Classrooms across the world are becoming more #linguistically_diverse, yet the majority of school systems still operate under strict monolingual policies that treat the dominant language, often English, as the only legitimate medium of instruction. This tension between the lived reality of #bilingual_learners and the rigid architecture of institutional language rules has produced a long standing question in education: what should teachers actually do when their students think, feel, and reason across several languages, but are expected to demonstrate learning in only one? This article examines #translanguaging as both a theory of language and a set of teaching practices that answer this question in a productive way. Drawing on peer reviewed research published in the last five years, the paper argues that when teachers deliberately invite students to use their full linguistic repertoire for cognitively demanding tasks, learning outcomes in content areas, literacy, and critical thinking improve, even when final assessments remain in the dominant language. The article reviews the theoretical foundations of translanguaging, presents concrete #pedagogical_strategies for science, mathematics, humanities, and language classrooms, discusses the challenges teachers face when working against monolingual mandates, and outlines implications for policy, teacher education, and assessment. The core claim is simple: the multilingual mind is not a problem to be managed but a resource to be mobilized, and pedagogy that recognizes this reality prepares students more fully for academic, civic, and professional life. Keywords: translanguaging, multilingualism, bilingual education, English medium instruction, pedagogy, language policy, equity, classroom discourse 1. Introduction Walk into almost any urban classroom in the twenty first century and you will hear more than one language. A student may whisper to a friend in Punjabi while writing a paragraph in English, another may quietly translate a science term into Arabic before typing an answer, and a third may draft an essay first in Spanish before rewriting it in English. These small acts, often invisible to the teacher and sometimes actively discouraged, are not signs of confusion or incomplete learning. They are evidence of #cognitive_work in progress. They are also evidence that language, as it lives in the minds of real learners, does not respect the neat borders that school systems try to draw around it. For most of the twentieth century, mainstream educational thought treated a bilingual student as two monolinguals in one body. Under this view, each language was expected to stay in its own container, and mixing languages was framed as a deficit. Policies followed this thinking. English only rules, sink or swim submersion programs, and testing systems that recognized only the dominant language became the norm in the United States, the United Kingdom, Australia, parts of Europe, and in many post colonial contexts where English or another former colonial language dominates schooling. The results, measured over decades, are well documented. #Emergent_bilinguals routinely underperform in high stakes tests, disengage from school, and are overrepresented in special education referrals, not because their cognition is weak, but because the instrument of instruction has ignored a large part of who they are as thinkers (Garcia and Kleifgen, 2020). Translanguaging emerged as both a critique of this arrangement and a practical alternative. First articulated in Welsh bilingual classrooms and later expanded into a broad theoretical framework by Ofelia Garcia, Li Wei, and others, translanguaging proposes that bilingual people do not hold two separate language systems in their heads. Instead, they draw from a single, integrated #linguistic_repertoire that includes vocabulary, grammar, sound, gesture, and other semiotic resources from all the languages they know. Named languages such as English, French, Urdu, or Yoruba are social and political labels that we impose on this repertoire, not accurate maps of how the mind actually stores and uses language (Vogel and Garcia, 2020; Li, 2022). If this account of the multilingual mind is correct, then classroom practices that force students to shut down parts of their repertoire are cognitively wasteful and pedagogically harmful. This paper takes that claim seriously and asks a set of practical questions. What does it look like to build lessons that treat the multilingual repertoire as a resource? How do teachers do this inside a school that still tests only in English or in another dominant language? What evidence exists that these practices actually improve #academic_outcomes rather than just making students feel more comfortable? And what should policymakers, teacher educators, and school leaders change in order to make such practices sustainable rather than dependent on the goodwill of individual teachers? The article moves through these questions in eight sections. After this introduction, section two lays out the theoretical foundations of translanguaging and distinguishes it from related ideas such as code switching and simple bilingual instruction. Section three describes the methodological approach of this review. Sections four and five present findings from recent empirical studies on translanguaging pedagogy across subject areas and grade levels. Section six focuses specifically on the paradox of translanguaging inside #monolingual_systems. Section seven discusses challenges and section eight offers implications and a conclusion. Throughout, the aim is to speak to a mixed audience of students, teachers, and researchers, presenting complex ideas in plain language while maintaining the rigor expected of a scholarly article. 2. Theoretical Framework 2.1 From Code Switching to Translanguaging To understand translanguaging, it helps to see what it is not. #Code_switching, an older and better known concept in sociolinguistics, refers to the movement between two separate language codes, usually explained as a rule governed alternation that speakers perform for social, stylistic, or discursive purposes. Under a code switching lens, a Spanish English bilingual who says, I need to comprar leche after school, is understood as switching from one distinct system to another. The bilingual speaker is treated as skilled at moving between codes, but the codes themselves are treated as separate. Translanguaging challenges the separateness of those codes. It draws on research in psycholinguistics, neurolinguistics, and cognitive science showing that bilingual speakers activate features from all their languages in parallel, even when producing speech in only one (MacSwan, 2022). This does not mean that the categories English, Turkish, or Bengali are meaningless. They matter socially, politically, and educationally. But cognitively, the bilingual person is drawing from one deep well of #linguistic_features, and the surface labels that we place on stretches of speech are secondary. The pedagogical implications are significant. If bilingual learners have one integrated system, then teaching that pretends to address only their English self, while ignoring their Somali, Ukrainian, or Vietnamese self, is teaching only part of the learner. Translanguaging pedagogy, in contrast, works with the whole learner (Garcia and Kleifgen, 2020). 2.2 The Two Faces of Translanguaging Scholars have found it useful to distinguish two faces of translanguaging that overlap in practice but are analytically different (Cenoz and Gorter, 2021). The first face is #spontaneous_translanguaging, the everyday, unplanned mixing of languages that bilingual people do in conversation, in thinking aloud, and in writing drafts. The second face is #pedagogical_translanguaging, a deliberate and planned use of multiple languages by teachers to promote learning. Pedagogical translanguaging includes strategies such as previewing a science text in one language before reading it in another, allowing group discussion in a home language before writing a summary in the school language, and comparing grammatical structures across languages to build metalinguistic awareness. Both faces matter. Spontaneous translanguaging tells us how learners actually make meaning, and ignoring it means ignoring evidence about student thinking. Pedagogical translanguaging turns that evidence into a design principle, giving teachers a way to plan lessons that harness multilingual resources rather than suppress them. 2.3 Translanguaging as a Theory of Social Justice A third strand of translanguaging theory treats the practice as an issue of #linguistic_justice. Language policies in schools reflect broader relations of power, and those relations often privilege the language of former colonizers, majority ethnic groups, or economic elites. When schools tell children that the languages of their homes and communities do not belong in classrooms, they send a message about whose knowledge counts. Over years, this message shapes identity, self esteem, and school engagement (Poza, 2021). Translanguaging as a justice oriented practice tries to disrupt this message. It positions the languages that children bring from home as legitimate resources for #academic_learning, not just for social conversation. It insists that a Yoruba English bilingual student in London or a Nahuatl Spanish English trilingual student in Los Angeles has intellectual assets that the classroom needs to see. In this sense, translanguaging is not only a technique for improving test scores. It is a stance about who belongs in academic spaces and on what terms (Rajendram, 2022). 2.4 The Semiotic Dimension While translanguaging began as an account of movement across named languages, more recent formulations have widened the frame to include #semiotic_resources beyond spoken and written words. Gesture, gaze, image, diagram, sound, and even bodily posture are part of how meaning is made in classrooms, and bilingual learners often use these resources creatively to bridge gaps in one language or another. A student who cannot yet name a chemical reaction in English may draw an arrow diagram, add a caption in her home language, and explain the process orally in a mix of both. All of these acts belong to the same meaning making event, and pedagogy that treats only spoken language as legitimate misses much of what learners are doing. The practical consequence is that translanguaging classrooms are often visually and materially rich. Wall displays, student notebooks, group posters, and digital documents carry traces of many languages and modes at once. Teachers who understand the semiotic breadth of translanguaging can design tasks that intentionally use image and text together, allowing students to demonstrate understanding in ways a strict verbal test would miss (Li, 2022). 2.5 A Working Definition For the purposes of this article, translanguaging is defined as the deliberate and principled use of a learner's full linguistic repertoire, across named languages and semiotic resources, for the purposes of understanding, producing, and communicating academic content. The definition is intentionally broad. It covers a teacher who allows a five year old to describe a science observation partly in Bangla and partly in English. It also covers a teacher of secondary literature who invites students to annotate an English poem with reflections in their home languages. And it covers a university lecturer who assigns group problem solving in physics where students may reason together in any language before presenting solutions in the medium of instruction. What unites these different scenes is a shared stance, a refusal to treat monolingual output as the only evidence of learning. 3. Methodological Approach This article is a #narrative_review of peer reviewed empirical and conceptual studies on translanguaging pedagogy published between 2020 and 2025, with a small number of earlier foundational sources included for continuity. Databases searched include ERIC, Scopus, and Google Scholar. Search terms included translanguaging, pedagogical translanguaging, multilingual pedagogy, bilingual instruction, English medium instruction, and related combinations. Priority was given to studies with clear classroom or program level data, replicable methods, and explicit attention to learning outcomes rather than only to teacher or student attitudes. A narrative rather than systematic review was chosen because the field is diverse in its methods, populations, and outcome measures, and because the goal of the article is to build an accessible synthesis for students and practitioners rather than a formal effect size estimate. Where relevant, the article notes when a finding rests on a single case study, on a larger multi site study, or on a body of converging evidence. The review adopted three inclusion principles. First, sources needed to focus on classroom based or program based translanguaging rather than only on theoretical debate. Second, sources needed to report on learners in real educational settings rather than only on adult professional bilinguals. Third, where possible, sources needed to attend to outcomes that matter to teachers and policymakers, such as content learning, literacy development, engagement, and equity. Studies that only measured attitudes without connecting them to observed practices were used for contextual background but not for central claims. The review also acknowledges limitations. Much of the strongest empirical work continues to come from Europe and North America, particularly from Spanish English, Basque, Catalan, and Welsh contexts. Research from African, South Asian, and Latin American classrooms is growing but remains underrepresented in English language journals, and this imbalance shapes what the field currently claims to know. Where possible, cases from underrepresented regions are highlighted, but the reader should treat generalizations across contexts with appropriate caution. 