must not be used to individualise structural problems: telling a student to be resilient in the face of poverty is not support, it is abdication. Work on the professional preferences of graduating social work students has found that they expect a thorough introduction to their roles, a reasonable workload and continuous support, and that working conditions weigh heavily on their career decisions (Blome and Espvall, 2024). The parallel with adult learners in transition is exact. People need induction, manageable demands and continuous support. This is true of new social workers and it is true of new students, and in the case of social work students they are frequently the same people. There is also a strong case for #reflective_practice and #supervision within educational social work roles. Practitioners supporting students in distress, in poverty and in crisis need supervision as much as their colleagues in child protection do, and are less likely to receive it because their host institution does not understand the work. 12.3 Implications for institutional and national policy At institutional level, four changes would make a measurable difference. Recognition of prior learning should be a designed and funded route rather than a discretionary exception. Timetabling and assessment scheduling should be evaluated for their effect on students with caring and work responsibilities. Hardship support should be fast and dignified. And there should be a named professional function, appropriately staffed, responsible for the whole person rather than a fragment of them. At national level, the questions are about #funding and about parity. Part time study is frequently funded worse than full time study, which penalises exactly the population discussed here. Vocational credit is often poorly portable across institutions. And the cultural devaluation of vocational qualifications, which persists in many countries and which stereotypes vocational students as academically weak, does material damage to the confidence of people who enter university through this route (McGrath et al., 2020). Longitudinal data infrastructure matters too. Panel studies that follow people across education and employment transitions, such as national cohort studies of school to work pathways, have shown how much can be learned when transitions are tracked over time rather than sampled at a single point (Hupka-Brunner, Meyer and Gomensoro, 2023). Equivalent longitudinal attention to the vocational to university transition would substantially improve the evidence base. 12.4 Implications for teaching staff Much of what determines an adult learner's fate is decided not by services but by the ordinary behaviour of lecturers. Research on adult attrition has found that the quality of the educator and learner relationship and the design of the programme are pivotal, and are frequently overlooked in favour of explanations that locate the problem in the student's personal circumstances. Relational pedagogy, the idea that connection and mattering are central to learning rather than decorative additions to it, is directly relevant (Gravett, 2023). Small things matter enormously. Learning names. Responding to emails within a reasonable time. Not making jokes that assume everyone is twenty. Saying explicitly, in a lecture, that anyone who is confused should come and speak to you, and meaning it. There is a reciprocal benefit here as well. Work on practising social workers who teach in university classrooms has found that the teaching relationship develops the practitioner as well as the student (Taylor and James, 2023). Support is rarely one directional. Adult learners bring experience into the classroom that improves the education of everyone in it, including the teacher. 13. Limitations and Directions for Research This article has clear limits. It is a conceptual synthesis rather than an empirical study, so the BRIDGE framework is a proposal, not a validated intervention. The literature drawn on is geographically uneven and dominated by high income Anglophone systems. Much of the qualitative evidence rests on small, single institution samples. And the category "adult learner from a vocational background" collapses a great deal of internal diversity: a former apprentice engineer and a former care worker face different worlds, and the differences are likely to be significant. Four research directions follow. First, there is a need for longitudinal work that follows vocational entrants across the whole degree rather than sampling them once in first year. The available evidence suggests that those who survive year one often do well afterwards, but this needs testing properly. Second, there is almost no evaluation of social work led support in this context. Pilot programmes with control comparisons, measuring completion, attainment and wellbeing, would be genuinely novel. Third, the intersection of #recognition_of_prior_learning with belonging and identity is underexplored. The policy literature treats recognition as a technical process. The psychological consequences of having one's knowledge validated or dismissed have barely been examined. Fourth, research should attend to the families and households of adult learners, who are currently almost entirely absent from the literature despite being central to the outcome. 14. Conclusion The person who leaves a vocational trade for a university degree is doing something difficult and admirable. She is betting time she does not have and money she cannot spare on a future she cannot guarantee. She is doing it while raising children, holding a job and managing a household. And she is doing it inside an institution that, for all its stated commitment to #widening_participation, was built for somebody else. This article has argued that her difficulties are not deficits. They are the predictable product of a mismatch between the capital she carries and the field she has entered, compounded by material pressures that no amount of study skills training can touch. It has argued that university support systems fail her not because the individual services are bad but because they are fragmented, and because nobody in the institution holds responsibility for her as a whole person. And it has argued that social work, with its person in environment lens, its dual commitment to individual support and structural change, and its long practice of advocacy and coordination, is the profession best equipped to fill that gap. The BRIDGE framework offered here is one attempt to specify what that might look like: preparation before arrival, genuine recognition of prior learning, integrated rather than siloed support, structural work on the environment, proactive attention to wellbeing, and honest evaluation. None of this is exotic or expensive. It requires an institution to decide that #adult_learners are not an anomaly to be tolerated but a population to be planned for. That decision, more than any single intervention, is what determines whether the shift from the workshop to the seminar room is a door or a wall. The stakes are worth stating plainly. Every adult learner who withdraws in the first year is not simply a statistic in a retention report. She is a person who took a serious risk, was quietly failed by a system that never told her the rules, and who now carries a second piece of evidence that education is not for people like her. That outcome is preventable. Preventing it is #social_work. Hashtags #Social_Work_and_Academic_Transitions #Adult_Learners_in_Higher_Education #Vocational_to_University_Pathways #Supporting_Mature_Students #Transition_from_Vocational_Training #Widening_Participation_in_Universities #Recognition_of_Prior_Learning_Matters #Student_Belonging_and_Retention #Lifelong_Learning_for_Adults #Social_Work_in_Education #Nontraditional_Student_Support #Academic_Literacy_Development #Equity_and_Inclusion_in_Higher_Education #Second_Chance_Education #Transition_Support_Frameworks References Ahn, M. Y., and Davis, H. H. (2023). Students' sense of belonging and their socio-economic status in higher education: A quantitative approach. Teaching in Higher Education, 28(1), 136-149. https://doi.org/10.1080/13562517.2020.1778664 Blome, H., and Espvall, M. (2024). Social work students' professional preferences. Social Work Education, 43(6), 1713-1731. https://doi.org/10.1080/02615479.2023.2216209 Brooks, R., Gupta, A., and Jayadeva, S. (2021). Higher education students' aspirations for their post-university lives: Evidence from six European nations. Children's Geographies. https://doi.org/10.1080/14733285.2021.1934403 Byrne, D. (2022). A worked example of Braun and Clarke's approach to reflexive thematic analysis. Quality and Quantity, 56(3), 1391-1412. Conrad, D. (2022). Accreditation and recognition of prior learning in higher education. In Handbook of Open, Distance and Digital Education. Springer. https://doi.org/10.1007/978-981-19-0351-9_44-1 Gehreke, L., Schilling, H., and Kauffeld, S. (2024). Effectiveness of peer mentoring in the study entry phase: A systematic review. Review of Education, 12(1). https://doi.org/10.1002/rev3.3462 Gravett, K. (2023). Relational pedagogies: Connections and mattering in higher education. Bloomsbury Academic. Hastings, J., and Noyes, A. (2023). Predicting outcomes in sport and exercise science degrees: The effect of qualification pathways. Journal of Further and Higher Education, 47(10), 1337-1350. https://doi.org/10.1080/0309877X.2023.2244434 Hayman, R., Coyles, A., Wharton, K., and Borkoles, E. (2024). Navigating the first year at an English university: Exploring the experiences of mature students through the lens of transition theory. International Journal of Lifelong Education, 43(1), 39-51. https://doi.org/10.1080/02601370.2023.2297671 Hitchcock, C., Hughes, M., McPherson, L., and Whitaker, L. (2021). The role of education in developing students' professional resilience for social work practice: A systematic scoping review. The British Journal of Social Work, 51(7), 2361-2380. https://doi.org/10.1093/bjsw/bcaa054 Hitchcock, C., McPherson, L., Whitaker, L., and Hughes, M. (2024). Social work students' understanding of professional resilience. Social Work Education, 43(1), 122-139. https://doi.org/10.1080/02615479.2022.2076826 Hupka-Brunner, S., Meyer, T., and Gomensoro, A. (2023). Data from the Swiss TREE panel study (Transitions from Education to Employment). Journal of Open Psychology Data. https://doi.org/10.5334/jopd.97 Jacobi, C. B., and Christensen, M. (2023). Functions, utilities and limitations: A scoping study of decision support algorithms in social work. Journal of Evidence-Based Social Work, 20(3), 323-341. https://doi.org/10.1080/26408066.2022.2159777 James, P., Lal, J., Liao, A., Magee, L., and Soldatic, K. (2023). Algorithmic decision-making in social work practice and pedagogy: Confronting the competency and critique dilemma. Social Work Education. https://doi.org/10.1080/02615479.2023.2195425 Katartzi, E., and Hayward, G. (2020). Transitions to higher education: The case of students with vocational background. Studies in Higher Education, 45(12), 2371-2381. Le, H.-G., Sok, S., and Heng, K. (2024). The benefits of peer mentoring in higher education: Findings from a systematic review. Journal of Learning Development in Higher Education, 31. https://doi.org/10.47408/jldhe.vi31.1159 McGrath, S., Ramsarup, P., Zeelen, J., Wedekind, V., Allais, S., and Lotz-Sisitka, H. (2020). Vocational education and training for African development: A literature review. Journal of Vocational Education and Training, 72(4), 465-487. OECD. (2023a). Spotlight on vocational education and training: Findings from Education at a Glance 2023. OECD Publishing. https://doi.org/10.1787/acff263d-en OECD. (2023b). Recognition of prior learning: A practical guide for policy makers. OECD Publishing. Pointon-Haas, J., Waqar, L., Upsher, R., Foster, J., Byrom, N., and Oates, J. (2024). A systematic review of peer support interventions for student mental health and well-being in higher education. BJPsych Open, 10(1), e12. https://doi.org/10.1192/bjo.2023.603 Poobalan, A., Barrow, J., and Cleland, J. (2021). The support needs of postgraduate taught students. MedEdPublish, 10(1). Taylor, M., and James, S. (2023). How social work managers can benefit from teaching social work students in a university. Irish Journal of Academic Practice, 11(1). Worsley, J. D., Harrison, P., and Corcoran, R. (2021). Bridging the gap: Exploring the unique transition from home, school or college into university. Frontiers in Public Health, 9, 634285. https://doi.org/10.3389/fpubh.2021.634285

Artificial intelligence has moved from research laboratories into classrooms and city halls. Schools now use adaptive tutoring engines, automated grading, early-warning systems for dropout risk, and generative writing assistants. Municipal governments now use eligibility screening engines, service chatbots, sensor networks, predictive maintenance tools, and automated benefit decisions. These systems promise speed, personalisation, and lower cost. They also assume something that is not true for a large part of the world population: that every user has a reliable device, an affordable connection, the skills to operate the interface, the language the model was trained on, and the confidence to challenge a machine decision. Where that assumption fails, the technology does not simply help less. It actively transfers cost, delay, and risk onto the people who already carry the heaviest social burdens. This article argues that the #digital_divide has crossed a threshold. It is no longer only a gap in technology adoption. It has become a #social_crisis, because access to #artificial_intelligence now mediates access to education, income support, housing services, health referrals, transport, and legal status. The article synthesises recent scholarship on digital inequality, algorithmic fairness, educational technology, and digital government to build an integrated account of how AI creates a new and steeper layer of stratification, which we describe as the #AI_divide. It then proposes a five-pillar mitigation framework covering infrastructure and #affordability, capability and #AI_literacy, inclusive design and representative data, governance and #accountability, and guaranteed human channels. The framework is translated into concrete responsibilities for national governments, city administrations, schools and universities, technology vendors, and civil society, and is paired with a measurement scheme so that progress can be tracked rather than asserted. The article closes with limitations and a research agenda. The central claim is simple. AI will not close the divide on its own. Left to default market and administrative logic, it will widen it. Closing it is a design choice, a budget choice, and a political choice, and those choices must be made deliberately and soon. Keywords: digital divide, artificial intelligence, educational technology, digital government, algorithmic fairness, digital inclusion, administrative burden, social policy 1. Introduction 1.1 A gap that changed shape For most of the last thirty years, the phrase "digital divide" described a fairly simple problem. Some people had a computer and an internet connection. Others did not. The policy answer was equally simple in principle: build networks, subsidise devices, and open public access points. Progress on that front has been real. The International Telecommunication Union estimates that roughly six billion people, about three quarters of the world population, used the internet in 2025, up from a revised estimate of 5.8 billion the year before. Yet about 2.2 billion people remain offline, and the overwhelming majority of them live in low-income and middle-income countries. #connectivity growth is also slowing. The remaining unconnected population is the hardest and most expensive to reach, because it is concentrated in rural areas, in low-income households, among older adults, among women in several regions, and among people with disabilities. The more important development is not the headline number. It is the change in what being connected means. Coverage figures conceal enormous differences in quality. Mobile broadband signals now reach most of humanity, but a typical user in a high-income country consumes far more data than a user in a low-income country, and fifth-generation networks cover a large majority of people in rich states while reaching only a small fraction in the poorest. Mobile data remains unaffordable in a majority of low-income and middle-income countries. Basic digital skills are widespread among users, but the more advanced skills, such as safe online behaviour, problem solving, and content creation, are spreading far more slowly. In other words, the frontier of exclusion has moved from the question of whether a person can get online at all to the question of what that person can actually accomplish once online. Scholarship anticipated this shift. A widely cited systematic review of the field concludes that the divide should be understood through three distinct segments, covering access, usage, and outcomes, and that the factors that shape it cluster into around nine categories spanning sociodemographic, socioeconomic, personal, social, cultural, motivational, material, political, and geographic variables (Lythreatis, Singh, and El-Kassar, 2022). The same review notes that the COVID-19 period sharpened every one of these gaps and that new forms of the divide are emerging faster than the literature can classify them. That warning has aged well. The newest form is the subject of this article. 1.2 Why AI changes the stakes AI is not simply another application layer sitting on top of the internet. It changes the relationship between the citizen and the institution in four ways that matter for equity. First, it changes who decides. When a municipal system automatically awards a child benefit based on registry data, the decision is made without any human reading the file. Research on the Norwegian child benefit system found that most recipients were awarded the benefit automatically, but that some citizens fell outside the automated path and had to apply manually, and that low-income citizens were disproportionately among those forced into the manual route (Larsson, 2021). Automation was distributed unequally. The people with the least time, the least administrative confidence, and the greatest need were the ones handed the extra paperwork. This is a precise illustration of how a technology designed to reduce effort can concentrate effort on the vulnerable. Second, it changes what counts as a person. AI systems act on data representations. If a citizen is poorly represented in the registry, if a student's home language is absent from the training corpus, if a neighbourhood has few sensors, then the system does not see them clearly. Under-representation is not neutral. It produces worse predictions, more errors, and more friction for exactly the groups that are already marginal. Third, it changes the cost of contesting a decision. Challenging a caseworker is difficult. Challenging a model whose logic is proprietary, whose training data is undisclosed, and whose output is presented as objective is much harder. #transparency problems in public sector algorithms are well documented, and explanation strategies carry their own pitfalls, since poorly designed explanations can create a false sense of understanding rather than genuine oversight (de Bruijn, Warnier, and Janssen, 2022). Fourth, it changes the speed at which advantage compounds. A student with a reliable laptop, fast broadband, a paid AI assistant, and a teacher confident enough to guide its use can now compress hours of work into minutes and receive detailed feedback on demand. A student without those things falls behind faster than before, and the gap is measured not in years of technology adoption but in weeks of coursework. Taken together, these four shifts mean that the divide is no longer only about missing out on a useful tool. It is about being governed, taught, assessed, and served by systems that were not built with you in mind and that you cannot easily question. 1.3 The case for calling it a crisis The word crisis is used deliberately and should be defended rather than assumed. Three tests are proposed. A gap becomes a crisis when it is structural, meaning it reproduces itself without further intervention. #digital_capital, understood as the accumulated stock of digital access, skills, devices, support networks, and confidence, converts into educational and economic advantage, and that advantage in turn buys more digital capital (Lybeck, Koiranen, and Koivula, 2024). The loop does not need a villain to keep running. A gap becomes a crisis when it is cross-cutting, meaning it damages more than one domain of life at once. Digital exclusion today reduces a person's ability to learn, to claim a benefit, to book a clinic appointment, to contest a fine, to find work, and to be counted in the data that shapes future budgets. #social_inclusion and digital inclusion have become hard to separate. A gap becomes a crisis when it is accelerating. The distinguishing feature of the current moment is that the capability frontier is moving faster than the inclusion frontier. Every quarter brings more capable models, deployed into schools and city systems, while the underlying inequalities in devices, bandwidth, skills, and language coverage move slowly if at all. The distance between the connected and the excluded is a function of both positions, and one of them is sprinting. On all three tests, the current situation qualifies. This is the analytical basis for the argument that follows. 1.4 Scope, aims, and contribution This article focuses on two service domains that touch nearly every citizen and that are being reshaped by AI at the same time. Education. Compulsory schooling, vocational training, and higher education, including adaptive learning platforms, automated assessment, #learning_analytics, dropout prediction, admissions screening, remote proctoring, and generative writing and tutoring assistants. Municipal services. The services delivered by city and local government, including welfare and benefit administration, housing allocation, permits and licensing, waste and transport, complaint handling, public health outreach, and the front-door channels through which citizens reach all of these. These two domains are chosen because they are where the state most directly shapes life chances, because they are heavily local, and because they are where automation has advanced fastest under budget pressure. The article has four aims: To develop a conceptual account of the #AI_divide that extends existing three-level models of digital inequality. To document the specific mechanisms through which AI-driven educational and municipal systems generate inequitable outcomes. To propose an actionable mitigation framework with clear allocation of responsibility. To specify indicators that allow progress to be measured and audited. The contribution is integrative rather than empirical. It brings together three research communities that rarely cite one another: digital inequality sociology, AI ethics and fairness, and digital government and public administration. Each has part of the picture. Read together, they describe a single problem. 1.5 Structure Section 2 reviews the conceptual foundations. Section 3 sets out the method and the analytical framework. Section 4 examines the divide in education. Section 5 examines the divide in municipal services. Section 6 explains the compounding mechanisms that turn a gap into a crisis. Section 7 presents the mitigation framework. Section 8 translates it into responsibilities for specific actors. Section 9 proposes measurement. Section 10 states limitations and a research agenda. Section 11 concludes. 2. Conceptual Background 2.1 The three levels of the divide The literature on digital inequality has converged on a layered model. The first level concerns physical access: whether a person has a device, a connection, and electricity. The second level concerns usage: whether a person has the skills, the time, the autonomy of use, and the support network to make productive use of the connection. The third level concerns outcomes: whether use actually translates into tangible benefits such as better grades, higher income, better health, or successful interaction with the state. The systematic review by Lythreatis and colleagues (2022) organises the field along these segments and shows that the third level is the least studied and the most consequential. Two people can have identical devices and identical connections and derive radically different returns from them. What separates them is a bundle of social resources, which Helsper (2021) describes as the link between digital and social disconnection, and which she argues cannot be understood by looking at technology alone. This is why device-distribution programmes so often disappoint. Handing a tablet to a household solves a first-level problem. It does nothing about the second level, where a parent may lack the skills to help with homework, or the third level, where the school may not have a #curriculum_reform strategy that turns tablet ownership into learning. 2.2 Digital capital and social reproduction A productive theoretical move has been to treat digital resources as a form of capital. #digital_skills, device quality, connection quality, informal support, and confidence accumulate over time and are convertible into economic, cultural, and social capital. Recent empirical work confirms that education and digital skills together shape the depth of participation in online life and that the pathway from digital divide to digital capital runs directly through formal schooling (Lybeck, Koiranen, and Koivula, 2024). This framing matters for policy because it predicts that one-off interventions will decay. Capital must be maintained. A subsidised connection that lapses after a year, a donated laptop that breaks in eighteen months, a training course that is never refreshed, all of these leave the household roughly where it started while allowing the funder to report a success. 2.3 Populations at the sharp end Certain groups appear repeatedly in the evidence. Older adults. The so-called grey divide is persistent and is widening in relative terms as services move online faster than older users adopt them. Digital exclusion among elderly populations has been described as a systematic form of exclusion in the information age rather than a matter of individual preference (Mubarak and Suomi, 2022). #older_adults are also disproportionate users of exactly the municipal services that are being automated first, including pensions, health, and housing support. Rural and remote residents. #rural_connectivity gaps remain wide. Urban internet use runs well ahead of rural use globally, and the difference is largest in the poorest regions. Women. The #gender_digital_divide has proved stubborn. Global gender parity in internet use has barely moved in recent years, and the gap is widest in low-income countries and in parts of Africa. People with disabilities. #accessibility failures compound with AI, because interfaces that assume speech, vision, fine motor control, or standard reading fluency create new barriers even as they remove old ones. #disability_inclusion is often an afterthought in procurement. Migrants, refugees, and people with irregular status. These groups face the identity problem. Automated systems key on official records. People with incomplete, inconsistent, or foreign documentation are pushed out of the automated lane and into the manual queue. Low-income households. #low_income_households experience every other divide at higher intensity, and they experience the affordability barrier directly, since a data plan competes with food and rent. These categories overlap. A rural, elderly woman with a disability and a low income is not experiencing five separate divides. She is experiencing one compounded exclusion, and single-issue programmes will miss her. 2.4 The AI divide as a fourth layer The core conceptual claim of this article is that AI adds a fourth layer to the model, one that does not reduce to the first three. Layer four: the algorithmic layer. A person may have full access (level one), strong skills (level two), and good outcomes from ordinary internet use (level three), and still be disadvantaged by AI systems because of who they are in the data. This layer has four components. Representation. Whether the person's language, dialect, name, appearance, household structure, and circumstances are well represented in the model's training data. Language technology performance is systematically unequal across the world's languages, with a small number of high-resource languages absorbing the overwhelming share of research and engineering effort (Blasi, Anastasopoulos, and Neubig, 2022). A citizen who speaks a #low_resource_language is not merely underserved by an AI chatbot. She is misunderstood by it, and the errors are invisible to the agency that deployed it. Prediction. Whether the model's outputs are equally accurate across groups. The evidence on #algorithmic_bias in education is extensive and shows concrete measurable harms across race, ethnicity, gender, nationality, socioeconomic status, and other categories, occurring at multiple stages of the machine learning pipeline (Baker and Hawn, 2022). Bias is not an exotic failure mode. It is the default outcome of training on historical data that encodes historical inequity (Mehrabi, Morstatter, Saxena, Lerman, and Galstyan, 2021). Agency. Whether the person can understand, question, and override the system's decision. This depends on transparency rules, appeal rights, and the practical availability of a human being. Amplification. Whether the person can use AI as a tool to increase their own capability, rather than only being subject to it as an object of assessment. The student who uses a model to draft, critique, and revise is amplifying. The student whose essay is scanned by a detection tool is being assessed. The same technology, two positions. This fourth layer is what makes the present moment different. It is possible to close the first three gaps and still be structurally disadvantaged by the fourth. That is why #digital_inclusion policy that stops at devices and broadband is now insufficient. 2.5 What the AI ethics literature contributes The ethics of algorithms literature has mapped the problem space carefully, identifying inconclusive evidence, inscrutable evidence, misguided evidence, unfair outcomes, transformative effects, and traceability as recurring concerns (Tsamados, Aggarwal, Cowls, Morley, Roberts, Taddeo, and Floridi, 2022). Relational approaches argue that algorithmic injustice cannot be fixed by better mathematics alone, because the harm arises from the relationship between the system and the people it classifies, and therefore requires attention to power, context, and consent (Birhane, 2021). Work on large language models has warned that scale itself creates risk, since the size of the training corpus makes it impossible to fully audit what is inside it and whose worldview it encodes (Bender, Gebru, McMillan-Major, and Shmitchell, 2021). Floridi (2023) argues that the ethics of AI must be treated as a design discipline rather than a compliance exercise, and that the question of who benefits should be built into the specification rather than assessed after deployment. Work on AI for social good similarly stresses that beneficial outcomes are not automatic and must be engineered against clear criteria (Cowls, Tsamados, Taddeo, and Floridi, 2021). #responsible_AI, on this reading, is not a communications strategy. It is an engineering and procurement requirement. What this literature does not yet do consistently is connect these abstract harms to the concrete institutional settings in which most citizens actually meet AI: their child's school and their local council office. That is the gap this article addresses. 2.6 What the digital government literature contributes Public administration research has documented both the promise and the risk. A systematic review of AI in public governance found the field to be largely exploratory and conceptual, with limited empirical work on actual implications, and called for research on the risks to equity, accountability, and legitimacy (Zuiderwijk, Chen, and Salem, 2021). A later review of AI adoption and diffusion in public administration reached a similar conclusion, noting that adoption is proceeding faster than the evidence base (Madan and Ashok, 2023). Implementation research shows that public organisations frequently underestimate the organisational work required, treating AI as a procurement event rather than a transformation of practice (Mergel, Dickinson, Stenvall, and Gasco, 2023). Research on #trust_in_government and automation finds that citizens' willingness to accept automated decisions depends heavily on the assurance that a human remains in the decision loop (Aoki, 2021), and that algorithmic transparency has measurable but conditional effects on perceived trustworthiness (Grimmelikhuijsen, 2023). Studies of #human_in_the_loop arrangements also caution that human oversight is not a guaranteed safeguard, since officials can defer to algorithmic advice uncritically or apply it selectively in ways that track existing prejudices (Alon-Barkat and Busuioc, 2023). Meanwhile, the administrative burden literature has shown that digital self-service does not remove work, it relocates it. Citizens become, in effect, unpaid caseworkers who must learn the system, assemble the evidence, and navigate the interface, and those with the fewest resources bear the highest cost (Madsen, Lindgren, and Melin, 2022). This is the single most important insight for the argument here, because it explains precisely how a labour-saving technology can increase the total suffering in the system while reducing the agency's costs. 3. Method and Analytical Framework 3.1 Approach This is an integrative conceptual review. It does not report new primary data. Its purpose is to synthesise findings that currently sit in separate literatures and to build a framework that practitioners can act on. This method is appropriate when a field is fragmented, when the phenomenon is changing faster than empirical studies can be completed, and when the practical need for guidance is urgent. All three conditions hold here. 3.2 Sources The review draws on peer-reviewed journal articles, scholarly books, and official statistical and policy reports published mainly between 2021 and 2025, in four bodies of work: digital inequality and digital inclusion research, algorithmic fairness and AI ethics, educational technology and AI in education, and digital government and public administration. Priority was given to systematic reviews, empirical studies with clear methods, and authoritative statistical sources such as the annual connectivity measurements published by the International Telecommunication Union. 3.3 Analytical framework The framework used throughout the article has three dimensions. Dimension one: the four layers. Access, usage, outcomes, and the algorithmic layer described in Section 2.4. Every finding is located in one or more layers. Dimension two: the direction of the AI relationship. For each system, the question is asked whether the citizen is a user of AI (able to amplify their own capability), a subject of AI (assessed, sorted, or decided about), or both. Systems in which disadvantaged groups are predominantly subjects while advantaged groups are predominantly users are flagged as equity-hostile by design. Dimension three: the burden transfer. For each system, the question is asked who absorbs the cost when the system fails. If the answer is the citizen, the design has externalised its errors onto the public, and the more vulnerable the citizen, the higher the real cost of that externalisation. These three questions form a simple diagnostic that any school district or city department can apply to any proposed deployment, and they recur in the mitigation framework in Section 7. 4. The AI Divide in Education 4.1 The material floor Before any discussion of pedagogy, there is a hard material floor. AI-driven learning requires a device capable of running a modern browser, a connection fast enough for real-time interaction, electricity that does not fail during the lesson, and a physical space in which a child can concentrate. None of these can be assumed. The device gap is the most visible. Household surveys consistently show that shared smartphones are the primary or only internet device in a large share of low-income households. A smartphone is adequate for messaging and video. It is poor for extended writing, for coding, for working with a document alongside a tutoring assistant, and for any task requiring sustained composition. Students on phones are not slightly disadvantaged. They are performing a different, harder task. The bandwidth gap is less visible but equally real. Conversational AI tools are interactive and latency-sensitive. On a congested or metered connection, the experience degrades from a fluid dialogue into a series of frustrating waits, and metered data makes extended use financially painful. Where mobile broadband remains unaffordable for a majority of households, as it does across much of the low-income and middle-income world, the practical effect is rationing. Students self-limit. Teachers do not see this happening and interpret reduced use as reduced interest. The electricity gap is the least discussed in the AI literature and one of the most decisive in practice. #school_infrastructure in many regions includes intermittent power, which makes any always-on digital pedagogy unreliable and forces teachers back to paper, which is the correct professional judgement in the circumstances but which then leaves those students without the AI-supported practice that their peers elsewhere are receiving daily. The result is a #compute_divide inside education. Access to capable models, especially the paid tiers with higher limits and stronger reasoning, is now a purchasable advantage. Where an institution does not provide equal access to a high-quality tool, the market allocates it by household income, and the school unintentionally ratifies that allocation by setting assignments that reward its use. 4.2 The language and representation gap Educational AI systems perform best in the languages and registers that dominate their training data. Performance across the world's languages is systematically unequal, and the inequality tracks economic and research investment rather than the number of speakers (Blasi, Anastasopoulos, and Neubig, 2022). A tutoring assistant that is fluent, patient, and accurate in one language may be clumsy, wrong, or simply unavailable in another. The consequences for education are direct. A learner working in a #low_resource_language receives lower-quality explanations, more factual errors, and fewer available resources. A learner whose dialect differs from the standard written form may find that automated writing feedback marks their natural constructions as errors, teaching them that their own speech community is wrong. Automated scoring systems trained largely on the writing of majority-language students can systematically undervalue the work of second-language writers, whose prose may be structurally sound but stylistically unfamiliar to the model. This is a #training_data problem, and it cannot be fixed by the school. It can only be mitigated by the school, by insisting on human review of automated scores for affected groups, and it can only be solved upstream, by investment in corpora, evaluation sets, and model development for underserved languages. That upstream investment is a public good and is therefore unlikely to be supplied at adequate scale by commercial actors alone. 