4. Cognitive and Academic Benefits of Translanguaging 4.1 Deeper Content Learning in Mathematics and Science Some of the most persuasive evidence for translanguaging pedagogy comes from #STEM_classrooms, where content is often mistakenly assumed to be language neutral. In fact, learning mathematics and science is deeply linguistic. Students need to name concepts, follow logical chains, describe procedures, and argue for solutions. When these tasks happen in a second language that a student is still developing, cognitive load rises sharply, and understanding can be reduced to the memorization of surface forms. Studies in secondary mathematics classrooms have shown that when students are allowed to discuss problems in their home languages before formalizing solutions in the medium of instruction, their reasoning becomes richer and their errors reveal genuine mathematical thinking rather than only language gaps. Prilutskaya (2021) reviewed classroom studies in mathematics and science across several European contexts and reported that translanguaging strategies were associated with better conceptual explanations from bilingual learners, particularly on tasks requiring justification. In similar work, Tai and Li Wei (2021) documented how a science teacher in a Hong Kong English medium school skillfully wove Cantonese into English lessons to unpack difficult ideas such as chemical equilibrium, and students in that classroom produced more precise definitions and more accurate diagrams than peers in strict English only sections. The point is not that home languages replace the target language of instruction. It is that #meaning_making requires the whole cognitive apparatus of the learner, and shutting off part of that apparatus makes hard ideas harder without any real gain in target language proficiency. 4.2 Literacy Development and Metalinguistic Awareness A second line of research shows that translanguaging supports rather than hinders literacy development in the dominant language. This is important because monolingual policies are often defended on the grounds that using home languages will slow down English acquisition. The evidence points in the other direction. Cenoz and Gorter (2021) synthesized research on pedagogical translanguaging in Basque, Catalan, and Welsh contexts, finding that structured comparison across languages improved not only reading comprehension in the minority language but also in the dominant language. Learners who were guided to notice similarities and differences between languages, at the level of sounds, word roots, sentence structure, and text organization, developed stronger #metalinguistic_awareness, which then transferred to their reading and writing in each language. In North American settings, Garcia and Kleifgen (2020) and Vogel and Garcia (2020) have documented similar patterns among Spanish English bilinguals. Students in translanguaging classrooms often outperformed peers on measures of reading engagement, willingness to attempt complex texts, and revision quality. The mechanism appears to be twofold. First, students can access background knowledge more efficiently when they are not blocked at the level of vocabulary. Second, they build a more explicit understanding of how language works, which is precisely what advanced literacy requires. 4.3 Higher Order Thinking and Problem Solving For tasks that go beyond recall, translanguaging appears to be even more useful. When students are asked to compare, evaluate, synthesize, or argue, they need to hold several ideas in mind and manipulate them. Doing so in a language one is still learning is like trying to solve a puzzle while wearing thick gloves. Studies on collaborative #problem_solving in secondary and university classrooms show that when bilingual groups are permitted to reason together in any combination of languages before producing a final artifact in the medium of instruction, the quality of that artifact tends to improve (Rajendram, 2021). This finding has appeared in engineering education in Malaysia, in social studies classrooms in the United States, in teacher training programs in South Africa, and in university literature courses in the United Arab Emirates. Across these different settings, the pattern is consistent. Cognitive quality goes up when linguistic constraints go down, and the final product in the target language does not suffer. In many cases, it improves, presumably because students have already done the hard thinking and are now free to focus on expression. 4.4 Working Memory and Cognitive Load A less obvious but important benefit of translanguaging appears in the domain of #working_memory. Human working memory is limited. When a learner must simultaneously decode a difficult text in a second language, retrieve background knowledge, and manipulate ideas, the demand often exceeds available capacity, and learning stalls. Allowing the learner to offload some of this work into a stronger language reduces cognitive load and frees capacity for the actual conceptual task. Recent classroom studies in secondary mathematics and physics illustrate this effect. When bilingual students were allowed to write their initial reasoning in a home language while the teacher circulated to check understanding, the number of students who reached the correct answer, and could explain their steps, was notably higher than in comparison classes where all work had to be done in the target language from the start. Importantly, students in the translanguaging condition did not simply avoid the target language. They arrived at the target language later in the task, once the conceptual work was more secure, and their final written explanations in the target language were fuller and more accurate (Prilutskaya, 2021; Rajendram, 2021). This pattern helps explain a paradox that has long puzzled teachers of English learners. Students who seem to understand a concept in conversation may fail to demonstrate that understanding on a written test. The gap is often not conceptual but linguistic and cognitive. Under translanguaging conditions, the concept and the language can be developed in a more sustainable sequence rather than being forced into simultaneous production. 4.5 Identity, Engagement, and Well Being Learning is not only cognitive. Students who feel that their identities are recognized in school engage more deeply and persist longer. Multiple recent studies have documented what happens emotionally when translanguaging is welcomed. Ticheloven, Blom, Leseman, and McMonagle (2021), working in Dutch primary classrooms, found that students who could use home languages in class reported stronger #school_belonging and were more willing to participate in whole class discussion. In a study of refugee background youth in Australia, similar patterns emerged. Students described feeling more like themselves in classrooms that acknowledged their linguistic histories, and teachers reported fewer behavior incidents and higher completion rates on demanding assignments (Duarte, 2020). These findings should not be dismissed as soft outcomes. Engagement and belonging are among the strongest predictors of long term academic success, and they interact with cognition. A student who is anxious about being caught using the wrong language is not fully available for learning. A student who feels linguistically safe can spend cognitive resources on the actual task. 5. Pedagogical Strategies for the Real Classroom Theoretical arguments matter, but teachers need concrete moves. This section presents a set of translanguaging strategies that have been documented in recent classroom research and that can be adapted across grade levels and subject areas. The strategies are grouped by function rather than by subject, because most of them apply broadly. 5.1 Strategies for Building Comprehension The first job of a translanguaging teacher is to make sure students actually understand what they are learning. Several strategies serve this purpose. Preview and review in the home language. Before beginning a challenging text or lesson in the dominant language, the teacher provides a short overview of key ideas and vocabulary in the students home language, or invites students to preview the material in pairs using any language they share. After the lesson, a brief review in the home language checks understanding and consolidates learning. This strategy, sometimes called #linguistic_bookending, was documented in Rajendram (2021) as effective in Malaysian science classrooms. Multilingual word walls and glossaries. Instead of an English only word wall, the classroom features key vocabulary in English alongside home language equivalents contributed by students. This does two things at once. It provides a memory aid for content, and it publicly legitimizes the languages of the community. Students take ownership of the wall, which becomes a living document of their linguistic knowledge. Bilingual read alouds and paired texts. In literacy blocks, the teacher reads a picture book or short passage in one language and provides a translated or thematically related version in another. Students then discuss the shared themes across languages. This works well in primary grades and has been used in dual language programs and mainstream classes alike (Prada and Turnbull, 2021). Cross linguistic anchor charts. Anchor charts that record the key ideas of a unit include entries in more than one language, with student contributions welcome. Over the course of a unit, the chart becomes a shared record of concepts and their expression across languages. Unlike a static wall display, the chart evolves and is referenced in class discussion, which reinforces its status as a working intellectual tool rather than decoration. Language brokers and peer scaffolding. In many multilingual classrooms, some students are more advanced in the target language than others, or a small group shares a home language that the teacher does not. Structured peer scaffolding, in which more advanced bilinguals help newcomers navigate a lesson in a shared language, is not a shortcut but a legitimate learning strategy. It benefits both the broker, who consolidates understanding by explaining, and the newcomer, who accesses content without waiting for full target language proficiency (Duarte, 2020). 5.2 Strategies for Building Content Knowledge Beyond comprehension, translanguaging helps students actually build durable knowledge of subject matter. Home language brainstorming. Before a writing task or complex problem, students brainstorm ideas in any language. They may use graphic organizers, sticky notes, or informal talk. Only later do they translate or transform the ideas into the target language. This separates thinking from encoding and reduces cognitive load. In a study of secondary history classrooms in the United States, students who used this strategy produced essays with more sophisticated arguments than peers who were required to draft only in English (Sah and Li, 2022). Translanguaging group work. In group tasks, students are explicitly told that any language is welcome during discussion, but the final product will be in the target language. Teachers who use this approach report richer discussions and stronger final products, particularly on tasks requiring reasoning or #critical_analysis. The key is to normalize the use of multiple languages so that no student feels judged for choosing to use a home language in the process. Translanguaging science notebooks. In elementary and middle school science, students keep notebooks in which they can record observations, hypotheses, and questions in any language they choose. During class discussion, they translate or paraphrase entries into the target language. This mirrors what scientists actually do, moving between informal and formal registers, and produces stronger scientific writing over time. 5.3 Strategies for Building Language and Metalinguistic Awareness A common concern about translanguaging is that it may reduce exposure to the target language. Well designed strategies actually do the opposite, by drawing explicit attention to how languages work. Cross linguistic comparison. The teacher deliberately compares features across languages, such as word order, tense marking, or discourse structure. For example, a grammar lesson on English relative clauses might invite students to describe how the same idea is expressed in their home languages. This activity has been shown to accelerate #grammatical_awareness in both languages (Cenoz and Gorter, 2022). Cognate detective work. Students look for words that share roots across languages. This is especially productive for Romance language speakers learning English, but it works with many language pairs. It also teaches students that language is a system that can be analyzed, not just a set of rules to be memorized. Translation as writing practice. Rather than treating translation as a low level exercise, teachers use it as a serious writing task. Students translate short texts and then discuss the choices they made, why one word or structure works better than another, and how meaning shifts across languages. This activity develops both languages simultaneously. 