4.3 Algorithmic bias in learning systems The empirical record on #algorithmic_bias in education is now substantial. The comprehensive review by Baker and Hawn (2022) documents bias affecting groups defined by race and ethnicity, gender, nationality, language, socioeconomic status, disability, and military-connected status, and traces how bias enters at each stage of the pipeline, from the choice of what to measure, through data collection and labelling, to model selection and deployment. Several patterns deserve emphasis for practitioners. Dropout and risk prediction. #dropout_prediction models are widely deployed in higher education and increasingly in schools. They are trained on historical outcomes. If students from a particular background historically left at higher rates because the institution served them badly, the model learns to flag them as risky. The flag can then trigger interventions that range from helpful (extra tutoring) to harmful (steering into easier courses, exclusion from selective programmes, or a self-fulfilling reduction in staff expectations). The model has learned the institution's past failure and is now recommending that it be repeated. Automated assessment. Essay scoring and short-answer grading systems can show differential accuracy across groups. Where a scoring model is less accurate for a group, that group is systematically mis-graded, and because the score carries institutional authority, the error is difficult to contest. Remote #proctoring. Systems that monitor students during examinations have been shown to generate more false flags for students with darker skin tones, for students with movement disorders or neurodivergent behaviours, for students who share a room with family members, and for students with unstable connections. Each false flag is an accusation of cheating. The burden of disproving it falls on the student. Recommendation and pathway systems. #adaptive_learning engines that route students to content based on predicted ability can lock in early misjudgements. A student who performs poorly in the first week because of a broken device may be routed to easier material for a year. The pattern across all four is the same. The student who is already disadvantaged is more likely to be misread by the system, and the misreading has institutional force. 4.4 The teacher capacity gap The most under-resourced variable in the entire equation is the teacher. AI in education does not work without professional judgement wrapped around it. Reviews of the field consistently find that the technology is developing far ahead of the pedagogical and ethical infrastructure needed to use it responsibly (Zhang and Aslan, 2021; Crompton and Burke, 2023). A large meta-review of AI in higher education found persistent weaknesses in ethical treatment, in collaboration between technologists and educators, and in methodological rigour, and called explicitly for these gaps to be closed (Bond, Khosravi, De Laat, and colleagues, 2024). #teacher_training is unevenly distributed in exactly the way one would predict. Well-funded schools release staff for professional development, purchase institutional licences, and hire specialists. Under-funded schools do not. The teacher in the under-funded school is then blamed for low adoption, when the real constraint is time and support. Selwyn (2022) offers a necessary caution here. The debate about AI in education is often conducted in the language of inevitability, which forecloses the political questions of who benefits, who pays, and who decides. Treating adoption as a technical necessity rather than a policy choice removes the very deliberation that equity requires. #teacher_capacity is not a barrier to be overcome on the way to adoption. It is the condition on which responsible adoption depends. 4.5 Generative AI and the acceleration of the gap Generative systems changed the picture in a specific and important way: they made a genuinely powerful cognitive tool available at low marginal cost, and simultaneously made cheating anxiety the dominant institutional response. The upside is real. Reviews of large language models in education identify substantial potential benefits, including personalised explanation, immediate feedback, language support, and assistance for learners with disabilities, alongside serious risks around accuracy, bias, over-reliance, and #academic_integrity (Kasneci and colleagues, 2023). Student perspectives show a similarly mixed picture, with strong interest tempered by concerns about fairness, accuracy, and unclear institutional rules (Chan and Hu, 2023). The equity problem lies in how institutions have responded. Where the response has been prohibition and detection, the effect has been regressive. Detection tools produce false positives, and the students least able to defend themselves against a false accusation are those without advocates, without confidence, and without the cultural fluency to argue with authority. Second-language writers are flagged more often. Meanwhile, students at well-resourced institutions with permissive, well-designed policies learn to use the tools skilfully and enter the labour market with an advantage. The result is the amplification asymmetry described in Section 2.4. Advantaged students become users of #generative_AI. Disadvantaged students become subjects of it. Bulathwela and colleagues (2024) make the point sharply: AI will not democratise education by itself, and a techno-solutionist framing actively obscures the structural inequalities that determine who benefits. The correct institutional response is neither prohibition nor laissez-faire. It is universal provision plus explicit instruction plus assessment redesign. Provide every student with equal access to a capable tool. Teach its use, its failure modes, and its ethics as a core competence. Redesign assessment so that it measures what the student can do with the tool and what the student can do without it, deliberately and separately. Any policy that leaves access to the market will reproduce household income in the classroom. 4.6 Summary of the education layer Applying the three diagnostic questions from Section 3.3: Layers affected: all four. Access (devices, bandwidth, power), usage (skills, teacher support), outcomes (grades, progression, confidence), and algorithmic (bias, language, detection). User or subject: advantaged students are increasingly users; disadvantaged students are increasingly subjects. Burden transfer: when the system errs, the student absorbs the cost, in the form of a wrong grade, a false accusation, or a diminished pathway. By its own diagnostic, the current default configuration of AI in education is equity-hostile. It is not irredeemably so. Section 7 sets out what would change it. 5. The AI Divide in Municipal Services 5.1 Why the local level is decisive National governments write digital strategies. Municipalities deliver the services that people actually touch: waste collection, housing, local benefits, permits, parking, schools admissions, social care, libraries, and the complaint channels that connect all of them. Cities are also under acute fiscal pressure, which makes automation attractive as a cost-reduction measure rather than as a service-improvement measure. That motivation shapes the design. A system built to reduce staff costs will optimise for deflecting contact. A system built to improve service would optimise for resolving need. These produce very different architectures, and citizens can tell the difference. Responsible use of AI by local government therefore requires an explicit framework rather than ad hoc adoption, covering purpose, capability, accountability, and community engagement (Yigitcanlar, Corchado, Mehmood, Li, Mossberger, and Desouza, 2021). Very few municipalities have one. 5.2 Eligibility automation and the manual lane The clearest documented harm is the creation of a two-tier service. When eligibility is determined automatically from registry data, some citizens are covered and some are not. The uncovered are not randomly distributed. In the Norwegian case study of automated child benefit awards, citizens who fell outside the automated path were disproportionately low-income, and they were required to complete a manual application to receive what they were already entitled to (Larsson, 2021). This is #welfare_automation producing an inverse care effect. The people whose circumstances are typical, stable, and well documented, which is to say the more advantaged, glide through. The people whose circumstances are atypical, which includes those with irregular work, multiple addresses, complex family structures, recent migration, or incomplete documentation, are pushed into the slow lane. Their need is greater and their service is worse. The mechanism generalises. Any automated eligibility system defines a normal case and handles it. Everyone else is an exception, and exceptions cost time. #automated_decision_making thus quietly redefines the atypical citizen as an administrative problem. 5.3 Administrative burden and the accidental caseworker Digital self-service is often justified as convenience. Research on citizens in need of financial assistance shows a different reality: self-service shifts the work of case handling onto the citizen, who must interpret the rules, gather the evidence, complete the forms correctly, and diagnose errors, without the training that a caseworker would have received (Madsen, Lindgren, and Melin, 2022). The citizen becomes an accidental caseworker. The costs of this transfer are unequally distributed, and the burden literature names them precisely: learning costs (working out what the rules are and how the system works), compliance costs (time, documents, travel, fees), and psychological costs (stress, stigma, and the sense of being treated as a suspect). For a confident professional with a laptop and spare time, these costs are trivial. For a person working two shifts, managing a health condition, or navigating the process in a second language, they are prohibitive. Non-take-up of benefits is the predictable result. #administrative_burden is not a side effect of digital service design. In many systems it is the load-bearing structure. Layering AI on top can go either way. A well-designed assistant that explains eligibility in plain language, in the citizen's own language, and helps assemble a claim, would reduce learning costs substantially. A poorly designed #chatbots deployment that exists to reduce call volume will increase them, because the citizen now has to defeat the chatbot before reaching the human who can actually help. 5.4 Chatbots, front doors, and the deflection trap Municipal chatbots are the most common first AI deployment. They are cheap, visible, and easy to justify. They are also the point at which the divide is most often manufactured. Three failure patterns recur. First, language coverage: the assistant works well in the dominant national language and poorly in the languages actually spoken in the neighbourhoods with the greatest need. Second, register and literacy: the assistant assumes the citizen can name their problem in bureaucratic vocabulary, which is precisely the skill that people unfamiliar with the system lack. A citizen who types "they are throwing me out" needs help; a system waiting for the phrase "possession proceedings" will not give it. Third, the deflection trap: the path to a human being is hidden, delayed, or removed, on the assumption that successful containment equals successful service. The evidence on trust is clear enough to guide design. Citizens' acceptance of AI in public decisions depends substantially on the assurance that a human remains in the loop (Aoki, 2021). A front door that removes the human does not merely inconvenience. It signals to the citizen that the institution has withdrawn. 5.5 Smart city sensing and data deserts #smart_city programmes rest on sensing: traffic counters, air quality monitors, footfall sensors, reporting apps, camera networks. Sensing determines what the city can see, and what it can see determines where it invests. Sensor coverage is rarely equal. Instrumentation follows investment, and investment follows existing affluence and existing political attention. Reporting apps compound this, because the people who report potholes, broken lights, and illegal dumping through the app are disproportionately the people who have smartphones, data, time, civic confidence, and the expectation of being listened to. A city that allocates repair crews by app-reported demand will therefore direct resources toward the neighbourhoods that were already better served, and it will do so while believing itself to be data-driven and objective. This is the phenomenon of #data_deserts. Under-measured neighbourhoods appear, in the dashboard, to have fewer problems. Their invisibility is then read as adequacy. The remedy is not more technology but a deliberate policy of representativeness: proactive sensing in under-instrumented areas, weighting of demand signals by known under-reporting, assisted reporting channels, and periodic ground-truthing surveys conducted by people, not devices. #urban_data equity has to be engineered. 5.6 Predictive systems and disproportionate scrutiny Municipalities increasingly use #predictive_analytics for fraud risk, child protection triage, housing inspection targeting, and enforcement prioritisation. These systems share a structural feature: they are trained on records of past institutional attention. Past attention was not distributed evenly. Where a group was historically scrutinised more, more of its members appear in the record, and the model learns to send scrutiny back to the same place. The loop closes. The literature on bias and fairness in machine learning is explicit that this feedback dynamic is one of the hardest to detect from inside the system, because the model's predictions look accurate against the data the system itself generated (Mehrabi and colleagues, 2021). Accuracy against a biased record is not fairness. It is fidelity to the bias. Human oversight is often offered as the safeguard, but the empirical picture is sobering. Officials may exhibit automation bias, accepting algorithmic advice without independent judgement, or selective adherence, accepting the advice when it matches their prior expectations and ignoring it when it does not (Alon-Barkat and Busuioc, 2023). Either failure mode can leave a system that looks supervised while functioning as if it were not. Meaningful oversight requires training, time, authority to override, and a record of overrides that is itself monitored. 5.7 Transparency, contestation, and redress The right that matters most in an automated administration is the right to be told why and the right to have the decision looked at again by a person. Both are commonly weak. #transparency in public sector algorithms is complicated by legitimate confidentiality concerns and by the practical difficulty of explaining a complex model to a lay audience. But the pitfalls of explanation are well documented, and a badly designed explanation can be worse than none, by producing the appearance of understanding without the substance of it (de Bruijn, Warnier, and Janssen, 2022). Experimental evidence indicates that transparency about algorithmic decision-making does affect perceived trustworthiness, though the effect is conditional on how the information is presented and on what the citizen already believes about the institution (Grimmelikhuijsen, 2023). The practical conclusion is that #right_to_human_review must be procedural rather than rhetorical. It requires: notification that an automated system was used; a plain-language statement of the main factors that drove the outcome; a named, reachable route to a human reviewer with the authority to change the decision; a deadline; and no penalty for asking. Systems lacking any one of these elements should not be deployed in decisions that affect entitlements. 5.8 Summary of the municipal layer Applying the diagnostic: Layers affected: all four, with the algorithmic layer dominant. Even fully skilled, fully connected citizens are exposed if they are atypical in the data. User or subject: citizens are overwhelmingly subjects of municipal AI, not users of it. This is the defining asymmetry of the public sector case, and it is why rights and oversight matter more here than in education. Burden transfer: almost total. When municipal automation fails, the citizen absorbs the delay, the paperwork, the travel, and the stress, while the agency records a cost saving. The municipal configuration is therefore more equity-hostile than the educational one, and it is subject to less public scrutiny, because the people it fails are the least likely to be heard. 6. Why This Is a Social Crisis: The Compounding Mechanisms Sections 4 and 5 described harms in two domains. This section explains why those harms do not stay in their domains, and why the total effect is greater than the sum of the parts. 6.1 Mechanism one: the feedback loop Every system described above learns from its own outputs. A dropout model that flags a student, triggering reduced expectations, produces a worse outcome, which becomes training data confirming the flag. A fraud model that targets a neighbourhood finds more fraud there, because that is where it looked, and the finding justifies more targeting. A city that repairs the streets where the app users live records higher satisfaction among app users, which justifies further investment in the app. The loop is closed and self-reinforcing. It does not require anyone to hold a prejudiced belief. It only requires that the system be evaluated against data that the system itself shaped. This is the central reason why #algorithmic_audit must be external and must use independently collected outcome data. Internal validation will always confirm the model, because the model wrote the exam. 6.2 Mechanism two: cross-domain accumulation A single disadvantage in one domain would be recoverable. The problem is that the same household is exposed in all domains at once. Consider a plausible composite. A household has one shared smartphone and a metered data plan. The child cannot do sustained written work, so their assignments are weaker. The school's writing tool flags their prose as machine-generated because the child, a second-language writer, uses simple, regular sentence structures. The child is placed under suspicion, disengages, and is flagged by the early-warning system, which recommends a less demanding pathway. Meanwhile the parent applies for a housing benefit. Their employment is informal and their records are inconsistent, so the automated eligibility check does not cover them, and they are directed to a manual application requiring documents they must take unpaid leave to obtain. The council's chatbot cannot understand their question and does not offer their language. They give up. The benefit goes unclaimed, household income falls, and the data plan is cut further. Every step in that sequence is documented in the literature reviewed above. None of it requires bad faith by any individual. It is the emergent behaviour of a set of independently reasonable systems interacting with a household that is atypical in all of them. #equity failures compound multiplicatively, not additively. 6.3 Mechanism three: the shrinking of the alternative Historically, exclusion from one channel was survivable because other channels existed. You could go to the office, phone, write a letter, or ask a neighbour. Digital-by-default policy has closed those channels, often explicitly, as a cost measure. When the online route is the only route, exclusion from it is exclusion from the service, full stop. This is the point at which a technology question becomes a rights question. If a service is legally an entitlement, and the only practical way to obtain it requires capabilities that a citizen does not possess and cannot easily acquire, then the entitlement has been quietly revoked by an implementation decision. Digital social policy scholarship has made exactly this point: the choice of delivery mechanism is a policy choice with distributive consequences, and it should be evaluated as such rather than treated as neutral administration (Henman, 2022). The digital divide as a human rights concern is now an established framing in social work and social policy literature (Sanders and Scanlon, 2021). 6.4 Mechanism four: the erosion of trust and the withdrawal of participation Repeated failure teaches a lesson. A citizen who has been misjudged by an automated system, or who has spent hours defeating a chatbot to reach a human, learns that the institution is not for them. The rational response is withdrawal: stop applying, stop reporting, stop engaging. Withdrawal then removes the person from the data, which reduces the measured demand from their group, which reduces future investment. Trust research shows that acceptance of AI in the public sector depends on human involvement and on the perceived fairness of the process (Aoki, 2021; Grimmelikhuijsen, 2023). Once trust is lost, it is expensive to rebuild, and the cost of rebuilding falls on the same public budgets that automation was meant to protect. #citizen_participation is not a soft outcome. It is the input to the data that governs everything else. 6.5 Why the crisis framing has practical consequences Naming the situation a crisis is not rhetorical escalation. It changes three things about how institutions should behave. It changes the burden of proof. In ordinary technology adoption, the sceptic must prove harm. In a crisis, the deployer must prove that the system does not worsen an existing emergency. This inverts the default and justifies pre-deployment #impact_assessment as a requirement rather than a courtesy. It changes the time horizon. Crisis responses do not wait for a complete evidence base. They act on the best available understanding, with monitoring, and correct as they learn. Given how far adoption is running ahead of evidence (Zuiderwijk, Chen, and Salem, 2021; Madan and Ashok, 2023), waiting for certainty means accepting years of avoidable harm. It changes the unit of accountability. In a crisis, the outcome is what matters, not the effort. A city cannot report success because it published an AI strategy. It must report whether the gap narrowed. 7. A Mitigation Framework The framework below is organised into five pillars, with two cross-cutting enablers. It is designed to be implementable by a national ministry, a city administration, or a school district, at different levels of ambition, without requiring capabilities that ordinary public bodies do not have. Pillar One: Infrastructure and Affordability Objective: ensure that no person is excluded from an AI-mediated service by lack of a device, a connection, power, or money. Actions. Universal service obligations extended to meaningful connectivity. Coverage targets are insufficient. The target should be meaningful connectivity, defined as access at adequate speed, with adequate data, on an adequate device, at an affordable cost, in a safe and usable place, whenever needed. This is the standard the ITU now promotes and it should be written into national and municipal policy. Affordability as a regulated outcome. Where mobile broadband remains unaffordable for a large share of households, market competition alone has failed and direct intervention is warranted: social tariffs, zero-rating of essential public and educational services, subsidised data allowances tied to school enrolment or benefit receipt, and public review of pricing. #affordability is the barrier most often ignored by digital strategies and the one most often decisive at the household level. Device adequacy, not device counts. Programmes should specify a minimum usable device for the intended task, with a repair and replacement pathway, rather than reporting the number of tablets distributed. A device that cannot run the tools the school assigns is not access. #community_networks and public access. Community-owned and municipally supported networks, alongside #public_libraries, community centres, and school facilities open outside teaching hours, provide a floor that commercial provision does not reach. These are not relics of an earlier era of digital inclusion. In an AI era they become more important, because the bandwidth and device requirements have risen. Electricity and physical space. Any AI-in-education programme in a region with unreliable power must include a power plan. Any home-learning expectation must account for households without a quiet place to work. Pillar Two: Capability and AI Literacy Objective: ensure that people can use AI as a tool, not only be used by it. Actions. Define and teach AI literacy. AI literacy is a distinct competence, encompassing understanding what these systems are, what they can and cannot do, how they fail, how to prompt and evaluate outputs, and how to recognise when a decision about you has been automated (Ng, Leung, Chu, and Qiao, 2021). It should appear in the school curriculum, in adult education, and in the induction of public servants. Invest in teachers before tools. No AI procurement in education should be approved without a funded professional development component covering pedagogy, bias, privacy, and assessment redesign, with protected time. The evidence that the field lacks pedagogical and ethical grounding is strong (Bond and colleagues, 2024). Buying software does not close that gap. #digital_navigators. Trained human intermediaries, based in libraries, clinics, schools, community organisations, and council offices, who sit with a person and help them complete a task, are among the most cost-effective interventions available. They address the second and third levels of the divide simultaneously and they generate intelligence about where systems fail. Capability for public servants. Officials who use algorithmic advice need training in when to override it, and organisations need to reward overrides that turn out to be correct. Without this, human-in-the-loop safeguards degrade into rubber stamps (Alon-Barkat and Busuioc, 2023). Pillar Three: Inclusive Design and Representative Data Objective: ensure that systems work for the people at the margins, not only the median user. Actions. Design for the hardest case first. If a service works for a person on an old phone, on a slow connection, in their second language, with low administrative confidence, then it works for everyone. If it is designed for the median user and then patched, it will not. #universal_design is cheaper when it is upstream. Language coverage as a procurement requirement. Any citizen-facing or student-facing AI system should be evaluated for performance in every language spoken by more than a defined threshold of the served population, with published results. Where performance is inadequate, the system should not be the only channel. Representative evaluation, not just representative training. It is not always possible to fix the training data, but it is always possible to test performance separately for each relevant group and publish the disaggregated results. Aggregate accuracy hides exactly the failures that matter. #co_design with affected communities. Involve the people who will be subject to the system in specifying it, and pay them for their time. This is not consultation theatre. It is the only reliable way to discover the failure modes that designers cannot imagine. Deliberate correction of #data_deserts. Where sensing or reporting is thin, actively instrument and actively survey. Weight demand signals to correct for known under-reporting. Never treat absence of complaint as absence of problem. Pillar Four: Governance, Audit, and Accountability Objective: ensure that harms are detected, attributed, and corrected, and that someone is answerable. Actions. Mandatory pre-deployment #impact_assessment. Before any AI system that affects entitlements, grades, or enforcement is deployed, the deploying body should publish an assessment covering purpose, data sources, known limitations, disaggregated performance, affected groups, fallback channels, and redress routes. The assessment should be short enough to be read and specific enough to be checked. Independent #algorithmic_audit. Audits must be conducted by parties without a commercial interest in the result, using outcome data collected independently of the system. Internal validation is not audit. #procurement as the main lever. Most public bodies do not build AI. They buy it. The contract is therefore the primary point of control. Contracts should require: disclosure of training data provenance at a meaningful level of detail; disaggregated performance reporting; the right to audit; the vendor's cooperation with appeals; data portability; and exit rights. A public body that cannot audit a system it has bought has outsourced its accountability along with its computation. A public register of automated decisions. Citizens should be able to find out, without asking, which decisions in their city are made or supported by automated systems. Clear liability. When an automated system produces a wrongful outcome, the responsible legal person must be identifiable in advance. Diffusion of responsibility between the vendor, the department, and the model is itself a design flaw. #AI_governance in the public sector must be capable of naming a defendant. Pillar Five: Guaranteed Human Channels and Redress Objective: ensure that no entitlement depends on the citizen's ability to operate a machine. Actions. The right to a person. For any service that constitutes a legal entitlement, a non-digital route must remain open, be publicised as prominently as the digital route, and not be slower or more onerous. If the offline route is deliberately degraded to push people online, the right is nominal. The right to human review of automated decisions, implemented procedurally as specified in Section 5.7: notification, plain-language reasons, a named route to a human with authority to change the outcome, a deadline, and no penalty for asking. #multichannel_service by design. Phone, in-person, paper, and assisted digital channels should be treated as permanent components of the service, funded accordingly, rather than as legacy costs to be eliminated. Proactive outreach for non-take-up. Where a system detects that an eligible person has not claimed, it should reach out. The same data infrastructure that enables automated denial enables automated invitation. Which of these a government builds first is a statement of values. Cross-cutting enabler A: Financing None of the above is free. But the relevant comparison is not with zero. It is with the cost of the harm. Non-take-up of benefits, school dropout, unnecessary enforcement, poor health outcomes, and lost tax revenue all carry public costs that exceed the price of doing this properly. Financing mechanisms include universal service funds, a levy on the operators and platforms that benefit from public digital infrastructure, ring-fencing a proportion of every AI procurement budget for inclusion measures, and treating #digital_public_infrastructure as capital investment rather than operating expense. A simple procurement rule is proposed: a fixed percentage of every public AI contract, not less than ten per cent of contract value, must be spent on inclusion measures, covering assisted channels, language coverage, staff training, audit, and redress. This makes inclusion a line item that cannot be cut quietly. Cross-cutting enabler B: Data protection and privacy Inclusion measures require data about disadvantage, and data about disadvantage is dangerous. #privacy and #data_protection are therefore not in tension with equity, they are a precondition for it. Communities that have been surveilled will not supply the data that inclusion planning requires. Strong protections, purpose limitation, independent oversight, and genuine community control over sensitive data are what make representative measurement politically possible. 8. Who Does What A framework without an owner is a wish. This section allocates responsibility. 8.1 National governments Legislate a right to human review of automated decisions affecting entitlements, and a right to a non-digital channel. Set and enforce meaningful connectivity and affordability targets, with published progress. Fund language and corpus development for national and minority languages as a public good. No commercial actor will do this at adequate scale. Establish an independent audit body with statutory access to public sector systems. Standardise impact assessment templates and a public register so that municipalities do not each invent their own. Fund the workforce: teachers, digital navigators, and the public servants who will supervise these systems. 8.2 Municipal governments Adopt a responsible local AI framework covering purpose, capability, accountability, and community engagement (Yigitcanlar and colleagues, 2021), and publish it. Apply the three diagnostic questions from Section 3.3 before any deployment: which layers, user or subject, who absorbs the failure. Map service demand against known under-reporting rather than against app usage. Correct for #data_deserts deliberately. Fund assisted channels permanently and count contact with a human as a success, not a failure. Publish disaggregated performance for every deployed system: who is being served well, who is not. Build the proactive-outreach capability before the automated-denial capability. 8.3 Schools, colleges, and universities Provide equal institutional access to a capable AI tool. Do not let the market allocate cognitive assistance by household income. Teach AI literacy explicitly as a core competence, including failure modes and ethics. Redesign assessment rather than policing it. Measure what students can do with the tools and, separately and deliberately, what they can do without. Refuse to act on automated integrity flags without human review, and never treat a detection score as evidence. Require human review of automated scores and risk flags for any group for which disaggregated performance is weaker. Invest in teachers before tools, with protected time. 8.4 Technology vendors Publish disaggregated performance by language, region, and relevant demographic group. Aggregate benchmarks are marketing, not evidence. Support audit contractually and practically. Build low-bandwidth and low-specification modes as first-class products, not as charity. Price for the public sector in a way that does not make inclusion a premium tier. Accept liability for the harms the system causes. If the vendor will not, the public body should read that as a statement about the product. 8.5 Civil society, researchers, and communities Conduct and publish independent audits, including adversarial ones. Document lived experience of automated exclusion, which is the evidence that administrative data structurally cannot capture, since the excluded are by definition the people the system did not record. Represent affected people in appeals, and use aggregate appeal data to identify systemic failure. Hold the register, the assessments, and the disaggregated numbers to account. Build and sustain #community_networks and local support infrastructure where the state and the market do not reach. 9. Measurement: How to Know If It Is Working Policy that cannot be measured will be claimed rather than achieved. The following indicator set is proposed, organised by layer. It is deliberately short, because indicator sets that are too long are not collected. 9.1 Access layer Share of households with a device adequate for the tasks their public services and schools require, disaggregated by income, region, and household type. Share of households for whom an adequate data plan costs more than two per cent of monthly income. Hours of usable connectivity per day, accounting for power interruptions, in schools and public facilities. 9.2 Usage layer Share of adults able to complete a defined set of essential civic and educational tasks unaided, measured by task, not by self-report. Share of teachers reporting adequate training and protected time for AI-supported teaching. Number of assisted-support contacts delivered per thousand residents, and their outcomes. 9.3 Outcome layer #non_take_up rate for each major benefit, measured against an independent estimate of eligibility, disaggregated. Attainment and progression gaps in AI-supported courses, disaggregated, compared against the same gaps in non-AI-supported courses. Time-to-resolution for a service request, disaggregated by channel and by neighbourhood. 9.4 Algorithmic layer Disaggregated accuracy, false positive rate, and false negative rate for every deployed decision-support system, published. Appeal rate, appeal success rate, and time to appeal resolution, disaggregated. Override rate by human reviewers, and the outcome of overrides, which tests whether the human in the loop is real. Language performance parity: measured performance in each significant local language relative to the dominant language. Indicators 10 through 13 are the innovation here. They are also the ones institutions are most reluctant to collect, precisely because they would reveal the problem. A jurisdiction that publishes them is serious. One that does not is not, whatever its strategy document says. #monitoring_and_evaluation is the difference between #evidence_based_policy and public relations. 10. Discussion, Limitations, and Research Agenda 10.1 What this article claims and does not claim The claim is not that AI is bad for education or for cities. The evidence that these systems can help is real: personalised explanation at scale, faster service, earlier identification of need, better allocation of scarce staff time. The claim is narrower and, we think, harder to dispute: that the default configuration of these systems, adopted under cost pressure without inclusion investment, systematically transfers benefit upward and burden downward, and that this default will hold unless it is actively overridden. Nor is the claim that technology causes inequality. Inequality preceded the technology. What the technology does is encode it, accelerate it, and give it the appearance of objectivity. That last property is the most dangerous, because it disarms the political response. A caseworker who denies a claim can be argued with. A model that denies a claim is presented as arithmetic. 10.2 Limitations This is a conceptual synthesis, not an empirical study. It inherits the biases of the literature it reviews, which is dominated by research from high-income countries and by studies of systems in English-speaking and European contexts. The municipal evidence base in particular is thin outside a handful of northern European cases, and the generalisability of findings from well-resourced welfare states to municipalities with weaker administrative capacity is untested and should not be assumed. The pace of change is a second limitation. Generative systems are being adopted faster than they can be studied, and any assessment of their equity effects made today rests partly on inference from prior technologies rather than on direct evidence. Third, the proposed framework has not been evaluated in the field. Its components are individually supported, but the framework as an integrated package is a proposal, not a validated intervention. The ten-per-cent procurement rule in particular is a reasoned starting point rather than an empirically calibrated figure, and it should be tested. Fourth, the article treats "the citizen" and "the student" somewhat generically, when in reality the experience of automated exclusion is highly specific to context, status, and identity. Qualitative work with affected populations would enrich, and might correct, parts of this account. 