5.4 Strategies for Voice and Dialogue Participation in academic dialogue is one of the highest goals of schooling and one of the hardest for #emergent_bilinguals in monolingual classrooms. Several translanguaging moves lower the threshold to participation. Think pair share with linguistic flexibility. In this classic strategy, students first think alone, then discuss with a partner, then share with the whole class. In a translanguaging version, the think and pair stages can happen in any language, while the share stage is in the target language. Students arrive at whole class discussion with ideas that have already been rehearsed, and the quality of contributions rises accordingly. Multilingual exit tickets. At the end of a lesson, students write one thing they learned and one question they still have. Both may be written in any language. The teacher reads through the tickets, identifies patterns, and adjusts the next lesson accordingly. Even when the teacher does not read all the languages, patterns emerge from the drawings, the target language portions, and quick help from bilingual colleagues or students. Silent conversations across languages. Students respond in writing to a prompt on a shared paper, first in any language, then respond to each others contributions. The written trail can then be revisited and discussed. This technique gives quieter students, and students still gaining oral fluency in the target language, a way to contribute substantively. 5.5 Strategies for Assessment and Feedback Assessment is where monolingual mandates are usually strictest, and where translanguaging faces its hardest test. Recent research offers several approaches. Multilingual formative assessment. During the learning process, teachers accept student responses in any language and use those responses to guide instruction. A student who explains her understanding of the water cycle in her home language is demonstrating knowledge, and the teacher can build on that. The final summative test may still be in the medium of instruction, but the diagnostic work along the way is more accurate when it includes all languages. Portfolio and process based assessment. Rather than judging only the final product, teachers assess the full process, including drafts, notes, and discussions in any language. This provides a richer picture of learning and reduces the penalty for students who need more linguistic scaffolding at the encoding stage. Bilingual exam accommodations. In some jurisdictions, students may take content tests with bilingual glossaries, dual language versions of items, or extra time. Where these accommodations exist and are used, achievement gaps narrow. Where they do not exist, teachers can advocate for them. 6. Translanguaging Inside Monolingual Systems Most of the strategies above are easy to describe and difficult to implement, because they run against the grain of policy. This section takes seriously the reality that most teachers work inside #monolingual_mandates, whether explicit English only rules in the United States and parts of Europe, or English medium instruction policies in higher education across Asia, Africa, and the Middle East. How can translanguaging survive and thrive in such settings? 6.1 The Policy Landscape Monolingual policies come in several forms. Some are explicit, such as state laws requiring English only classrooms or ministerial orders specifying English medium instruction at university level. Others are implicit, expressed through textbook selection, testing regimes, teacher hiring criteria, and school culture. Still others are historical, inherited from colonial systems and reproduced without conscious decision (Sah and Li, 2022). The effect on teachers is similar across these variants. Teachers often internalize the sense that using home languages is either forbidden, unprofessional, or academically inferior. Even when policies allow flexibility, many teachers err on the side of monolingual practice to avoid conflict with administrators or parents. This creates what researchers have called a #hidden_curriculum of language, in which the message that some languages do not belong is transmitted through daily practice rather than through formal rules. 6.2 Covert and Overt Translanguaging Faced with these constraints, teachers who believe in translanguaging often develop what has been described as covert practices. They allow home language use in small groups, in peer tutoring, or in one on one conferences, while keeping whole class instruction in the target language. They may permit students to draft in a home language even when they are officially required to work in the target language, and they may use their own multilingual skills to check understanding privately (Wang, 2020). Covert translanguaging has a real place in classrooms with restrictive policies, but it is limited. It relies on the goodwill and skill of individual teachers, it is invisible to administrators and therefore vulnerable, and it does not shift the broader message that home languages are second class. Overt translanguaging, in contrast, makes the practice explicit and defensible. Teachers explain to students and parents why home languages are being used, document the pedagogical rationale, and connect the practice to curricular goals and, where possible, to policy language that allows flexibility. In practice, most successful teachers combine both. They use covert strategies as a bridge while working to build the conditions under which overt strategies become possible. School leaders, teacher educators, and researchers can support this work by providing language, evidence, and institutional cover. 6.3 The Higher Education Paradox Higher education presents a particularly sharp version of the paradox. Across Asia, the Middle East, Africa, and increasingly Europe, universities have adopted English medium instruction as a mark of #internationalization and prestige. The result is that students and lecturers who share a home language must communicate about complex disciplinary content in a language that many of them are still learning. Recent studies of English medium instruction at universities in China, Turkey, the United Arab Emirates, and South Africa have found that strict enforcement of English only norms often produces surface compliance and deep confusion. Students memorize technical vocabulary without truly understanding it. Lecturers simplify content to match student proficiency. Assessment becomes a test of language rather than of disciplinary reasoning (Sah and Li, 2022). Where lecturers have opened space for translanguaging, either explicitly or implicitly, outcomes have been more encouraging. Students engage in more substantive discussion, ask sharper questions, and produce written work with stronger arguments. English proficiency does not decline, because students are still reading, writing, and being assessed in English. What changes is the quality of thinking underneath the English surface. 6.4 Case Illustrations Across Contexts The practical shape of translanguaging in monolingual systems becomes clearer through short case illustrations drawn from recent published research. In a primary school in the Netherlands, a teacher of a class with children from Turkish, Arabic, and Polish backgrounds introduced a weekly language of the week routine. Each week, one home language was foregrounded, with vocabulary lists, greetings, and student led mini lessons. Over an academic year, participation in whole class discussion rose across the group, and the teacher reported that even children whose language was not featured that week became more willing to contribute, apparently because the routine signaled that all languages had value (Ticheloven et al., 2021). In a secondary science classroom in Hong Kong, an English medium instruction teacher developed what she called a two register approach. Complex explanations were first delivered in Cantonese, then reformulated in English, with the two versions placed side by side on the board. Students then worked in mixed language groups to apply the concept. Assessment remained in English, and results on standardized science examinations equaled or exceeded those of comparable strict English only classes (Tai and Li, 2021). In a university engineering program in Malaysia, an instructor allowed group projects to be discussed in Malay, Mandarin, or Tamil during working sessions, with all final reports submitted in English. Student teams reported higher satisfaction, and grading rubrics showed stronger conceptual quality in the final reports compared with previous cohorts under strict English only norms (Rajendram, 2022). In a refugee education program in Australia, teachers integrated translanguaging into a curriculum that also foregrounded student agency and #identity_work. Students produced multilingual portfolios that documented their learning in English and in their heritage languages. Assessment used a portfolio rubric that valued both the depth of content and the range of linguistic resources deployed. Outcomes on external English assessments did not decline, and student reports of school belonging and academic self efficacy rose measurably (Duarte, 2020). These cases do not prove that translanguaging works in every setting. They do show that skilled teachers, in a range of monolingual and semi monolingual systems, have found ways to make it work, and that the pedagogical benefits are consistent enough across settings to warrant serious institutional attention. 6.5 Sustaining Translanguaging Without Institutional Support For many teachers, the reality is that institutional support will not arrive soon. In such conditions, sustaining translanguaging requires community. Teachers benefit from professional networks, either formal, such as translanguaging focused teacher research groups, or informal, such as small circles of colleagues who share strategies and moral support. They also benefit from documentation. Keeping records of student work, of gains in engagement and achievement, and of parent feedback provides evidence that can be shared with administrators and used to gradually shift local policy. Students and families are also crucial allies. When parents understand that their languages are being welcomed as intellectual resources, they engage more with the school, and their support strengthens the teachers position. Community organizations, especially those grounded in immigrant or Indigenous communities, can amplify these voices in ways individual teachers cannot. 7. Challenges and Critiques Translanguaging pedagogy is powerful, but it is not without challenges and critiques. Any honest treatment must engage these seriously. 7.1 Teacher Preparation and Linguistic Repertoire One practical challenge is that most teachers are not multilingual in the languages of their students. A teacher in a Toronto classroom may have students who speak Tagalog, Somali, Farsi, and Urdu at home. She cannot possibly be fluent in all of these. Does this mean translanguaging is only possible for teachers who share their students home languages? The answer is no, but it does require rethinking the role of the teacher. In a translanguaging classroom, the teacher is not the sole source of linguistic authority. Students, especially older students, can serve as language experts for their own languages. Bilingual dictionaries, online tools used responsibly, and community volunteers can fill gaps. What matters is the teachers stance, the willingness to treat all languages as legitimate, and the ability to design tasks that make use of student expertise (Prada and Turnbull, 2021). Nonetheless, this raises important questions for #teacher_education. Programs that prepare teachers for #linguistically_diverse settings need to include explicit training in translanguaging pedagogy, in language awareness, and in strategies for working with languages the teacher does not speak. Recent research on teacher preparation in the United States, the United Kingdom, and the Netherlands suggests that this training remains uneven and often superficial. 