10.3 A research agenda Six priorities are proposed. One. Measure the fourth layer. The field lacks accepted instruments for the algorithmic layer of the divide. Developing and validating measures of algorithmic exposure, algorithmic disadvantage, and effective contestability is the most urgent methodological task. Two. Study the manual lane. Larsson's finding that automation coverage is itself unequally distributed deserves replication across many systems and countries. How large is the manual lane, who is in it, and what does it cost them? Three. Test the interventions. Digital navigators, universal institutional AI access in schools, proactive outreach for non-take-up, and mandatory disaggregated reporting are all plausible. Almost none have been evaluated with strong designs. Randomised and quasi-experimental evaluation is feasible here and is not being done. Four. Follow the compounding. The composite scenario in Section 6.2 is constructed from separate findings. Longitudinal work following households across education and municipal systems simultaneously would test whether the compounding operates as theorised. Five. Extend the evidence base geographically. The literature is skewed. Research led from low-income and middle-income contexts, and from cities in the global south, is essential and currently scarce. Six. Study the vendors. Very little independent research examines how educational and municipal AI products are actually designed, what assumptions are built into them, and how procurement processes shape those assumptions. The point of leverage is upstream of deployment, and it is almost unstudied. 11. Conclusion The digital divide used to be a queue. Some people were at the front and some at the back, but everyone was in the same line, and the policy task was to move the line forward. Artificial intelligence has ended that arrangement. There are now two lines. In one, people use these systems to think faster, learn more, and get what they are entitled to. In the other, people are read, scored, sorted, and sometimes refused by systems they did not choose, cannot see into, and cannot argue with. Movement between the lines is not automatic, and the gap between them is widening every quarter. That is why the word crisis is appropriate. Not because the technology is malevolent, and not because the outcome is fixed, but because the mechanisms described in this article are self-reinforcing, cross-cutting, and accelerating, and because the window in which the architecture of these systems is still open is closing. Once eligibility engines, tutoring platforms, and municipal decision systems are embedded, they become the infrastructure, and infrastructure is very hard to change. The remedies are not mysterious. Connect people meaningfully and affordably. Teach them to use these tools rather than only to be used by them. Design for the person at the margin rather than the person in the middle. Audit what you deploy and publish what you find. And keep a person on the other end of the line, always, for anyone who needs one. None of this is technically difficult. All of it costs money, and all of it requires an institution to accept that the point of a public service is not to reduce its own contact volume but to meet a need. The choice being made right now, in thousands of school districts and city halls, is not a choice about software. It is a choice about whether the next generation of public institutions will treat the people who need them most as their central case or as their exception. Everything in the evidence reviewed here suggests that if that choice is left to default, it will be made badly. It should be made deliberately, and it should be made now. Hashtags #digital_divide #social_crisis #AI_divide #digital_inclusion #digital_exclusion #algorithmic_bias #educational_technology #digital_government #municipal_services #AI_literacy #administrative_burden #algorithmic_accountability #meaningful_connectivity #public_value #inclusive_AI References Alon-Barkat, S., and Busuioc, M. (2023). Human-AI interactions in public sector decision making: Automation bias and selective adherence to algorithmic advice. Journal of Public Administration Research and Theory, 33(1), 153-169. https://doi.org/10.1093/jopart/muac007 Aoki, N. (2021). The importance of the assurance that humans are still in the decision loop for public trust in artificial intelligence: Evidence from an online experiment. Computers in Human Behavior, 114, 106572. https://doi.org/10.1016/j.chb.2020.106572 Baker, R. S., and Hawn, A. (2022). Algorithmic bias in education. International Journal of Artificial Intelligence in Education, 32(4), 1052-1092. https://doi.org/10.1007/s40593-021-00285-9 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, 610-623. https://doi.org/10.1145/3442188.3445922 Birhane, A. (2021). Algorithmic injustice: A relational ethics approach. Patterns, 2(2), 100205. https://doi.org/10.1016/j.patter.2021.100205 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, 5486-5505. https://doi.org/10.18653/v1/2022.acl-long.376 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, 4. https://doi.org/10.1186/s41239-023-00436-z Bulathwela, S., Perez-Ortiz, M., Holloway, C., Cukurova, M., and Shawe-Taylor, J. (2024). Artificial intelligence alone will not democratise education: On educational inequality, techno-solutionism and inclusive tools. Sustainability, 16(2), 781. https://doi.org/10.3390/su16020781 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 Cowls, J., Tsamados, A., Taddeo, M., and Floridi, L. (2021). A definition, benchmark and database of AI for social good initiatives. Nature Machine Intelligence, 3, 111-115. https://doi.org/10.1038/s42256-021-00296-0 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, 22. https://doi.org/10.1186/s41239-023-00392-8 de Bruijn, H., Warnier, M., and Janssen, M. (2022). The perils and pitfalls of explainable AI: Strategies for explaining algorithmic decision-making. Government Information Quarterly, 39(2), 101666. https://doi.org/10.1016/j.giq.2021.101666 Floridi, L. (2023). The ethics of artificial intelligence: Principles, challenges, and opportunities. Oxford University Press. Grimmelikhuijsen, S. (2023). Explaining why the computer says no: Algorithmic transparency affects the perceived trustworthiness of automated decision-making. Public Administration Review, 83(2), 241-262. https://doi.org/10.1111/puar.13483 Helsper, E. J. (2021). The digital disconnect: The social causes and consequences of digital inequalities. SAGE Publications. Henman, P. (2022). Digital social policy: Past, present, future. Journal of Social Policy, 51(3), 535-550. https://doi.org/10.1017/S0047279422000162 Holmes, W., Porayska-Pokorny, K., Holstein, K., Sutherland, E., Baker, T., Shum, S. B., Santos, O. C., Rodrigo, M. T., Cukurova, M., Bittencourt, I. I., and Koedinger, K. R. (2022). Ethics of AI in education: Towards a community-wide framework. International Journal of Artificial Intelligence in Education, 32, 504-526. https://doi.org/10.1007/s40593-021-00239-1 International Telecommunication Union (2025). Measuring digital development: Facts and figures 2025. International Telecommunication Union. Kasneci, E., Sessler, K., Kuchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Gunnemann, S., Hullermeier, E., Krusche, S., Kutyniok, G., Michaeli, T., Nerdel, C., Pfeffer, J., Poquet, O., Sailer, M., Schmidt, A., Seidel, T., Stadler, M., Weller, J., Kuhn, J., and Kasneci, G. (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103, 102274. https://doi.org/10.1016/j.lindif.2023.102274 Larsson, K. K. (2021). Digitization or equality: When government automation covers some, but not all citizens. Government Information Quarterly, 38(1), 101547. https://doi.org/10.1016/j.giq.2020.101547 Lybeck, R., Koiranen, I., and Koivula, A. (2024). From digital divide to digital capital: The role of education and digital skills in social media participation. Universal Access in the Information Society, 23(4), 1657-1669. https://doi.org/10.1007/s10209-022-00961-0 Lythreatis, S., Singh, S. K., and El-Kassar, A. N. (2022). The digital divide: A review and future research agenda. Technological Forecasting and Social Change, 175, 121359. https://doi.org/10.1016/j.techfore.2021.121359 Madan, R., and Ashok, M. (2023). AI adoption and diffusion in public administration: A systematic literature review and future research agenda. Government Information Quarterly, 40(1), 101774. https://doi.org/10.1016/j.giq.2022.101774 Madsen, C. O., Lindgren, I., and Melin, U. (2022). The accidental caseworker: How digital self-service influences citizens' administrative burden. Government Information Quarterly, 39(1), 101653. https://doi.org/10.1016/j.giq.2021.101653 Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., and Galstyan, A. (2021). A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6), 1-35. https://doi.org/10.1145/3457607 Mergel, I., Dickinson, H., Stenvall, J., and Gasco, M. (2023). Implementing AI in the public sector. Public Management Review. https://doi.org/10.1080/14719037.2023.2231950 Mubarak, F., and Suomi, R. (2022). Elderly forgotten? Digital exclusion in the information age and the rising grey digital divide. INQUIRY: The Journal of Health Care Organization, Provision, and Financing, 59. https://doi.org/10.1177/00469580221096272 Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., and Qiao, M. S. (2021). Conceptualizing AI literacy: An exploratory review. Computers and Education: Artificial Intelligence, 2, 100041. https://doi.org/10.1016/j.caeai.2021.100041 Nguyen, A., Ngo, H. N., Hong, Y., Dang, B., and Nguyen, B. P. T. (2023). Ethical principles for artificial intelligence in education. Education and Information Technologies, 28(4), 4221-4241. https://doi.org/10.1007/s10639-022-11316-w Organisation for Economic Co-operation and Development (2023). OECD digital education outlook 2023: Towards an effective digital education ecosystem. OECD Publishing. Sanders, C. K., and Scanlon, E. (2021). The digital divide is a human rights issue: Advancing social inclusion through social work advocacy. Journal of Human Rights and Social Work, 6(2), 130-143. https://doi.org/10.1007/s41134-020-00147-9 Schiff, D. (2021). Out of the laboratory and into the classroom: The future of artificial intelligence in education. AI and Society, 36(1), 331-348. https://doi.org/10.1007/s00146-020-01033-8 Schiff, D. (2022). Education for AI, not AI for education: The role of education and ethics in national AI policy strategies. International Journal of Artificial Intelligence in Education, 32, 527-563. https://doi.org/10.1007/s40593-021-00270-2 Selwyn, N. (2022). The future of AI and education: Some cautionary notes. European Journal of Education, 57(4), 620-631. https://doi.org/10.1111/ejed.12532 Selwyn, N. (2022). Education and technology: Key issues and debates (3rd ed.). Bloomsbury Academic. Tsamados, A., Aggarwal, N., Cowls, J., Morley, J., Roberts, H., Taddeo, M., and Floridi, L. (2022). The ethics of algorithms: Key problems and solutions. AI and Society, 37, 215-230. https://doi.org/10.1007/s00146-021-01154-8 UNESCO (2021). AI and education: Guidance for policy-makers. United Nations Educational, Scientific and Cultural Organization. UNESCO (2023). Global education monitoring report 2023: Technology in education. United Nations Educational, Scientific and Cultural Organization. UNESCO (2023). Guidance for generative AI in education and research. United Nations Educational, Scientific and Cultural Organization. Vassilakopoulou, P., and Hustad, E. (2023). Bridging digital divides: A literature review and research agenda for information systems research. Information Systems Frontiers, 25(3), 955-969. https://doi.org/10.1007/s10796-020-10096-3 World Bank (2021). World development report 2021: Data for better lives. World Bank. Yigitcanlar, T., Corchado, J. M., Mehmood, R., Li, R. Y. M., Mossberger, K., and Desouza, K. (2021). Responsible urban innovation with local government artificial intelligence: A conceptual framework and research agenda. Journal of Open Innovation: Technology, Market, and Complexity, 7(1), 71. https://doi.org/10.3390/joitmc7010071 Zhang, K., and Aslan, A. B. (2021). AI technologies for education: Recent research and future directions. Computers and Education: Artificial Intelligence, 2, 100025. https://doi.org/10.1016/j.caeai.2021.100025 Zuiderwijk, A., Chen, Y. C., and Salem, F. (2021). Implications of the use of artificial intelligence in public governance: A systematic literature review and a research agenda. Government Information Quarterly, 38(3), 101577. https://doi.org/10.1016/j.giq.2021.101577

This article offers a full, student friendly, and academically organised study of the framework known as the thirty-six dramatic situations, first published in French in 1895 by the theatre critic #Georges_Polti and made widely available in English translation from 1916 onward. The framework claims that the conflicts of drama, although endless in surface detail, fall into a limited number of repeatable structures, each built from a small set of roles that stand in tension with one another. The article traces the intellectual history of the idea from #Carlo_Gozzi through Goethe and Schiller to Polti, and then to later attempts at classification such as the two hundred thousand situations proposed by Etienne Souriau. It presents all thirty-six situations one by one, listing the dynamic elements that Polti considered technically necessary for each, explaining the internal logic in plain language, and giving examples drawn from widely studied works. It then groups the situations into larger families in order to show that the list is not a flat inventory but a structured map of #conflict. The article evaluates the framework against modern #narratology, and reviews how it is currently used in #screenwriting education, in game design, in #computational_narratology, and in recent research on story generation with large language models. Finally, it sets out the main criticisms of the framework, including overlap between categories, cultural narrowness, the tragic bias of the sample, and the absence of any test of reliability, and it proposes practical guidance for students who want to use the list as an analytical instrument rather than as a recipe. The conclusion argues that the thirty-six situations remain valuable not because the number thirty-six is correct, but because the framework forces the analyst to name the roles, the stakes, and the pressure that make a scene dramatic. Keywords: Georges Polti, dramatic situations, plot taxonomy, narrative theory, conflict, screenwriting, computational narratology, creative writing pedagogy 1. Introduction Every student of literature, theatre or film eventually meets the same uncomfortable question. If there are only a few basic stories in the world, why do we keep reading and watching new ones? The question is old, and it is not merely a puzzle for critics. It matters for the working writer, for the teacher who must explain why one scene works and another does not, and for the researcher who now asks a machine to produce a story that a human being will accept as a story. One of the most durable answers to that question is the list of thirty-six #dramatic_situation types published by the French theatre critic Georges Polti at the end of the nineteenth century. The claim of the list is simple and bold. Polti argued that beneath the endless variety of characters, settings, periods and styles, drama is powered by a limited number of structural arrangements, each one defined by a small cast of functional roles placed under pressure. Change the names, the century and the costumes, and the same engine still runs. A person begs a hostile power for mercy, which is #supplication. A person avenges a killing. A person is torn between love and duty to family. A person discovers that the crime they punished was their own. This article treats the framework as an object of study rather than as a manual of tricks. That distinction matters. Popular writing about the thirty-six situations often reduces the framework to a checklist that promises plot ideas on demand. Serious study asks harder questions. What exactly is a situation, as opposed to a theme, a plot, a motif or a genre? On what evidence was the list built? Is the list exhaustive, as its author suggested, or merely convenient? Do the categories overlap, and if they overlap, what does that tell us about the nature of #conflict itself? And what is the framework good for today, in an era when narrative is analysed statistically and generated automatically? The article has four aims. The first is descriptive: to present the thirty-six situations accurately and completely, including the functional roles that Polti called the dynamic elements of each situation. The second is historical: to show where the idea came from and how it was transmitted. The third is critical: to test the framework against the standards of modern #narrative_theory and to state its weaknesses without either dismissing it or worshipping it. The fourth is applied: to show how students, writers, teachers and researchers can use the framework responsibly in #teaching_writing, in #literary_analysis, in #playwriting, in #screenwriting, in game design and in current work on #artificial_intelligence and storytelling. The argument advanced here is that the value of the framework does not depend on the accuracy of the number thirty-six. The number is arbitrary, and Polti himself did not defend it as a law of nature. The value lies elsewhere. The framework is one of the earliest systematic attempts to describe drama as a set of relations rather than as a set of subjects, and this relational approach is precisely what makes it usable a century later. A situation in this sense is not a topic. It is a configuration of forces. Once a student learns to see configurations rather than topics, the student can analyse any narrative, in any medium, including narratives that Polti could never have imagined. 2. Historical background and intellectual lineage 2.1 The anecdote that started everything The story of the thirty-six situations begins not with Polti but with an anecdote. The Venetian dramatist Carlo Gozzi, who worked in the eighteenth century and is remembered for his theatrical fables, is reported to have claimed that the number of tragic situations available to a dramatist is limited, and that the limit is thirty-six. The claim survives mainly through report rather than through a preserved demonstration. It was repeated in German literary circles, where Goethe recorded that Schiller tried to find more than thirty-six situations and could not manage even that many. This anecdote is important for two reasons. First, it shows that the intuition of a finite dramatic repertoire is much older than Polti, and that it belongs to the practical world of the theatre rather than to abstract philosophy. Second, it shows how weak the evidential foundation was. A remark attributed to a playwright, passed through a poet, becomes a number, and the number acquires an authority that no one ever earned by argument. Polti inherited both the intuition and the number, and much of the later criticism of his work is really criticism of Gozzi's unsupported figure. 2.2 Polti and his method Polti was a theatre critic and writer associated with Parisian literary journals of the late nineteenth century, an environment in which the ambition to make criticism scientific was widely shared. His book appeared in French in 1895, with later French editions, and reached the English speaking world through a translation published in 1916 and reprinted repeatedly through the twentieth century. The book is short, dense, and written in a style that modern readers often find difficult, but its structure is clear. For each of the thirty-six situations, Polti lists the roles required, describes the situation, subdivides it into variants, and cites works that exemplify each variant. The method was, in intention, empirical. Polti read very widely in Greek tragedy, in French classical and contemporary drama, and in a smaller selection of other European literatures, and he sorted what he found. The sorting produced a much larger number of variants, usually given as a few hundred, which he then grouped under thirty-six headings. This is the crucial and much criticised move. The variants are the observation; the thirty-six headings are an interpretation imposed on the observation, and they were imposed partly in order to vindicate Gozzi. In other words, the empirical work is genuine, but the famous number is a decision rather than a finding. It is also worth noting what Polti was arguing against. He was not celebrating the poverty of the theatre. He was complaining about it. His view was that dramatists had explored a shockingly small portion of the space available to them, and his list was meant to expose that neglect and to provoke writers into occupying the empty regions. The framework was therefore intended as a generative and critical instrument, not as a comforting proof that everything has already been written. 