7.2 Balancing Home Language and Dominant Language Development A second concern raised by critics is that translanguaging may inadvertently slow down the development of the dominant language, which students need in order to succeed in wider society. This is a fair concern, and it deserves a careful answer. The empirical evidence, taken as a whole, does not support the worry. Studies over the last decade consistently show that structured translanguaging supports rather than hinders dominant language development. However, this is true only when translanguaging is planned and purposeful. If home language use becomes an escape route that students use to avoid engaging with the target language entirely, gains in the target language may stall. The solution is not to ban home languages but to design activities where all languages are working together toward specific learning goals (MacSwan, 2022). There is also a question of equity within multilingualism. In many settings, students are pressured to become proficient in a global language, often English, even at the cost of losing their home language. Translanguaging can be part of a strategy to sustain home languages across generations, which matters for individual well being, family cohesion, and cultural continuity. 7.3 Risk of Superficial or Symbolic Implementation A third challenge is that translanguaging can become a slogan rather than a practice. Schools may announce that they value multilingualism while doing little to change instruction. A translation of the welcome sign into several languages is not translanguaging pedagogy. Neither is an occasional heritage language celebration. Real translanguaging requires structural change in how lessons are planned, how discussion is managed, how assessment is conducted, and how student work is valued (Poza, 2021). Researchers have documented cases where the term translanguaging was adopted in policy documents without corresponding changes in classroom practice. Teachers were told to be inclusive but were not given the training, materials, or time needed to implement the pedagogy. Predictably, outcomes did not improve, and translanguaging was blamed for a failure that was really a failure of implementation. 7.4 Assessment and Accountability Systems The largest structural challenge is assessment. As long as high stakes tests are administered in a single language, students who need translanguaging support are penalized twice, once by the difficulty of the content in a second language and again by the expectation that they demonstrate mastery in that language. Some jurisdictions have begun to experiment with #multilingual_assessment, offering bilingual versions of items or allowing home language responses on portions of tests. These experiments are important and should be studied and scaled where they succeed. Absent such changes, translanguaging teachers face a practical dilemma. If they invest heavily in home language use, will their students perform well enough on English only tests to advance? The evidence suggests yes, provided the pedagogy is well designed. But the pressure of accountability testing is real, and it shapes teacher behavior even when the underlying pedagogy would benefit from more flexibility. 7.5 The Question of Named Languages A more theoretical critique of translanguaging comes from scholars who worry that the framework may erase the importance of named languages themselves. If everything is one repertoire, do we lose the ability to talk about Igbo, Tamil, or Quechua as distinct languages with their own histories, communities, and rights? Some critics have argued that a strong version of translanguaging risks flattening the political struggles of language communities, particularly Indigenous communities, whose fight is often precisely to have their languages recognized as separate and legitimate (MacSwan, 2022). Most translanguaging theorists reject this critique as based on a misreading. Translanguaging does not deny the reality of named languages as social and political categories. It only questions their cognitive reality as separate systems in the bilingual mind. But the critique points to a real risk in practice. Translanguaging pedagogy should never be used as an excuse to neglect specific instruction in minority or Indigenous languages. In many contexts, students need both, translanguaging as a general pedagogical stance and dedicated instruction in each of their languages as distinct systems. 8. Implications and Conclusion 8.1 Implications for Teachers For classroom teachers, the implications of this review are practical. First, treat every students full linguistic repertoire as an asset. Even if you do not share their languages, communicate through your planning and your talk that these languages belong in learning. Second, design tasks that separate thinking from encoding, giving students space to reason in any language before producing final work in the target language. Third, make cross linguistic connections explicit, both to build metalinguistic awareness and to model the intellectual work of comparison. Fourth, document what you do. Records of student work and progress are your best defense against pressure to return to monolingual practice. 8.2 Implications for School Leaders School and district leaders shape the conditions in which teachers work. They should audit their own policies for hidden monolingual assumptions, including in hiring, communication with families, and student placement. They should invest in ongoing #professional_development on translanguaging pedagogy, not as a one time workshop but as sustained learning. They should build partnerships with families and communities that treat home languages as intellectual capital. And they should protect teachers who take pedagogical risks, providing air cover when parents or policymakers raise concerns. 8.3 Implications for Teacher Educators Teacher education programs must move beyond a token session on English learners toward a fully integrated approach in which #multilingual_pedagogy is treated as core, not peripheral. This means changing course syllabi, reading lists, and field placements. It also means recruiting and supporting a more #linguistically_diverse teacher workforce, one that reflects the students in schools rather than remaining a monolingual majority. 8.4 Implications for Policymakers For policymakers, the message from recent research is clear. Rigid monolingual policies are not producing the outcomes they promise. Students continue to underperform, gaps persist, and the linguistic wealth of communities is wasted. Policymakers should revise language of instruction policies to allow flexibility, invest in bilingual and multilingual programs where communities want them, develop assessment systems that recognize multilingual competence, and treat home language proficiency as an educational outcome worth pursuing rather than a distraction from real learning. 8.5 Implications for Researchers Finally, researchers have several open tasks. There is a need for more studies with rigorous outcome measures, comparing translanguaging and monolingual pedagogies on matched content. There is a need for longitudinal work that tracks students over years rather than a single semester. There is a need for research in underrepresented contexts, including African and Indigenous language settings, where translanguaging may look different from its usual portrayal in North American and European research. And there is a need for careful attention to the conditions under which translanguaging succeeds or fails, so that the pedagogy does not become a magic word that is applied without understanding. 8.6 Directions for Future Inquiry Despite the encouraging picture painted in this review, important questions remain open. The relationship between translanguaging and #digital_learning environments, for example, deserves closer attention. Online and hybrid instruction has expanded rapidly, and digital platforms create new opportunities for translanguaging through translation tools, multilingual chat, and multimodal composition, but they also risk automating away the human dimensions of language learning. Careful research is needed on how digital tools can support rather than replace pedagogical judgment. A second open area concerns very early childhood settings, where the child is still developing foundational language and literacy in more than one language at once. Some research suggests that early childhood translanguaging supports vocabulary breadth across languages, but longitudinal data are still limited. A third area is the intersection of translanguaging with subject specific literacies at the secondary and tertiary level. Different disciplines have different linguistic demands, and translanguaging strategies that work well for narrative writing may need adaptation for scientific argument, historical analysis, or mathematical proof. Comparative work across disciplines could help teacher educators develop more differentiated guidance. A fourth area concerns assessment innovation. Building fair, valid, and practical multilingual assessment instruments is a substantial technical challenge, but the alternative, continuing to test only in the dominant language, systematically underestimates what bilingual students know and can do. Educational measurement researchers have much to contribute here. Finally, more work is needed in contexts outside the well studied European and North American cases. Translanguaging in African, South Asian, Southeast Asian, Latin American, and Middle Eastern classrooms may look and function differently, and the field will be richer when these voices are more fully represented in the international research conversation. 8.7 A Note for Students and Early Career Educators For readers who are just beginning their journey as teachers, researchers, or graduate students, a final reflection may be useful. Translanguaging can feel intimidating when read at a theoretical level, but at the everyday level it is remarkably practical. It starts with listening. Notice how your students actually talk when they think no one is watching. Notice what languages they whisper in, what languages they scribble notes in, what languages they laugh in. Those are the languages of their thinking, and thinking is what schooling is supposed to develop. From that noticing, small changes follow. Allow a two minute pair discussion in any language before a class discussion. Accept a first draft in a home language and coach the student through revision into the target language. Post a word in more than one script on the board. Ask students to teach you a term you do not know. These small acts change the atmosphere of a classroom in ways that formal policy cannot always reach, and they build the professional judgment that will make you a stronger educator in whatever system you find yourself. Teaching in a monolingual system while believing in multilingualism is not easy. It requires patience, documentation, and community. But every classroom that opens up in this way is a small correction of a long historical mistake, the mistake of treating the ordinary multilingual life of most of humanity as a problem to be solved rather than a foundation to be built on. 8.8 Conclusion The world is multilingual and always has been. Schools that treat it otherwise are working against evidence, against equity, and against the actual minds of the children they serve. Translanguaging pedagogy does not solve every problem in education, but it does something important. It insists that the languages children bring with them to school are not obstacles to be managed but resources to be mobilized. It gives teachers concrete moves for turning that insistence into daily practice. And it aligns education more honestly with the world outside the classroom door, where communication happens across languages, cultures, and communities in ways that no single language can capture. Encouraging bilingual students to use their full linguistic repertoire, especially for the hardest #cognitive_tasks, does not weaken their command of the dominant language. It strengthens it, because thinking in more than one language produces sharper thinking overall. Students who translanguage well are not less prepared for a monolingual test, they are more prepared, because they understand ideas more deeply and can express them more precisely. The advantage is real, measurable, and available to any teacher willing to reorganize the linguistic architecture of the classroom. Institutional systems will not change overnight. But every classroom that opens its doors to the full repertoire of its students is a small act of change, and enough small acts, sustained over time, can shift what is possible. The multilingual student sitting in a monolingual system is not a problem to be fixed. She is a scholar who already knows things her teacher does not. The task of translanguaging pedagogy is to make sure that what she knows can enter the conversation, be built on, and become part of what the classroom, and eventually the system, understands. References Cenoz, J. and Gorter, D. (2021). Pedagogical translanguaging. Cambridge University Press. https://doi.org/10.1017/9781009029384 Cenoz, J. and Gorter, D. (2022). Pedagogical translanguaging and its application to language classes. RELC Journal, 53(2), 342 to 354. https://doi.org/10.1177/00336882221082751 Duarte, J. (2020). Translanguaging in the context of mainstream multilingual education. International Journal of Multilingualism, 17(2), 232 to 247. https://doi.org/10.1080/14790718.2018.1512607 Garcia, O. and Kleifgen, J. A. (2020). Translanguaging and literacies. Reading Research Quarterly, 55(4), 553 to 571. https://doi.org/10.1002/rrq.286 Li, W. (2022). Translanguaging as a political stance: Implications for English language education. ELT Journal, 76(2), 172 to 182. https://doi.org/10.1093/elt/ccab083 MacSwan, J. (Ed.). (2022). Multilingual perspectives on translanguaging. Multilingual Matters. https://doi.org/10.21832/9781800415690 Poza, L. E. (2021). Adding flesh to the bones: Translanguaging, embodiment, and critical consciousness. Multilingua, 40(4), 493 to 514. https://doi.org/10.1515/multi-2019-0143 Prada, J. and Turnbull, B. (2021). The role of translanguaging in the multilingual turn: Driving philosophical and conceptual renewal in language education. EuroAmerican Journal of Applied Linguistics and Languages, 8(1), 8 to 23. https://doi.org/10.21283/2376905X.13.204 Prilutskaya, M. (2021). Examining pedagogical translanguaging: A systematic review of the literature. Languages, 6(4), 180. https://doi.org/10.3390/languages6040180 Rajendram, S. (2021). Translanguaging as an agentive pedagogy for multilingual learners: Affordances and constraints. International Journal of Multilingualism, 20(2), 595 to 622. https://doi.org/10.1080/14790718.2021.1898619 Rajendram, S. (2022). The affordances of translanguaging as a pedagogical resource for multilingual English language classrooms. TESOL Quarterly, 56(2), 623 to 651. https://doi.org/10.1002/tesq.3131 Sah, P. K. and Li, G. (2022). Translanguaging or unequal languaging? Unfolding the plurilingual discourse of English medium instruction policy in Nepal's public schools. International Journal of Bilingual Education and Bilingualism, 25(6), 2075 to 2094. https://doi.org/10.1080/13670050.2020.1849011 Tai, K. W. H. and Li, W. (2021). Constructing playful talk through translanguaging in English medium instruction mathematics classrooms. Applied Linguistics, 42(4), 607 to 640. https://doi.org/10.1093/applin/amaa043 Ticheloven, A., Blom, E., Leseman, P. and McMonagle, S. (2021). Translanguaging challenges in multilingual classrooms: Scholar, teacher and student perspectives. International Journal of Multilingualism, 18(3), 491 to 514. https://doi.org/10.1080/14790718.2019.1686002 Vogel, S. and Garcia, O. (2020). Translanguaging. In G. Noblit (Ed.), Oxford Research Encyclopedia of Education. Oxford University Press. https://doi.org/10.1093/acrefore/9780190264093.013.181 Wang, D. (2020). Studying Chinese language in higher education: The translanguaging reality through learners eyes. System, 95, 102394. https://doi.org/10.1016/j.system.2020.102394 #translanguaging #multilingual_education #bilingual_learners #pedagogical_translanguaging #linguistic_repertoire #English_medium_instruction #language_policy #monolingual_mandates #classroom_diversity #linguistic_justice #metalinguistic_awareness #multilingual_pedagogy #emergent_bilinguals #language_and_learning #equity_in_education

High-stakes assessment has become the default gatekeeper of academic progress, credentialing, and opportunity across schools and universities worldwide. Yet the standard timed, heavily formatted examination rests on assumptions about attention, working memory, reading speed, and executive function that do not reflect the full range of learners in modern classrooms. This article examines how conventional testing systematically disadvantages students with #ADHD and #autistic processing styles, and it reviews the growing evidence base supporting Universal Design for Learning as a credible alternative for measuring #mastery. Drawing on recent scholarship in #inclusive_assessment, #neurodiversity studies, and higher education pedagogy, the paper argues that fairness and validity are not competing values but the same value seen from two angles. When an assessment penalizes a student for slow processing rather than weak knowledge, it produces a score that is both unjust and psychometrically noisy. The article proposes a redesigned protocol grounded in flexible timing, multiple response formats, transparent rubrics, and process-based evidence of learning. It closes by discussing the ethical, institutional, and policy conditions under which such reforms can move from pilot projects to mainstream practice, and by identifying the research gaps that still separate promising practice from settled science. Keywords Neurodiversity; Universal Design for Learning; high-stakes testing; ADHD; autism; assessment reform; inclusive pedagogy; executive function; test validity; higher education. 1. Introduction For most of the twentieth century, the timed written examination was treated as a neutral instrument. A student who knew the material would show it inside the allotted minutes, and a student who did not would fail. This view survives in policy documents, in admissions handbooks, and in the daily practice of thousands of classrooms. It also survives, quietly, in the design of the newer computer-based tests that increasingly replace paper booklets. Speed, silence, and a narrow band of acceptable formats remain the shared language of #high_stakes_testing. The trouble is that this language was never truly neutral. It was written for a specific kind of #test_taker: one who reads quickly, writes fluently under pressure, holds several instructions in mind at once, filters out background noise without effort, and recovers from small mistakes without spiraling. Students who match this profile perform as expected. Students whose #cognitive_profile differs, including many students with #attention_deficit_hyperactivity_disorder and many #autistic students, produce scores that under-represent what they actually know (Pellicano & den Houting, 2022; Dwyer, 2022). The score gap in these cases is not a knowledge gap. It is a design gap. This distinction matters because #assessment scores drive consequential decisions. They determine progression, scholarship eligibility, professional licensure, and, in many systems, the tracking of children into academic pathways from an early age. When a testing format systematically mis-measures a subgroup of learners, the harm is compounded across years and across institutions. A student who is repeatedly told, through low grades, that they cannot succeed academically will internalize that message even when the underlying cause is a mismatch between their #processing_style and the test format (Hamilton & Petty, 2023). The #neurodiversity paradigm offers a different starting point. It treats variation in cognition, attention, and sensory processing as an expected feature of the human population rather than a set of deficits to be corrected (Dwyer, 2022; Pellicano & den Houting, 2022). From this vantage point, the question is not how to help #neurodivergent students survive assessments designed for someone else. The question is how to build assessments that give every student a fair chance to demonstrate what they have learned. This reframing is not a rhetorical flourish. It has direct implications for test design, scoring, and interpretation. #Universal_Design_for_Learning, developed originally in the special education and cognitive neuroscience literature and refined over three decades of classroom application, provides a practical framework for that redesign (CAST, 2024). UDL asks educators to plan for variability from the start, to offer multiple means of engagement, representation, and expression, and to treat #accessibility as a property of the assessment rather than an accommodation bolted on afterward. Recent empirical work suggests that UDL-aligned assessments can preserve rigor while reducing the performance penalty associated with neurodivergent profiles (Nieminen, 2022; Tai et al., 2023). This article makes three contributions. First, it synthesizes recent scholarship on how conventional #timed_testing disadvantages ADHD and autistic learners, distinguishing effects on #construct_validity from effects on well-being. Second, it maps those findings onto specific UDL principles and shows where the evidence for each principle is strong, promising, or still thin. Third, it proposes a redesigned assessment protocol that institutions can pilot without abandoning the accountability functions that #standardized_testing is asked to perform. The aim is not to romanticize alternative assessment nor to dismiss the legitimate uses of timed tests. The aim is to make assessment do the job it claims to do: measure #learning, accurately, for the students actually in the room. 2. Literature Review 2.1 The Neurodiversity Paradigm and Its Educational Consequences The term #neurodiversity was introduced in the late 1990s by the autistic sociologist Judy Singer and has since moved from advocacy into mainstream academic discourse (Chapman & Botha, 2023). At its core, the paradigm makes two related claims. The descriptive claim is that human cognition varies along multiple dimensions, including attention regulation, sensory sensitivity, social processing, memory, and executive function, and that this variation is a stable feature of the population rather than a temporary disorder in a small minority. The normative claim is that this variation should be accommodated by social and institutional design, in the same way that architectural design accommodates variation in physical mobility (Pellicano & den Houting, 2022). In education, the paradigm shifts the burden of adjustment. Under the older #medical_model, an ADHD or autistic student was expected to receive treatment, learn compensatory strategies, and eventually approximate the behavior of a #neurotypical peer, at which point the environment could remain unchanged. Under the neurodiversity paradigm, the environment itself is scrutinized for features that create avoidable friction (Dwyer, 2022). A classroom with harsh fluorescent lighting, an assessment protocol that penalizes slow handwriting, or a lecture format that requires sustained silent attention for ninety minutes are not neutral defaults. They are choices, and they can be chosen differently. Recent scholarship has emphasized that the paradigm is not a claim that #neurodivergent people experience no difficulty or need no support (Dwyer, 2022; Chapman & Botha, 2023). ADHD is associated with genuine challenges in sustained attention, time perception, and working memory, and autistic experience often includes sensory overload, difficulty with abrupt transitions, and exhaustion from social masking. The paradigm's claim is narrower: that these challenges are shaped by environmental fit as much as by individual biology, and that when the environment can be redesigned, it usually should be. Assessment is one of the environments most open to redesign, because its features are set by institutional choice rather than by physical constraint. 2.2 ADHD, Autism, and the Cognitive Demands of Timed Testing To see why conventional testing produces distorted scores for many #neurodivergent students, it helps to look closely at what a timed exam actually requires. A standard two-hour written examination demands sustained #attention across the full window, working memory to hold instructions and partial answers in mind, executive function to plan and pace responses across sections, fluent reading to process items at speed, and self-regulation to manage anxiety and to recover from difficult items without losing time on later ones (Sedgwick-Müller et al., 2022). Each of these demands is a separate cognitive task, and each is a task on which many ADHD and autistic students diverge from the neurotypical average. For students with ADHD, the most consistent findings concern #time_perception and pacing. Research on adult and adolescent ADHD documents difficulties in estimating elapsed time, in resisting distraction across long tasks, and in maintaining consistent output when a task is not intrinsically engaging (Sedgwick-Müller et al., 2022; Weyandt et al., 2023). In an exam context, these difficulties compound. A student may spend too long on early items, notice too late that the clock is against them, and rush the remainder in a state of heightened stress. The resulting score reflects an interaction between the student's knowledge, their pacing, and the strictness of the time limit. Removing the pacing variable, either by extending time meaningfully or by allowing scheduled breaks, tends to raise scores for ADHD students without raising them for #neurotypical peers, which is the expected pattern when a design change targets a source of #construct_irrelevant_variance (Lovett & Lewandowski, 2023). For autistic students, the picture is more heterogeneous. Some autistic students perform strongly on timed tests, particularly when the format is highly structured and the content aligns with intense interest areas. Others struggle for reasons that have less to do with time and more to do with #sensory processing, ambiguous instructions, and the social demands of the testing environment (Clouder et al., 2020; Hamilton & Petty, 2023). Fluorescent lighting, other students' movements, the pressure of an invigilator's presence, and small variations in question phrasing that require inference about the examiner's intent can all consume cognitive resources that would otherwise go to the question itself. A test that asks "briefly discuss" or "give a few examples" without specifying how many or how long is not a neutral instruction. It is a request for #social_inference that some autistic students find genuinely opaque. There is also a mental health dimension that recent studies have brought into clearer focus. Both ADHD and autistic students report high rates of test anxiety, and the anxiety is not simply a byproduct of poorer expected performance (Hamilton & Petty, 2023; Scheef et al., 2023). It reflects accumulated experience of being penalized for cognitive features they cannot easily change, and it interacts with the cognitive load of the test itself. A student who has failed timed exams before enters the next one with an elevated baseline of physiological stress, which further degrades working memory and attention (Weyandt et al., 2023). This is a feedback loop that assessment design can either amplify or dampen. 