2.3 Rival taxonomies Polti is one entry in a long series of attempts to classify narrative structure, and students should see him in that series rather than in isolation. The Russian folklorist Vladimir Propp, working in the 1920s on a corpus of Russian wonder tales, produced a very different kind of model. Instead of situations, Propp identified functions, defined as acts of a character considered in terms of their significance for the course of the action, and he argued that these functions occur in a fixed order. Propp's model is sequential; Polti's is not. A Proppian analysis tells you what happens next. A Poltian analysis tells you what is at stake now. The contrast is instructive, because it shows that a taxonomy encodes a theory of what narrative fundamentally is, whether the author admits it or not. The French aesthetician Etienne Souriau, writing in mid century, deliberately exploded Polti's economy by proposing that the number of dramatic situations, understood combinatorially, runs into the hundreds of thousands. Souriau's move is a reminder that the size of a taxonomy depends entirely on the grain of description chosen. If you describe situations by the relations among a handful of dramatic functions, and you allow those functions to combine freely, the numbers grow very fast. Thirty-six is not a discovery about the world; it is a statement about resolution. Later popular taxonomies continued in the same tradition, offering twenty plots, seven plots, or a single mythic journey with many stages. These lists differ in number but share the assumption that narrative variety is generated from a small combinatorial base. Recent scholarship in #computational_narratology has taken this assumption seriously and tested it on large corpora, and has argued that computational work on narrative should be linked explicitly to narratological theory rather than proceeding without it (Piper, So, and Bamman, 2021). 2.4 Reception and afterlife The framework has had three lives. In its first life it was a tool of theatre criticism. In its second it became a practical resource for writers, especially in the twentieth century boom of instructional writing manuals, where the list circulated widely, often stripped of Polti's commentary and reduced to a bare enumeration. In its third and current life it has been picked up by researchers in interactive storytelling and #artificial_intelligence, who value it precisely because it offers a compact and role based vocabulary for describing dramatic tension in a form that a system can reason about (Lessard and Pare-Chouinard, 2022; Wu, Wu, Xu, Zhang, and Zhao, 2025). This third life is the most interesting for current students, because it changes the criteria of evaluation. A framework used by a critic must be true. A framework used by a writer must be fertile. A framework used by a computational system must be operational, which means that its categories must be applied consistently by different observers and must be capable of being represented in data. Polti's list has never been tested against that last standard, and the gap is one of the most promising research openings available today. 3. Research aims and questions This study is a conceptual and analytical review. It does not report new empirical data. Its purpose is to consolidate, clarify and evaluate an old framework so that students can use it well. The study is organised around five questions. First, what is a dramatic situation, defined precisely enough to distinguish it from neighbouring concepts such as theme, motif, plot and genre? Second, what are the thirty-six situations, and what are the dynamic elements or roles that each requires? Third, how do the situations relate to one another? Are there natural families, and does the structure of those families tell us anything about the underlying logic of #dramaturgy? Fourth, what are the limits of the framework, considered as a scientific classification rather than as a writing aid? Fifth, what are the legitimate contemporary applications of the framework in #literary_analysis, in creative practice, in teaching, and in computational research? 4. Methodology The method used here is a structured conceptual review with an analytical component, which is the standard approach for work of this kind in the humanities. The primary text is Polti's own book in its English translation, which supplies the canonical list, the roles, and the subdivisions. The historical context is reconstructed from the transmission history of the Gozzi anecdote and from the framework's reception in later theory. Secondary literature was gathered by searching for recent work that engages seriously with dramatic situations, plot taxonomy, and narrative structure, with particular attention to the last five years, since this is the period in which the framework has been revived within interactive storytelling and language technology research. The secondary literature is therefore weighted towards two fields: contemporary #narratology and computational narrative research. This weighting is deliberate. It reflects where the live argument about the framework is actually taking place. The analytical component proceeds in three steps. In the first step, each of the thirty-six situations is restated in plain language together with its roles. In the second step, the situations are grouped into families using a simple criterion: two situations belong to the same family when the source of pressure on the protagonist is of the same kind, for example pressure from kinship, pressure from desire, pressure from power, or pressure from knowledge. In the third step, the framework is stress tested against known narratives that fit awkwardly, since the interesting information about any classification lies at its boundaries rather than at its centre. A note on evidence is necessary. Because Polti did not publish his coding decisions, we cannot reproduce his classification of any given play. This is a real methodological weakness of the original work, and it should be stated openly rather than hidden. Any student who applies the framework is therefore performing an interpretive act, and should say so. 5. What a dramatic situation is Before the catalogue, the central concept must be defined, because most misunderstandings of the framework come from a loose sense of the word situation. A dramatic situation, in Polti's sense, is a configuration of roles in tension. It has three defining features. It is relational. A situation is not a person and not an event. It is a relation among positions. Vengeance requires an avenger and a guilty party. #supplication requires someone who begs, someone who threatens, and someone with the power to decide. Remove one position and the situation collapses. It is dynamic. The positions are not stable. They exert force on one another, and that force demands resolution. This is why Polti called the roles the dynamic elements. A situation contains within itself a pressure toward change, which is what distinguishes it from a mere state of affairs. Two people who dislike each other are in a state. Two brothers who both need the same inheritance and cannot both have it are in a situation. It is neutral about outcome. A situation does not specify how it ends. #Ambition may end in triumph or in ruin. This neutrality is what makes the framework a tool of analysis rather than a set of plots. It also explains why a single long narrative usually contains many situations in sequence, and why two works that share a situation can feel entirely different. From these three features we can distinguish the situation from its neighbours. A theme is an idea that a work explores, such as justice or forgiveness, and it is not necessarily relational. A motif is a recurring concrete detail, such as a locked room or a poisoned cup. A plot is the particular ordered chain of events in a particular work. A genre is a set of conventions and audience expectations. A situation sits underneath all of these. It is the mechanical core that the theme interprets, the motif decorates, the plot arranges, and the genre dresses. 6. The thirty-six dramatic situations What follows is the complete catalogue. For each situation the article gives the traditional title, the dynamic elements, an explanation of the internal logic, notes on common variants, and examples. Examples are drawn from works that are widely taught, so that students can verify the analysis for themselves. 6.1 Supplication Elements: a Persecutor; a Suppliant; a Power whose decision is in doubt. This is the founding situation of the list and the clearest illustration of the relational principle. A weak party, under threat, appeals to an authority that could save them but might not. The drama lives entirely in the uncertainty of the third role. If the authority is certain to help, there is no tension; if it is certain to refuse, there is only despair. The situation therefore requires a doubtful power, which is why courtroom scenes, appeals to a king, prayers to a god, and pleas to an indifferent institution are so dramatically productive. Variants turn on who the authority is and what it wants. The authority may be bound by law, by love, by fear, or by self interest. In one common variant the persecutor and the authority are the same person, which compresses the triangle and raises the stakes, since the suppliant must beg the very party that threatens them. Examples include the tragedies of the Greek theatre in which exiles or refugees beg a city for protection, and the many modern courtroom dramas in which a defendant's fate rests on a judgement that could go either way. 6.2 Deliverance Elements: an Unfortunate; a Threatener; a Rescuer. #Deliverance is the mirror image of #supplication. Here the emphasis falls not on the plea but on the intervention. Someone is in danger and someone else acts. The dramatic weight shifts from the victim to the rescuer, and the interesting questions become questions about the rescuer: what does the rescue cost, what does it reveal, and does the rescuer arrive in time? Because the arrival of help is inherently suspenseful, this situation is the engine of much popular narrative, from the rescue in the last reel of an action film to the unexpected ally in a war story. Its weakness is that a rescuer who is too powerful destroys the tension, which is why writers usually handicap the rescuer with distance, doubt, injury, or a competing obligation. 6.3 Crime pursued by vengeance Elements: an Avenger; a Criminal. Here the moral engine is retribution outside the machinery of official justice. The situation is minimal in cast, since only two positions are strictly required, and yet it is one of the most emotionally powerful, because the audience is invited to want something it may also fear. The dramatic interest usually comes from the tension between the justice of the cause and the corruption of the pursuit. The avenger may become what they hunt. The criminal may become sympathetic. The law may be either absent or complicit. Variants are organised by what was done and to whom: the murder of a friend, the destruction of a livelihood, the betrayal of a trust, or a professional wrong. 6.4 Vengeance taken for kindred upon kindred Elements: an Avenging Kinsman; a Guilty Kinsman; the Remembrance of the Victim. This is #vengeance turned inward. The avenger and the guilty party belong to the same family, and so the act of justice is also an act of self destruction. The third element, the remembrance of the victim, is not a person but a pressure, and this is a subtle and important point about Polti's method: he allows a memory or an obligation to occupy a structural position. The situation is the core of the Greek tragic house and of a large part of the Western tragic tradition after it, in which a child must avenge a parent upon a parent, or a sibling upon a sibling. It is also, in modern narrative, the engine of crime sagas centred on families, where loyalty and justice pull in opposite directions and every choice is a betrayal of something. 6.5 Pursuit Elements: a Fugitive; a Punishment or Pursuer. #Pursuit strips drama down to motion. Someone flees; something follows. Polti's interest is in the moral colouring of the flight, because the situation changes entirely depending on whether the fugitive deserves capture. A guilty fugitive produces a study in conscience. An innocent fugitive produces a study in injustice and paranoia. A fugitive who is guilty of a crime the audience approves of produces the most complex effect of all. The situation is medium friendly to an unusual degree, since it converts naturally into physical movement, and it therefore dominates film and interactive media, where it can be experienced rather than merely observed. 6.6 Disaster Elements: a Vanquished Power; a Victorious Enemy or a Messenger. Disaster concerns the moment of collapse: the defeat of an army, the fall of a house, the loss of a fortune, the sudden reversal of a life. Polti's inclusion of the Messenger as a possible element is significant, because it captures a specifically theatrical fact. Catastrophe is often too large to stage, and so it enters the drama as news. The scene of the messenger is therefore not a limitation of ancient theatre but a dramatic form in its own right, and it survives today in the phone call, the telegram, the broadcast, and the notification. The situation is about the experience of reversal rather than about the reversal itself. What the audience watches is a person absorbing a fact that has already destroyed them. 6.7 Falling prey to cruelty or misfortune Elements: an Unfortunate; a Master or a Misfortune. Where #disaster is a single blow, this situation is a condition. The protagonist is subject to sustained cruelty or grinding misfortune with no immediate prospect of relief. Polti groups here the abandoned, the exploited, the imprisoned, and the forgotten. This is a difficult situation to sustain dramatically, because pure suffering without agency becomes static. Successful uses give the sufferer some small field of action, some dignity, or some hope that the audience can hold on to, which is why so many works in this category are structured around endurance rather than around change. 6.8 Revolt Elements: a Tyrant; a Conspirator. #Revolt is the political situation in its purest form. Power is illegitimate, or is felt to be, and someone acts against it. The essential dramatic questions are about the conspirator rather than the tyrant: what is the cost of acting, who is betrayed by the act, and does the rebel become the thing they overthrow? Polti's variants cover the lone rebel, the conspiracy of a few, and the mass rising. The three are structurally different. The lone rebel produces tragedy of conscience. The conspiracy produces tragedy of trust, since conspiracies are destroyed from within. The mass rising produces epic, and tends to convert the individual into a representative. 6.9 Daring enterprise Elements: a Bold Leader; an Object; an Adversary. This is the situation of the #daring_enterprise, or the plan. A goal is set, obstacles are known, and a group or an individual attempts something difficult. It is the structural core of the #heist, the raid, the expedition, the rescue mission, and the campaign. Its #dramatic_economy is unusually clear, because the audience can measure progress against a stated objective. This is also its risk: an enterprise without internal conflict becomes mere procedure. Writers therefore load the enterprise with a second situation, usually rivalry, betrayal, or love, so that the plan and the person are in tension. 6.10 Abduction Elements: an Abductor; the Abducted; a Guardian. #Abduction is a violation of possession or protection. Polti treats the category more broadly than a modern reader expects, since he includes elopements and rescues in which the abducted party consents. The structural point is that a person is removed from the sphere of one authority into the sphere of another, and the guardian must respond. Read structurally, the situation exposes an uncomfortable assumption in much traditional narrative: the abducted party is often treated as an object between two claimants rather than as an agent. Contemporary uses of the situation frequently work precisely by breaking that assumption and giving the abducted person the decisive agency. 6.11 The enigma Elements: an Interrogator; a Seeker; a Problem. Here the pressure is epistemic rather than physical, and the engine is the #enigma. Something is unknown, and knowing it matters. Polti's version includes riddles, hidden identities, buried facts, and searches for a missing person or object. This situation is the ancestor of the modern #detective_story, but it is much wider than crime. Any narrative in which the audience is made to want an answer is running on this engine, including scientific narratives, mysteries of parentage, and the slow revelation of a character's past. Its special property is that it can be laid over almost any other situation as a second layer, which is why it is so common in hybrid forms. 6.12 Obtaining Elements: a Solicitor; an Adversary who refuses; or an Arbitrator and opposing parties. Obtaining concerns the effort to get something from someone who does not want to give it. Polti insists that the situation is defined by the method of getting rather than by the thing got, which is an unusually modern insight, because it locates the drama in tactics: persuasion, cunning, force, seduction, bargaining, and threat. This is the situation of negotiation scenes, of pitches, of proposals, of interrogations that seek a confession, and of any encounter in which two wills meet across a table. Its presence explains why dialogue can be as tense as combat. 6.13 Enmity of kinsmen Elements: a Malevolent Kinsman; a Hated or reciprocally hating Kinsman. Hatred inside the family is dramatically potent because it cannot be escaped. Strangers can walk away; relatives cannot, or at least cannot without cost. Polti's variants cover hatred between brothers, between parent and child, between in laws, and across generations. The situation is worth close study because it demonstrates that the framework encodes not just conflict but the conditions that trap the parties inside the conflict. What makes a situation dramatic is not only opposition but the impossibility of simple exit. 6.14 Rivalry of kinsmen Elements: a Preferred Kinsman; a Rejected Kinsman; the Object of the rivalry. #Rivalry differs from #enmity because it is triangular. Two relatives want the same thing, and the thing may be a person, an inheritance, a throne, or the love of a parent. The rejected kinsman is one of the most productive roles in all of drama, and it generates a recognisable psychological arc: exclusion produces resentment, resentment produces action, action confirms the exclusion. Because the object is often love or approval rather than property, this situation carries deep emotional charge and appears constantly in family drama, in dynastic narrative, and in stories about siblings. 6.15 Murderous adultery Elements: two Adulterers; a betrayed Spouse who is killed or destroyed. This situation combines betrayal with violence: the illicit couple destroy the person who stands between them. Polti's placement is instructive, because he does not treat this as a variant of ordinary #adultery. The addition of murder changes the structure and not merely the intensity, since it converts the third party from an obstacle into a victim and converts the lovers from transgressors into criminals. It is one of the darkest categories in the list, and its persistence in crime fiction and film noir shows how efficiently it fuses desire, guilt and dread. 