2.3 Construct Validity and the Fairness Argument The strongest argument for redesign is not a compassion argument. It is a #validity argument. Every assessment claims to measure a specific construct, whether that is mastery of a body of content, ability to reason in a discipline, or readiness for the next stage of study. The score is a #proxy for the construct, and the proxy is only useful to the extent that it isolates the construct from irrelevant sources of variance (Nieminen, 2022). A test that measures both content mastery and reading speed will produce different scores for two students with identical content mastery but different reading speeds. In psychometric terms, reading speed has become part of the construct, whether or not the test designers intended it to be. This is the concept of #construct_irrelevant_variance, and it is where the neurodiversity discussion meets classical test theory. When a timed test produces a lower score for an ADHD student who knows the material as well as a peer, the difference between the two scores is not measuring knowledge. It is measuring the interaction between the student's pacing and the test's clock. If a scholarship committee treats the score as a measure of knowledge, they are drawing a conclusion the test cannot support (Nieminen, 2022; Tai et al., 2023). The problem is not that the test is unfair in a vague moral sense. The problem is that it does not measure what it claims to measure. This framing has practical value because it reframes accommodation as validity work rather than as a favor. Extended time, quiet rooms, or the option to type rather than handwrite are often described as accommodations that #level_the_playing_field. Under the validity framing, they are corrections to a measurement instrument that would otherwise produce misleading scores (Lovett & Lewandowski, 2023). This framing also clarifies why accommodations should not necessarily be limited to students with formal diagnoses. If a design feature produces construct-irrelevant variance for a broad group of learners, changing that feature improves the measurement for everyone. 2.4 Universal Design for Learning as an Assessment Framework #UDL emerged from work at the Center for Applied Special Technology in the 1990s and has been refined into a set of guidelines that address three broad networks of learning: engagement, representation, and action and expression (CAST, 2024). Applied to assessment, the framework asks designers to consider variability at each stage. How will students be engaged with the task? How will information be presented to them? And how will they be allowed to demonstrate what they know? The engagement dimension has received less attention in the assessment literature than the other two, but it is not trivial. A task that is genuinely engaging reduces the executive burden required to sustain attention, which particularly benefits ADHD students (Lambert et al., 2021). This does not mean assessments should be entertainment. It means that when the content of a task connects to a purpose the student can see, the student is drawing on #intrinsic_motivation rather than fighting against boredom. Portfolio assessments, project-based tasks, and authentic problem-solving scenarios tend to score higher on this dimension than sit-down examinations, although the trade-offs in comparability and administrative cost are real. The representation dimension addresses how test content reaches the student. A single format, typically dense prose read silently under time pressure, disadvantages students whose reading is slower, who process better with visual scaffolding, or whose comprehension improves when they can hear the text as well as read it (Griful-Freixenet et al., 2020). Offering the same content in multiple formats, or allowing text-to-speech tools by default, reduces this bottleneck. The evidence here is stronger for classroom assessment than for high-stakes standardized testing, where equivalence across formats is harder to guarantee (Nieminen & Pesonen, 2022). The #action_and_expression dimension is where UDL and traditional assessment collide most directly. Traditional testing typically permits one output mode: handwritten prose or a set of multiple-choice selections. UDL asks whether the same knowledge could be demonstrated through a spoken response, a diagram, a structured project, or a combination of formats. For autistic students who find open-ended written prose difficult but who can produce sophisticated structured responses, this flexibility can be transformative (Hamilton & Petty, 2023; Nieminen, 2022). The design challenge is to ensure that different output modes measure the same construct at comparable difficulty, which is achievable but requires deliberate rubric work. 2.5 Evidence on Alternative Assessment Formats The empirical literature on alternative assessment has grown substantially in the last five years, driven partly by pandemic-era shifts to remote learning and partly by broader interest in #inclusive_assessment. Several patterns are now reasonably well-established (Tai et al., 2023; Nieminen, 2022; O'Neill & Padden, 2022). First, extended time is the most-studied accommodation and the one with the strongest evidence base. Studies comparing standard time to extended time consistently find that students with ADHD and specific learning differences benefit more from additional time than neurotypical peers do, which is the pattern expected if the extension is correcting for construct-irrelevant variance rather than inflating scores generally (Lovett & Lewandowski, 2023). The evidence does not support the once-common concern that extended time provides an unearned advantage. It suggests that the standard time limit was too tight for a subgroup of students all along. Second, format flexibility appears to help autistic students in particular, although the studies are smaller and more varied. Allowing students to choose between essay and structured-response formats, or between oral and written presentation, tends to raise reported measures of #self_efficacy and to reduce test anxiety without depressing measured achievement (Hamilton & Petty, 2023). The main methodological challenge in this literature is that self-selection into formats can confound comparisons, and more randomized designs are needed. Third, #process_based_assessment, in which students demonstrate mastery through a body of work developed over time rather than through a single high-stakes event, shows promise for #neurodivergent learners but faces significant institutional barriers (Nieminen, 2022; Boud et al., 2023). Portfolios, capstone projects, and cumulative competency demonstrations reduce the acute cognitive load of a single test day and give students the chance to show growth over time. The trade-off is that portfolio assessment is harder to standardize across large populations and more labor-intensive to score reliably. Fourth, the evidence on #open_book and #take_home examinations is genuinely mixed. Some studies find that removing the memory retrieval component and allowing reference to materials during the exam benefits students with working memory difficulties without reducing measurement of higher-order reasoning (Johanns et al., 2022). Others find that the reduced time pressure of take-home formats can paradoxically increase anxiety by extending the exam window across days and by inviting perfectionism. The design details, particularly the specificity of prompts and the clarity of scope, appear to matter more than the format label. 2.6 What the Literature Does Not Yet Settle Several important questions remain open. The first is the extent to which UDL-aligned assessments can be scaled to genuinely high-stakes contexts such as university admissions, medical licensure, and legal bar examinations. Most published studies are conducted in classroom or program contexts where the assessment is consequential but not gatekeeping in the strongest sense (Tai et al., 2023). Scaling to gatekeeping contexts raises legitimate concerns about #comparability, security, and public trust that pilot studies have not fully answered. The second open question concerns the interaction of design changes with other identity variables. Students who are neurodivergent and also from #low_income backgrounds, or who are neurodivergent and second-language learners, or who are neurodivergent and physically disabled, may experience assessments in ways that a single-axis analysis misses (Griful-Freixenet et al., 2020). The literature has not yet caught up with this #intersectional complexity. The third open question is the durability of design changes over time. A UDL-aligned assessment that works well in a small program with committed faculty may lose fidelity when rolled out across a large institution with variable buy-in. The implementation science of assessment reform is thinner than the design literature, and this gap has practical consequences for policy (O'Neill & Padden, 2022). 3. Methodology This article is a #narrative_review synthesis rather than a systematic review, and the methodological choice is deliberate. Systematic reviews are strongest when a well-defined empirical question can be answered by pooling comparable studies. The question addressed here is broader: how does the design of high-stakes assessment interact with neurodivergent cognitive profiles, and what does this imply for redesign? Answering this question requires drawing on empirical studies, theoretical accounts of neurodiversity, psychometric analyses of construct validity, and applied work on UDL, none of which sit comfortably in a single meta-analytic framework. The literature was identified through targeted searches of major education, psychology, and disability studies databases, focusing on work published between 2020 and 2026. Search terms combined variants of #neurodiversity, ADHD, autism, and high-stakes assessment with terms for UDL, inclusive assessment, and test accommodation. Additional sources were identified through backward citation tracing from key review articles and through examination of major journals in the field. Studies were included when they addressed either the effect of assessment design on neurodivergent learners or the conceptual framework of inclusive assessment. Grey literature was consulted for context but not treated as evidence for empirical claims. The synthesis prioritized peer-reviewed empirical work for causal claims and used theoretical and conceptual literature to frame the discussion. Two limitations of this approach should be acknowledged. First, narrative reviews are more susceptible to selection bias than systematic reviews, and readers should treat the specific studies cited as illustrative rather than exhaustive. Second, the empirical literature on neurodivergent assessment is still developing, and some claims that appear well-supported in current work may not survive replication. 4. How Standard Assessment Design Penalizes Neurodivergent Learners 4.1 The Time Variable The clock is the single most consequential feature of a conventional examination. Yet the choice of any specific time limit is rarely justified in a technical sense. Time limits are typically inherited from prior versions of a test, adjusted by rule of thumb, or set by administrative convenience rather than by systematic study of how long students actually need to demonstrate mastery (Lovett & Lewandowski, 2023). This is a striking gap in a domain that otherwise takes psychometric rigor seriously. For ADHD students, the time variable produces two related problems. The first is pacing. Difficulties in time perception mean that ADHD students often misjudge how long has passed or how long a specific item is taking, which produces the familiar pattern of strong performance on early items and rushed or incomplete performance on later items (Sedgwick-Müller et al., 2022). The second is #executive_function load. A tight time limit forces the student to run a background process that monitors elapsed time, evaluates progress, and adjusts strategy accordingly. This monitoring competes for the same executive resources that the test content requires, and the competition is more costly for students whose executive function is already stretched. The design implications are clearer than the policy implications. From a design perspective, the answer is either to extend the time meaningfully, to remove the time limit for many assessment types, or to break the assessment into shorter segments with scheduled breaks. From a policy perspective, extending time on a high-stakes standardized test raises questions about score comparability across administrations. These questions are addressable, but they require deliberate psychometric work that most systems have not yet undertaken. 4.2 Format Rigidity A second consequential design choice is the rigidity of the required response format. A test that permits only handwritten prose answers is measuring at least three things: content mastery, prose production under pressure, and #handwriting fluency. For most students the second and third contribute little variance, but for students with dysgraphia, hand tremor, or autistic profiles that make extended prose production difficult, they can dominate the score (Clouder et al., 2020). The problem is not that any specific format is wrong. The problem is that a single required format converts every difference in students' expressive strengths into a difference in measured achievement. A student who can explain a complex proof orally in five minutes but who struggles to write it out in twenty is not less capable in mathematics. They are less capable at one specific expressive channel. When the test requires that channel and only that channel, the score confounds mathematical reasoning with prose fluency. Format flexibility does not require abandoning rigor. It requires deliberate work on #equivalent_rubrics that assess the same underlying construct across multiple response modes. This work is non-trivial, but it is doable, and several disciplines have developed multi-format assessment protocols that maintain reliable scoring (Nieminen, 2022; Boud et al., 2023). 4.3 The Sensory and Environmental Design of Testing Rooms The physical environment of testing is often treated as a background condition rather than a design choice. In practice it is a design choice, and one that particularly affects autistic students. Standard testing rooms tend to feature fluorescent overhead lighting, hard reflective surfaces, ambient noise from ventilation and other students, and the physical presence of invigilators moving through the space. Each of these features is a potential source of #sensory_load, and the load consumes cognitive resources that the test itself requires (Hamilton & Petty, 2023). The literature on autistic sensory processing has documented substantial individual variation, which means that a single alternative environment cannot serve all autistic test takers. But the pattern of design choices in standard rooms is not neutral either. It tends to maximize sensory input rather than minimize it, on the assumption that a neurotypical student can filter the input at low cognitive cost. For students whose filtering is more effortful, the standard room is not a level baseline. It is an active source of interference. Simple design changes can reduce this interference. Softer or task-focused lighting, sound-absorbing surfaces, permission to use noise-canceling headphones, smaller room sizes, and reduced invigilator movement all appear in the literature as low-cost modifications with meaningful effects (Clouder et al., 2020). None of these changes threaten the security or comparability of the assessment itself. 4.4 The Hidden Curriculum of Test Instructions Test instructions are the interface between the examiner's intent and the student's response. In principle they should be transparent. In practice they often carry a #hidden_curriculum of expectations that are legible to students familiar with the genre and opaque to students who are not. Phrases like "briefly explain," "give some examples," "discuss critically," or "in your own words" ask the student to infer what the examiner wants without stating it directly. Familiar test takers have learned that "briefly" typically means one paragraph, that "some examples" typically means three, and that "critically" typically means presenting more than one view. Students who read instructions literally, including many autistic students, may spend cognitive resources trying to determine what the examiner actually intends, or may produce responses that meet the instruction as written but not as intended (Hamilton & Petty, 2023; Nieminen, 2022). The remedy is instruction transparency. Specifying word counts, numbers of examples, and expected structural features reduces the load of inference and lets students focus on content. Some assessment designers worry that this level of specificity constrains creativity, but the evidence suggests it does not. Students who know what is expected can meet the expectation and then go beyond it, whereas students who do not know what is expected often produce responses that miss the target for reasons unrelated to knowledge. 4.5 The Cumulative Cost Any single design feature discussed above is a modest source of #construct_irrelevant_variance. The cumulative cost is larger than the sum, because the features interact. A student who is already fatigued by sensory load will find pacing harder. A student who is anxious about pacing will find ambiguous instructions harder to parse. A student who has parsed ambiguous instructions incorrectly will lose time correcting course, which compounds the pacing problem. The result is that #neurodivergent students often lose ground for reasons that no single feature would fully explain, and that accommodation of a single feature only partially addresses. This cumulative logic is the strongest argument for a whole-system redesign rather than a menu of individual accommodations. A UDL-aligned assessment protocol addresses the features together, and the benefits interact in the opposite direction. Reduced sensory load makes pacing easier, clear instructions reduce anxiety, format flexibility reduces the cost of any single expressive weakness, and the whole becomes measurably fairer without becoming easier (Nieminen, 2022; O'Neill & Padden, 2022). 5. Redesigning Assessment Protocols: A Proposed Framework The framework proposed here is not a finished protocol. It is a set of design principles that institutions can adapt to their context, along with practical illustrations of how each principle might be operationalized. The framework treats assessment as a system with four components: the task, the environment, the response, and the scoring. Redesign should address all four, and the strongest gains come when they are addressed together. 5.1 Task Design Task design is the starting point because it determines what the assessment claims to measure. A well-designed task isolates the target construct as far as possible from irrelevant demands. In practice this means several things. First, the task should be aligned to specified #learning_outcomes rather than to a generic content area. If the outcome is "explain the mechanism of action of a specific class of drugs," the task should ask exactly that, in language that closely tracks the outcome. If the outcome is "apply a framework to an unfamiliar case," the task should present an unfamiliar case. Vague prompts that require students to guess what content is expected introduce irrelevant variance from the start (Boud et al., 2023). Second, the task should specify its own structural expectations. Word counts, number of parts, expected format, and criteria for a strong response should be visible in the task itself, not inferred from convention. This transparency benefits all students and particularly benefits students who read literally or who lack familiarity with the specific assessment genre. Third, the task should minimize #reading_load that is not part of the target construct. If the assessment is measuring mathematical reasoning, dense prose framing of the problems introduces reading speed as an unintended factor. Simpler framing, visual scaffolding, or the option of audio access to the same content reduces this bottleneck without reducing the difficulty of the underlying task. Fourth, where feasible, the task should offer more than one response format, with equivalent scoring across formats. A student might choose between a written essay and a structured oral presentation, or between a traditional problem set and a design-and-defend project. This is where the design work is most demanding, because equivalence across formats requires rubrics that are format-neutral. The investment in developing these rubrics pays off in more accurate measurement across a wider student population (Nieminen, 2022; Tai et al., 2023). 5.2 Environmental Design Environmental design attends to the physical and temporal conditions under which the assessment takes place. Even when the task is well designed, a hostile environment can dominate the resulting score. The most important environmental variable is time. The framework here recommends that time limits be justified rather than assumed. A time limit is defensible when it is measuring a construct that includes speed. It is not defensible when it is measuring a construct that does not, in which case the limit is a source of construct-irrelevant variance. For most academic knowledge and reasoning tasks, the construct does not include speed, and the time limit should be either loose enough that pacing is not a significant factor for any student, or removed entirely in favor of a fixed-effort model (Lovett & Lewandowski, 2023). Sensory conditions are the second important environmental variable. The framework recommends that testing rooms be designed with sensory variability in mind. This includes moderated lighting, sound treatment, reduced invigilator movement, and permission for personal sensory tools such as noise-canceling headphones and #stim_toys where they do not interfere with security. These changes are inexpensive and reduce a documented source of load for autistic and sensory-sensitive students (Hamilton & Petty, 2023). The framework also recommends #scheduled_breaks for assessments longer than sixty minutes. A short break every hour, during which students can move, drink water, or use the bathroom, reduces the executive load of sustained attention without providing any content advantage. This is one of the most cost-effective changes an institution can make, and it benefits a broad range of students including but not limited to those with ADHD. 5.3 Response Design Response design addresses what students are permitted to do to demonstrate what they know. The framework's central recommendation is #response_flexibility within construct-preserving constraints. Concretely, this might look like the following. For a task assessing knowledge of a body of content, students might choose to respond in written prose, structured short-answer format, or annotated diagram, with rubrics that treat each format as an equivalent demonstration of content mastery. For a task assessing analytical reasoning, students might choose between written analysis and an oral defense with a marker, with the oral defense scored on the same reasoning criteria. For a task assessing #applied_skill, students might submit a project artifact with an accompanying reflection that explains the reasoning behind key decisions. The framework does not recommend that every assessment offer every possible format. That would be impractical and would dilute rubric quality. It recommends that the assessment offer at least two formats where feasible, and that the choice between formats be a genuine choice rather than a fallback that carries implicit stigma. Students should not have to disclose a diagnosis to choose an oral format over a written one. An additional response design consideration is the treatment of #metacognition. Assessments that include a brief structured reflection, in which students explain how they approached the task and what they found difficult, generate additional information about the student's reasoning process and reduce the score's dependence on a single point-in-time performance (Boud et al., 2023). Metacognitive components should be short, structured, and separately scored so they do not become an additional writing task in disguise. 