6.16 Madness Elements: a Madman; a Victim. Polti treats #madness as a dramatic situation rather than as a character trait, and the distinction matters. What creates the situation is that a person whose judgement has failed still has power to act, and someone else lies within reach of that power. The victim may be a stranger, a loved one, or the madman himself. Modern readers should handle this category with care, since nineteenth century categories of insanity carry assumptions that current medical and ethical understanding rejects. Read structurally, however, the category identifies something real: narratives in which the ordinary rules of motive and consequence are suspended, and in which the audience cannot predict action from character in the usual way. 6.17 Fatal imprudence Elements: the Imprudent person; the Victim or the Object lost. A single careless act destroys something irreplaceable. There is no villain. There is only a decision that seemed small and was not. Polti includes curiosity, carelessness, credulity and haste among the causes. This is arguably the most underused situation in the list and one of the most modern in feeling, because it captures a world in which harm arises from ordinary human weakness rather than from evil intention. It is the structure behind narratives of #negligence, of accidents, of systems failure, and of the small lie that grows. 6.18 Involuntary crimes of love Elements: a Lover; a Beloved; a Revealer. Two people love, and then learn that their love is forbidden by a fact they did not know, most classically a fact of kinship. The essential element is the Revealer, the agent or event that brings the knowledge, because the situation only exists at the moment of discovery. The category demonstrates a general principle of great importance to students: in Polti's system, knowledge is a dramatic force. The facts of the world do not create drama by themselves. Drama is created when a character learns them. 6.19 Slaying of a kinsman unrecognised Elements: the Slayer; the unrecognised Victim; the Revealer of the identity. This is the tragic engine of misrecognition. A person kills, injures or condemns someone who turns out to be a relative, a child, or a friend. The horror is not in the act but in the discovery after the act, when the deed can no longer be undone. Together with the previous category, it shows that Polti built a whole region of his map around the gap between what a character knows and what is true. This region is the structural home of #dramatic_irony, since the audience often possesses the fatal knowledge before the character does, and watches the character walk toward it. 6.20 Self sacrifice for an ideal Elements: a Hero; an Ideal; the Person or Thing sacrificed. Someone gives up something precious, possibly life itself, for a cause, a faith, a principle, or a duty. The structure requires that the thing given up be genuinely valuable, or the sacrifice is empty. The dramatic tension in this situation is not between hero and enemy but between two goods. The most powerful examples are those in which the ideal demands the sacrifice of a person the hero loves, which places this situation next to its darker relatives later in the list. 6.21 Self sacrifice for kindred Elements: a Hero; a Kinsman; the Person or Thing sacrificed. The same structure of #sacrifice, redirected from an abstract ideal to a concrete family bond. A parent ruins themselves for a child; a sibling takes the blame for a sibling; a spouse gives up a future for a partner. The pairing of these two categories is one of the clearest signs that Polti's list is organised by relational logic rather than by subject matter. The act is the same. The addressee of the act is different, and that difference changes everything about the meaning of the work. 6.22 All sacrificed for a passion Elements: a Lover; the Object of the fatal passion; the Person or Thing sacrificed. Here the ordering of values collapses. A person destroys career, family, honour and safety for a desire, and the desire need not be romantic: it may be for money, for revenge, for status, or for an obsession of any kind. The situation is the structural home of tragic obsession, and its dramatic requirement is severe: the passion must be made comprehensible to an audience that would not choose it. Without that, the work becomes a case study rather than a drama. 6.23 Necessity of sacrificing loved ones Elements: a Hero; a Beloved Victim; the Necessity for the sacrifice. This is the dilemma in its most brutal form. A person must destroy what they love, because a duty, a law, a god, or a hard arithmetic of survival demands it. Notice the third element. Necessity occupies a structural role, exactly as remembrance did in situation four. Polti's willingness to give a non human pressure the status of an element is one of his most useful contributions, because it lets the framework describe stories in which the antagonist is a circumstance rather than a person. 6.24 Rivalry of superior and inferior Elements: a Superior Rival; an Inferior Rival; the Object of the rivalry. Two people compete for the same object, but they are unequal in power, rank, wealth, beauty, or status. Polti's list of variants is long, covering rivals of different class, of different age, of different fortune, and of different nature. The inequality is what generates the drama, because it converts a simple contest into a question of justice. The audience is invited to ask not only who will win but who ought to, and this is why the situation has been so central to narratives of class, of romance across social boundaries, and of the outsider who challenges an establishment. 6.25 Adultery Elements: a deceived Spouse; two Adulterers. Betrayal within a bond of exclusive trust. The situation is entirely about the triangle, and its dramatic force depends on the audience being able to see all three positions at once, which is why the form so often uses dramatic irony, with the audience knowing what the deceived party does not. Polti's variants sort the triangle by sympathy: the rival may be more attractive or less, the betrayed spouse may be worthy or contemptible, and the adulterer may act from love, weakness, revenge, or boredom. The sorting shows how much of the meaning of a situation is set by whom the audience is invited to pity. 6.26 Crimes of love Elements: a Lover; a Beloved. Polti groups here transgressive attachments that violate strong social prohibitions. The categories he used reflect the moral assumptions of his era, and some of them are today understood as abuse rather than as transgressive love, which must be stated plainly. For the modern student, the useful abstraction is this: the situation describes narratives in which desire collides with a boundary that the surrounding society treats as absolute. What counts as absolute changes across cultures and periods, and tracking those changes is one of the most powerful ways to use the framework as a tool of cultural analysis rather than as a fixed moral chart. 6.27 Discovery of the dishonour of a loved one Elements: a Discoverer; the Guilty One. Someone learns that a person they love has done something shameful, and the drama becomes one of #discovery. The situation is again epistemic, and again the drama belongs to the moment of knowledge and to what follows it: concealment, denunciation, forgiveness, or complicity. This is one of the most contemporary categories in the entire list. It is the structure of narratives about a parent's hidden past, a partner's secret life, a mentor's corruption, and an institution's cover up. Its central question is not what the guilty person did, but what the discoverer will now do. 6.28 Obstacles to love Elements: two Lovers; an Obstacle. The most familiar situation in popular narrative, and the home of #obstacles_to_love. Two people wish to be together; something prevents it. The obstacle may be a family, a law, a difference of rank or wealth, a war, an illness, a prior commitment, or an inner refusal. Polti's insistence that the obstacle is a structural element, and not merely a plot detail, is worth pausing over. The nature of the obstacle determines the genre. An external obstacle that can be defeated produces romance. An external obstacle that cannot be defeated produces tragedy. An internal obstacle produces psychological drama. The same two lovers, with three different obstacles, produce three different works. 6.29 An enemy loved Elements: a Lover; the Beloved who is an enemy; a Hater who is kin to the lover. Love across a line of enmity, with the added pressure of a third party who is bound to the lover and hates the beloved. This is the classic structure of the feud crossed by love, and it is one of the few situations in which Polti explicitly builds the social group into the triangle. Its power comes from the impossibility of a private solution, since the #feud is public. The lovers cannot simply agree to love each other, because their bond is a public act with consequences for others, and the third element ensures that the world will not let them alone. 6.30 Ambition Elements: an Ambitious person; the Thing coveted; an Adversary. Ambition drives a person upward through or over others. Polti includes ambition for power, for wealth, for reputation, and for achievement, and he notes the important variant in which the adversary is a relative or a friend, since the cost of rising is then measured in relationships. The situation is neutral about morality, which makes it especially useful for analysis. Ambition can produce the hero of a founding narrative or the villain of a tragedy, and the difference lies not in the situation but in what the ambitious person is prepared to sacrifice, which links this category directly to the sacrifice family. 6.31 Conflict with a god Elements: a Mortal; an Immortal. A human being contends with a power vastly greater than themselves, whether a god, a fate, a natural order, or a moral law of the universe. Polti's variants include the struggle against a deity, the punishment for defying one, and the wager or contest with one. Students often assume this situation is obsolete in secular narrative. It is not. Once the immortal element is understood functionally, as an overwhelming and non negotiable power, the category covers stories of human beings against fate, against an implacable system, against death itself, and against forces of nature that cannot be reasoned with. It is arguably the situation with the greatest untapped potential in contemporary writing, since it is the natural structure for narratives of climate, catastrophe, technology and mortality. 6.32 Mistaken jealousy Elements: the Jealous one; the Object of jealousy; a supposed Accomplice; the Cause or Author of the mistake. Four elements, and the fourth is the key. Jealousy that is merely felt is a state. Jealousy that has been engineered by another person, or by a misleading accident, is a situation, because it contains a hidden agent and a false belief. The category is one of Polti's finest analytical achievements, because it separates the emotion from the mechanism. What drives the drama is the manufactured error, and this is why the manipulator figure who plants a suspicion is one of the most feared characters in the theatre. 6.33 Erroneous judgement Elements: the Mistaken one; the Victim of the mistake; the Cause or Author of the mistake; the Guilty person. A person judges wrongly, and someone innocent suffers for it, which is #erroneous_judgement. The elements make the moral architecture explicit: there is a person who errs, a person harmed by the error, a source of the error, and often a guilty party who escapes because of it. This situation is the structural core of narratives of miscarried justice, of wrongful accusation, of prejudice, and of institutional failure. It is also the sober counterpart of the enigma: the enigma is a search for truth, and erroneous judgement is what happens when the search fails or is not conducted at all. 6.34 Remorse Elements: the Culprit; the Victim or the Sin; the Interrogator. #Remorse turns the drama inward. The conflict is between a person and their own past act. The presence of an Interrogator among the elements is telling, because Polti understood that guilt becomes dramatic only when it is externalised, whether by another character, by a threat of exposure, or by a demand for confession. This is the situation that most directly connects Polti's list to modern psychological narrative, and it is the one that most clearly rewards a slow, interior treatment in prose fiction. 6.35 Recovery of a lost one Elements: the Seeker; the One found. Someone lost is found, which is the situation of #recovery: a missing child, a vanished parent, a separated sibling, a friend believed dead. The situation is the structural inverse of loss, and it explains the enduring power of the reunion scene. Polti notes that the search itself is often more dramatically productive than the discovery, which is a practical warning to writers. The dramatic energy is in the seeking, and a reunion that arrives too easily discharges the tension it was meant to release. 6.36 Loss of loved ones Elements: a slain Kinsman; a Kinsman who witnesses; an Executioner. The final situation is #grief, and Polti characteristically defines it not by the emotion but by the positions: someone dies, someone must watch and survive, and some agent is responsible. That third element deserves attention. By insisting on an Executioner, Polti keeps even mourning within the field of conflict, since grief in drama tends to demand an object, whether a killer, a system, a god, or a self to blame. When there is no one to blame, the survivor frequently invents someone, and that invention is itself a dramatic act. 7. Analysis: families of situations and the hidden structure of the list Presented as a numbered list, the thirty-six situations look like an inventory. Studied carefully, they look more like a map. The following grouping is analytical rather than historical: Polti did not organise his book this way, but the groups emerge naturally once we ask what kind of pressure creates the conflict in each case. This is the core analytical contribution of the present article, and students may use it as a template for their own coding work. 7.1 The family of power and authority Situations one, two, five, eight, twelve and twenty four share a structure in which the decisive force is authority, whether legitimate or not. Someone can grant or withhold, punish or pardon, permit or forbid. Supplication, deliverance, pursuit, revolt, obtaining and the rivalry of superior and inferior all turn on an asymmetry of power, and their dramatic question is always the same: will the powerful yield? This family explains why so much political and legal narrative feels similar in shape even when its content differs. It also explains why writers who want to escape the shape must attack the asymmetry itself, by making the powerful vulnerable or the powerless resourceful. 7.2 The family of kinship Situations four, thirteen, fourteen, nineteen, twenty one, twenty three and thirty six are all generated by the bond of kinship. They form the densest cluster in the entire list, and the density is itself a finding. It tells us that in the corpus Polti studied, the family was the primary machine of tragedy. The reason is structural. Kinship supplies three things that drama needs and that few other relations supply at once: obligation that cannot be dissolved, proximity that cannot be avoided, and love that survives injury. When conflict occurs inside such a bond, no simple resolution is available, and the result is the specific tragic pressure that the Greek theatre discovered and that family sagas continue to exploit. 7.3 The family of desire Situations fifteen, twenty two, twenty five, twenty six, twenty eight, twenty nine and thirty are driven by wanting. The object may be a person, a position, or a possession. What unites them is that the protagonist's desire is in conflict either with a social boundary or with another person's claim. Within this family, the key analytic variable is the location of the obstacle. When the obstacle is external, the story becomes a struggle. When it is internal, the story becomes a study of character. When the obstacle is a moral prohibition that the protagonist accepts and violates anyway, the story becomes tragedy of transgression. 7.4 The family of knowledge Situations eleven, eighteen, nineteen, twenty seven, thirty two and thirty three all turn on what a character knows, does not know, or wrongly believes. This family is the most theoretically interesting in the list, because it shows that Polti had grasped something that later #narratology would formalise: that the distribution of information among characters, and between characters and audience, is itself a source of dramatic form. Everything we now call suspense, surprise and dramatic irony can be described in terms of this family. Suspense arises when the audience knows a danger the character does not. Surprise arises when neither knows. Irony arises when the audience knows the meaning of an act that the actor misunderstands. Polti did not use these terms systematically, but his categories encode them. 7.5 The family of catastrophe and endurance Situations six, seven, sixteen, seventeen, thirty one, thirty four and thirty five concern forces that are not fully controllable: disaster, cruelty, disordered mind, accident, the overwhelming power of gods or fate, the weight of a past act, and loss followed by search. Here the antagonist is often not a person at all. The existence of this family is the strongest refutation of the common claim that Polti's framework reduces all drama to person against person conflict. It does not. It has room for the impersonal, and this is precisely what makes it usable for contemporary narratives about systems, environments, illnesses and technologies. 7.6 What the map shows Three conclusions follow from the grouping. First, the list is unevenly distributed. Kinship and desire are heavily populated; impersonal catastrophe is thinly populated. This reflects the corpus that Polti read, which was dominated by classical and French tragic drama, and it warns us that the framework is not a neutral instrument but a portrait of a particular literary tradition. Second, situations combine. Real narratives almost never contain one situation. They stack them. A single film may open with an enigma, develop through a daring enterprise, turn on an erroneous judgement, and close with remorse. The framework is therefore best used as a vocabulary for describing a sequence, not as a label for a whole work. Third, the boundaries are soft. Ambition shades into rivalry; rivalry shades into enmity; enmity shades into vengeance. This softness is a defect if we want a strict taxonomy, but it is a virtue if we want an analytical language, because it mirrors the way conflicts actually mutate inside a narrative as characters act and are changed by acting. 8. Applications 8.1 Literary and film analysis Used carefully, the framework produces a specific and testable kind of reading. Instead of asking what a work is about, the analyst asks which configurations of roles are active, when they change, and who occupies which position at each stage. This produces observations that can be checked. If two readers disagree about whether a scene is #supplication or #obtaining, they can argue about the roles, and the argument will be productive because the categories force them to say what they think the power relations are. The most valuable analytical move is to track role migration. Characters do not stay in one position. The suppliant of act one becomes the power in authority of act three; the avenger becomes the criminal; the rescuer becomes the threatener. Charting these migrations is a rigorous way to describe what critics loosely call a character arc. 8.2 Creative writing and screenwriting For writers, the framework is a diagnostic instrument, not a generator. Its most reliable use is to ask, of any scene that feels flat, a single question: what is the situation here, and who occupies each role? Scenes that feel flat usually turn out to contain no situation at all. Two characters are exchanging information, and nothing is at stake for either. A second use is combinatorial invention. Because situations can be stacked, a writer can deliberately layer a second situation over a scene that already works. A negotiation scene, which runs on #obtaining, becomes far more charged if one of the two parties is also inside a #remorse situation, or if a third party in the room is engineering a mistaken jealousy. A third use is deliberate underuse. Polti's own complaint was that writers crowded into a few situations and abandoned the rest. Students who want original material should look at the thin regions of the map, particularly fatal imprudence, erroneous judgement, conflict with an overwhelming power, and self sacrifice for an ideal, all of which are structurally rich and comparatively neglected in popular contemporary narrative. 8.3 Teaching and curriculum In the classroom, the framework has a specific pedagogical value that is often missed. It gives beginners a vocabulary of relations at exactly the moment when they are most tempted to think in terms of subjects. A beginning writer says the story is about jealousy. The framework asks: who is jealous, of whom, on the basis of what belief, and who planted that belief? The second question produces a story; the first does not. Research on writing instruction has increasingly examined how structured supports and, more recently, generative tools shape the composing process of student writers, including their motivation and confidence (Fang, Ng, Leung, and Chu, 2023; Oruc, Aktas, and Kurt, 2026). A role based framework fits naturally into this line of work, because it provides a scaffold that is explicit enough to teach and to assess, while remaining open enough not to dictate content. Similar findings in digital storytelling contexts suggest that collaborative and structured approaches can strengthen students' writing outcomes (Tanrikulu, 2022). A practical classroom exercise follows from the framework directly. Give students a single situation, fix the roles, and require four different works from it, changing only one variable at a time: the sympathies of the audience, the obstacle, the location of knowledge, and the outcome. Students discover through practice what the theory asserts, namely that the situation is a machine and not a message. 8.4 Game design and interactive narrative The framework has found an unexpected second home in interactive media. In a game, the author does not write a story; the author writes a system from which stories may emerge, and this creates a hard problem: how to design low level rules that reliably produce high level events that a human being will recognise as a story. Dramatic situations have been proposed as a solution to exactly this problem, functioning as an intermediate heuristic that is concrete enough to guide the design of underlying mechanics and abstract enough not to force a single outcome; an experimental system embedding a subset of Polti's situations found that recognising a dramatic situation prompts a player to reconstruct its causes and to anticipate its consequences, which is close to what we mean by experiencing a story (Lessard and Pare-Chouinard, 2022). The same work notes the central risk of the approach, which is that procedurally generated situations become repetitive, and that recognisability can turn into predictability. This application also explains why the concept of #tellability matters. A situation is worth telling because it is rare in ordinary life and because the stakes are high. That is, in effect, a design criterion, and it can be stated in advance. 8.5 Computational narratology and artificial intelligence The most active current use of the framework lies in language technology. Automatic story generation has been an ambition for decades, and surveys of the field show that the central difficulties are not fluency but structure, coherence, and interest (Alhussain and Azmi, 2021; Wang, Lin, Yu, Hu, and Karlsson, 2023). Systems can produce readable sentences long before they can produce a plot that goes anywhere. Because of this, researchers have looked for compact structural priors that can be given to a model, and Polti's situations are attractive precisely because they are compact, role based, and outcome neutral. Recent work on interactive drama with large language models uses a condensed set of dramatic situations, together with narrative techniques such as suspense and reversal, to guide generation, and reports improved narrative structure and improved player experience compared with unguided baselines (Wu, Wu, Xu, Zhang, and Zhao, 2025). Related work on planning for suspense shows that explicitly modelling a dramatic effect, rather than hoping it appears, produces measurably better stories (Xie and Riedl, 2024), and studies of role play systems have moved from character imitation toward genuine dramatic interaction (Wu, Wu, Jiang, Liu, Zhao, and Zhang, 2024). Two cautions belong here. First, evaluation is unsolved. The community has repeatedly shown that automatic metrics correlate poorly with human judgements of story quality, and that human criteria themselves need careful definition (Chhun, Colombo, Suchanek, and Clavel, 2022; Chhun, Suchanek, and Clavel, 2024). A framework that improves structure on paper may not improve the reading experience. Second, using a nineteenth century tragic taxonomy as a prior for a generative system bakes the biases of that taxonomy into the output. If the map over represents kinship tragedy and under represents impersonal catastrophe, so will the machine. Still, the direction of travel is clear. Narrative theory is being asked to become operational, and frameworks that specify roles and relations are the ones that survive that demand. This is why an idea from 1895 is being cited in engineering venues in the 2020s, and why calls to connect computational narrative work to narratological theory are increasingly heard (Piper, So, and Bamman, 2021). Human centred systems that scaffold rather than replace the writer are an especially promising line, since they use structure to prompt invention instead of substituting for it (Chung, Kim, Yoo, Lee, Adar, and Chang, 2022). 9. Criticisms and limitations Any responsible academic treatment must state the weaknesses of the framework clearly. There are six. 9.1 The number is arbitrary Thirty six is inherited, not derived. It comes from an anecdote about Gozzi, and Polti's own procedure produced a much larger set of variants which he then compressed into thirty six headings. Souriau's counter proposal of two hundred thousand situations makes the point unmistakably: the number is a function of the grain of description. Any student who repeats the claim that there are exactly thirty six situations is repeating a rhetorical flourish, not a finding. 9.2 The categories overlap Ambition, rivalry, enmity and vengeance are not cleanly separable. Neither are supplication, deliverance and obtaining. Neither are the enigma and erroneous judgement. In a strict taxonomy, categories are mutually exclusive; here they are not. This means that two competent analysts can classify the same scene differently, and neither is wrong. Overlap is not fatal to an analytical vocabulary, but it is fatal to any claim of scientific classification, and it is the reason the framework has never been validated in the way a research instrument must be validated. 9.3 There is no reliability evidence This is the most serious methodological gap. In modern #content_analysis, a coding scheme must demonstrate that independent coders, working from the same definitions, assign the same categories to the same material at a rate that exceeds chance. No such study of Polti's list exists in the published literature. Until it does, statements such as this film contains situation twenty four are informed interpretations, not measurements. This gap is also an opportunity. A well designed inter coder reliability study of the thirty six situations, on a defined corpus, would be a genuinely publishable contribution and is well within the reach of a good postgraduate project. 9.4 The corpus was narrow Polti read Greek tragedy and French drama, with a comparatively small selection of other traditions. The resulting map reflects the preoccupations of that corpus, which is why kinship tragedy is so heavily represented and why forms outside the European tragic tradition fit less comfortably. Narratives from oral traditions, from non Western dramatic forms, and from genres organised around cyclical rather than linear structure, may require categories that the list simply does not contain. The framework should therefore be treated as a description of a tradition, not as a description of humanity. This is a #cultural_bias in the strict methodological sense: the sampling frame determined the result. 9.5 The tragic bias The corpus was not only narrow but tonally skewed. The list is overwhelmingly a list of ways things go wrong. Comedy, satire, farce, and narratives of ordinary life are poorly served. There is no situation for the misunderstanding that resolves happily, for the reconciliation, for the joke that gets out of hand, or for the slow work of building something. Where comedy appears, it appears as an inversion of a tragic structure rather than as a structure in its own right. 9.6 Dated moral assumptions Several categories, particularly those concerning madness, crimes of love, and abduction, are framed in ways that encode nineteenth century assumptions about gender, consent, family authority and mental illness. Students must not import those assumptions along with the structure. The correct procedure is to keep the relational skeleton, which is what has lasting analytical value, and to discard the moral commentary, which does not. 10. Discussion If the framework has this many weaknesses, why does it persist? The answer lies in a distinction between two things that are easy to confuse: a classification and a heuristic. A classification aims to sort the world into exhaustive and mutually exclusive boxes, and it is judged by whether the boxes hold. A heuristic aims to make a difficult problem tractable, and it is judged by whether it produces good questions. Polti's list fails as a classification. It succeeds, remarkably well, as a heuristic. What does it make tractable? It makes conflict analysable. Before the analyst can ask any interesting question about a narrative, the analyst must be able to say what is pressing on whom, and why the pressure cannot simply be removed. Polti's roles do this work. They convert a vague intuition that a scene is tense into an explicit account of the tension: here is the party who wants, here is the party who refuses, and here is the third party whose decision is uncertain. This is also why the framework has migrated into fields Polti never anticipated. A game designer needs an explicit account of dramatic pressure in order to build systems that produce it. A researcher training a model to generate a story needs an explicit account in order to specify what a good story would even be. In both cases the requirement is the same: make the implicit structure explicit. That requirement is exactly what a role based situation supplies, and it is exactly what a thematic vocabulary, however sophisticated, cannot supply. There is a further theoretical point. The framework anticipates, in rough form, an insight that later narrative theory developed with much more precision, namely that narrative interest is a function of the distribution of knowledge and of the imbalance of power. Polti's knowledge family and power family, taken together, cover most of the mechanisms that produce suspense, irony and moral complexity. He arrived at these by sorting plays rather than by theorising, but the result is convergent, and convergence from a different direction is a reason to take a framework seriously even when its formal claims are indefensible. Finally, the framework is honest about something that a great deal of writing advice conceals. Stories are made of collisions. Not of ideas, not of atmospheres, not of voices, but of parties who want incompatible things and cannot walk away. A student who absorbs nothing else from Polti but this will have absorbed the most useful lesson in the book. 11. Implications for students and researchers Six practical recommendations follow from the analysis above. Use the framework as a language, not as a law. Say that a scene runs on #vengeance, and then defend the claim by naming the roles. Do not say that a work is situation number three, as though the number settled anything. Analyse sequences, not works. Situations change across a narrative, and the change is the finding. A study that maps the sequence of situations in a set of related works, and shows how the sequence differs across genres or periods, is far more interesting than a study that assigns one label to each work. Attend to role migration. The most revealing question a student can ask of any narrative is which character changes position, and at what cost. This is where the framework connects with characterisation, and it is where most good essays about it are actually written. Look for the missing situations. If the map is uneven, the empty regions are where original work lives. This applies to creative writers looking for material and to researchers looking for a topic. Test the framework rather than assuming it. The obvious research projects are open and unclaimed: a reliability study of the coding scheme; a comparative study of situation distribution across cultures or across a century of film; an examination of whether situation based conditioning improves the perceived quality of machine generated stories under rigorous human evaluation. State your interpretive act. Because the categories overlap and the original coding decisions are lost, every application of the framework involves judgement. Good scholarship says so, gives its criteria, and lets the reader disagree. 12. Conclusion The thirty six dramatic situations are neither a discovery nor a hoax. They are an instrument, built in the late nineteenth century by a critic who wanted to make the study of drama rigorous, using a number he borrowed from an anecdote and a corpus he chose from his own tradition. Judged as a scientific classification, the framework fails on several counts: the number is arbitrary, the categories overlap, the corpus was narrow and tonally skewed, and the coding has never been validated. Judged as an analytical heuristic, it has proved to be one of the most durable ideas in the study of narrative. It insists that drama is relational, that conflict is a configuration of roles rather than a subject matter, that knowledge is a dramatic force, and that impersonal pressures such as necessity, remembrance and fate can occupy structural positions just as persons do. These claims were unusual in 1895. They remain correct, and they are precisely the claims that make the framework usable in a research environment its author could not have imagined, from the design of emergent narrative systems to the structural conditioning of generative language models. For the student, the practical lesson is simple and worth stating without ornament. Do not ask what your story is about. Ask who wants what, who refuses, who decides, who knows, and who cannot walk away. The thirty six situations are thirty six answers to that question. There are almost certainly more. Finding them is the work. References Primary and historical sources Gozzi, C. (eighteenth century). The claim of a finite number of tragic situations, transmitted through later reports and discussed in Polti's introduction. Polti, G. (1921). The Thirty Six Dramatic Situations (L. Ray, Trans.). Franklin, Ohio: James Knapp Reeve. (Original French edition published 1895; later French editions issued by Mercure de France.) Propp, V. (1968). Morphology of the Folktale (2nd ed.). Austin: University of Texas Press. (Original work published 1928.) Souriau, E. (1950). Les deux cent mille situations dramatiques. Paris: Flammarion. Recent scholarship Alhussain, A. I., and Azmi, A. M. (2021). Automatic story generation: A survey of approaches. ACM Computing Surveys, 54(5), Article 103. https://doi.org/10.1145/3453156 Alhussain, A., and Azmi, A. M. (2024). Beyond event centric narratives: Advancing Arabic story generation with large language models and beam search. Mathematics, 12. Chhun, C., Colombo, P., Suchanek, F. M., and Clavel, C. (2022). Of human criteria and automatic metrics: A benchmark of the evaluation of story generation. In Proceedings of the 29th International Conference on Computational Linguistics. Chhun, C., Suchanek, F. M., and Clavel, C. (2024). Do language models enjoy their own stories? Prompting large language models for automatic story evaluation. Transactions of the Association for Computational Linguistics, 12, 1122 to 1142. Chung, J. J. Y., Kim, W., Yoo, K. M., Lee, H., Adar, E., and Chang, M. (2022). TaleBrush: Sketching stories with generative pretrained language models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, Article 209. https://doi.org/10.1145/3491102.3501819 Fang, X., Ng, D. T. K., Leung, J. K. L., and Chu, S. K. W. (2023). A systematic review of artificial intelligence technologies used for story writing. Education and Information Technologies, 28(11), 14361 to 14397. https://doi.org/10.1007/s10639-023-11741-5 Lessard, J., and Pare-Chouinard, S. (2022). Dramatic situations for emergent narrative system authorship. In M. Vosmeer and L. Holloway-Attaway (Eds.), Interactive Storytelling, ICIDS 2022, Lecture Notes in Computer Science, vol. 13762 (pp. 217 to 228). Cham: Springer. https://doi.org/10.1007/978-3-031-22298-6_14 Oruc, T., Aktas, I., and Kurt, M. (2026). The effects of artificial intelligence assisted creative writing on students' writing motivation, writing anxiety, and creative writing skills. Journal of Educational Computing Research. https://doi.org/10.1177/07356331251409998 Piper, A., So, R. J., and Bamman, D. (2021). Narrative theory for computational narrative understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 298 to 311). Association for Computational Linguistics. https://doi.org/10.18653/v1/2021.emnlp-main.26 Piper, A., and Wu, S. (2025). Evaluating large language models for narrative topic labeling. In Proceedings of the 5th International Conference on Natural Language Processing for Digital Humanities (pp. 281 to 291). Association for Computational Linguistics. Shen, J., Sap, M., Colon-Hernandez, P., Park, H., and Breazeal, C. (2023). Modeling empathic similarity in personal narratives. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 6237 to 6252). Association for Computational Linguistics. Tanrikulu, F. (2022). Students' perceptions about the effects of collaborative digital storytelling on writing skills. Computer Assisted Language Learning, 35(5 to 6), 1090 to 1105. Wang, Y., Lin, J., Yu, Z., Hu, W., and Karlsson, B. F. (2023). Open world story generation with structured knowledge enhancement: A comprehensive survey. Neurocomputing, Article 126792. Wu, H., Wu, W., Xu, T., Zhang, J., and Zhao, H. (2025). Towards enhanced immersion and agency for LLM based interactive drama. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. Wu, W., Wu, H., Jiang, L., Liu, X., Zhao, H., and Zhang, M. (2024). From role play to drama interaction: An LLM solution. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 3271 to 3290). Association for Computational Linguistics. Xie, K., and Riedl, M. (2024). Creating suspenseful stories: Iterative planning with large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics. Hashtags #Georges_Polti #Thirty_Six_Dramatic_Situations #dramatic_situation #plot_taxonomy #narratology #narrative_theory #dramaturgy #conflict #screenwriting #playwriting #creative_writing #story_structure #computational_narratology #literary_analysis #media_studies