5.4 Scoring Design Scoring design closes the loop. Even the best task, environment, and response design can be undermined by scoring that reintroduces bias. The framework recommends four scoring principles. First, #rubrics should be #transparent and shared with students in advance. Transparency is not a threat to rigor. It is a condition of fair measurement, because it allows students to direct their effort toward the criteria that will actually be scored. Second, rubrics should describe #performance in terms of demonstrated understanding rather than surface features of the response. A rubric that rewards "clear, well-organized prose" is measuring prose fluency alongside understanding, and disadvantages students whose understanding is strong but whose prose fluency is not. A rubric that describes "accurate identification of key concepts, correct application to the case, and coherent reasoning across steps" can be applied fairly across multiple response formats. Third, scoring should be #calibrated across markers, particularly for assessments that offer format flexibility. Multi-marker moderation is essential, because format equivalence depends on markers applying the rubric consistently across formats. Institutions that have piloted multi-format assessment have found that marker training is the single largest implementation cost, and that skipping it produces the very inconsistency the framework is designed to prevent (Tai et al., 2023). Fourth, #high_stakes_decisions should not rest on a single assessment point. Where consequential decisions must be made, the framework recommends #triangulation across multiple assessment types over time. A student's academic profile is more accurately captured by a combination of a timed test, a portfolio, a project, and a reflective component than by any one of these alone. Triangulation both improves measurement and reduces the impact of a single bad test day, which particularly benefits students whose performance varies more from day to day. 5.5 A Worked Example To make the framework concrete, consider a redesigned undergraduate assessment for an introductory biology course. The original assessment was a two-hour written examination with fifty multiple-choice items and two short essay questions, administered in a large exam hall. Under the framework, the redesigned assessment might comprise three components. The first is a #knowledge_check module completed online during a two-week window, consisting of item types drawn from the original test but with untimed response (subject to the two-week window) and audio access to items. The second is a #case_analysis module in which students receive a novel case and produce a structured response, with the choice of written analysis, annotated diagram with narration, or recorded oral analysis. The third is a #reflective_component in which students briefly explain how they approached the case and what they would do differently on a similar future task. The overall assessment is scored against the same learning outcomes as the original test. The rubric is shared in advance. Markers are trained on cross-format equivalence. Students who prefer a single sit-down format can complete all three components in a proctored session if they choose. The redesigned assessment maintains rigor, addresses the specific sources of #construct_irrelevant_variance discussed above, and produces a richer picture of student mastery than the original. Pilot implementations of this kind of redesign have reported comparable or higher overall pass rates, reduced gaps between neurodivergent and neurotypical students, and improved student reports of well-being, with acceptable marker workload after initial rubric development (Nieminen, 2022; O'Neill & Padden, 2022). These findings are promising but should be replicated across a wider range of disciplines and institutions before they support strong causal claims. 6. Implementation Challenges and How to Address Them Design frameworks are easier to write than to implement. The literature on assessment reform is candid that many well-designed pilots fail to scale, and that the reasons for failure are usually institutional rather than technical (O'Neill & Padden, 2022; Boud et al., 2023). Four challenges are especially common. 6.1 Faculty Capacity and Buy-In Redesigning assessments requires faculty time and expertise. Faculty members are typically evaluated on research output more than on teaching, and the incentive structure often does not reward the investment that redesign requires. Even faculty who are sympathetic to the goals of #inclusive_assessment may feel unable to commit the necessary time to rubric development, format expansion, and marker training. Addressing this challenge requires institutional signals that assessment redesign is valued. This can take several forms, including recognition in promotion criteria, small grants for redesign work, and shared institutional infrastructure for rubric development and marker training. Perhaps most importantly, it requires visible institutional protection of faculty who pilot new approaches, so that early implementation difficulties do not translate into negative teaching evaluations that penalize the very faculty doing the reform. 6.2 Concerns About Rigor and Comparability A recurring concern is that flexible assessments dilute rigor or make comparisons across students unfair. The concern deserves serious engagement rather than dismissal. In some poorly designed implementations, format flexibility has been introduced without corresponding investment in rubric equivalence, and the resulting scores have been genuinely non-comparable. This is a real failure mode, and one that critics of UDL are correct to point out. The response is not to abandon flexibility but to invest in the technical work that makes flexibility rigorous. Cross-format rubrics, multi-marker moderation, and periodic psychometric analysis of format effects are all established tools. When they are used, format flexibility does not compromise comparability. When they are skipped, it does. The rigor question is a real one, but it is a question about implementation quality rather than about the framework itself (Nieminen, 2022; Tai et al., 2023). 6.3 High-Stakes Standardized Contexts Adapting the framework to standardized admissions or licensure testing is genuinely harder than adapting it to classroom assessment. Standardized tests are administered to large populations, scored on tight timelines, and used to make decisions with significant legal and economic consequences. Change is slower in these contexts for reasons that are not purely inertia. Several avenues nonetheless appear promising. The most conservative involves generalizing existing accommodations. If extended time and quiet rooms benefit a broad group of students without disadvantaging any subgroup, expanding these features to more students by default is a low-risk change that could be piloted at scale. A middle path involves offering test takers a choice among a small set of validated formats, with careful psychometric equating across formats. The most ambitious path involves moving toward #modular_assessment in which the test is delivered in shorter sessions across a window, with more flexible pacing within each session. None of these paths is straightforward, but each has been discussed in the recent testing policy literature and each has partial precedents (Lovett & Lewandowski, 2023). The main obstacle is institutional risk aversion, which is understandable given the consequences of a testing failure, but which also produces a status quo bias that keeps in place features whose validity is more assumed than demonstrated. 6.4 The Diagnosis Question Under most current systems, formal accommodations require formal diagnoses. This produces two problems. First, diagnosis is expensive, time-consuming, and unevenly available, which means that students from #low_income backgrounds or in under-resourced regions may be neurodivergent but undiagnosed and therefore ineligible for support. Second, tying support to diagnosis reinforces the idea that the student is the problem, which contradicts the whole logic of the neurodiversity paradigm. UDL-aligned assessment addresses this by making many design features available to all students by default, without requiring disclosure. Extended time, format flexibility, and sensory-considered environments do not need to be individually justified if they are built into the standard protocol. This reduces the burden on students, reduces the administrative cost of case-by-case accommodation, and avoids the disclosure dilemma that many neurodivergent students face (Chapman & Botha, 2023; Dwyer, 2022). Diagnosis-based accommodation can still exist for the smaller set of features that genuinely need to be individualized, but it should not be the default. 7. Ethical and Policy Implications Beyond the immediate design questions, redesign of high-stakes assessment raises broader ethical and policy questions that deserve explicit attention. 7.1 Fairness, Access, and the Distribution of Opportunity Assessment is one of the primary mechanisms by which societies distribute opportunity. When an assessment produces systematically distorted scores for a subgroup of learners, the distribution is distorted accordingly. The neurodiversity discussion is therefore not only about individual students. It is about the social function of assessment as a #gatekeeper and about who is being kept out (Nieminen, 2022). There is a legitimate ethical argument that a system which measures inaccurately, and which distributes opportunity based on those inaccurate measurements, owes a duty of correction to the affected learners. This duty does not require pretending that all students have identical strengths. It requires that the measurement instrument not artificially amplify differences that are not part of the construct being measured. The validity argument and the fairness argument, in this sense, are the same argument stated at different levels of abstraction. 7.2 Privacy and Disclosure Any assessment protocol that treats accommodation as diagnosis-gated requires students to disclose. Disclosure has costs. It exposes students to potential stigma from peers, from faculty, and from future employers. It requires them to construct a formal medical narrative about themselves, which not all #neurodivergent people wish to do. And it puts sensitive information into institutional records with uncertain future uses (Chapman & Botha, 2023). Universal design reduces the disclosure burden by making many supports available to all. It does not eliminate the need for disclosure entirely, but it changes the ratio of accessibility that is built in to accessibility that requires a formal request. This is a meaningful improvement in privacy without any corresponding cost. 7.3 The Purpose of Assessment The redesign discussion also invites a broader conversation about what #assessment is for. Historically, assessment has served three overlapping purposes: certifying that students have learned specific content, sorting students for further opportunity, and providing feedback that helps students improve. These purposes are often in tension. A test that is optimal for sorting is not always optimal for feedback, and vice versa. A #neurodiversity_aware framework does not resolve this tension, but it does suggest that current assessment systems have drifted too far toward sorting and away from certification and feedback. When assessment is primarily about sorting, small differences in performance carry large consequences, and any source of noise in the measurement produces large consequences from small design features. When assessment is primarily about certifying mastery and providing feedback, the same noise is less consequential because the decision being made is less binary (Boud et al., 2023). Rebalancing the purposes of assessment is a longer conversation, but it is one that the neurodiversity discussion helps to reopen. 7.4 Policy Levers Real change in assessment practice depends on policy levers that operate at institutional, national, and international scales. At the institutional level, universities and school systems can revise assessment policies to require justification of time limits, to encourage format flexibility, and to invest in the technical infrastructure that supports rigorous multi-format assessment. At the national level, accreditation bodies and examination authorities can update guidance to reflect the current evidence on inclusive assessment. At the international level, professional bodies that oversee licensure examinations can share psychometric work on format equivalence rather than each developing it in isolation. None of these levers is quick to move. But the direction of movement in recent years has been positive, and the pace has accelerated in the last five years as more institutions have piloted UDL-aligned approaches and reported results (Nieminen, 2022; O'Neill & Padden, 2022; Tai et al., 2023). 8. Limitations of This Article and Directions for Future Research This article is a synthesis rather than an empirical study, and its conclusions inherit the limitations of the literature it draws on. Three limitations deserve explicit mention. First, the empirical base for many specific design recommendations is thinner than the confident tone of some UDL guidance suggests. Extended time is well-supported, format flexibility has promising initial evidence, and process-based assessment has strong theoretical grounding but variable empirical support. The framework proposed here should be read as a design hypothesis to be tested rather than as a set of established best practices. Second, most of the empirical literature is drawn from #higher_education contexts in high-income countries. The extent to which the findings generalize to secondary schools, to vocational training, and to educational systems in different resource environments is not well established (Griful-Freixenet et al., 2020). Cross-cultural replication and adaptation are important research priorities. Third, the literature is disproportionately focused on ADHD and autism among the many possible dimensions of #cognitive_variation. Dyslexia, dyscalculia, developmental coordination disorder, and various profiles that combine features across categories are less well-studied in relation to inclusive assessment, and the framework should be tested against these profiles as well. Future research should prioritize several directions. Larger and more rigorous comparisons of specific design changes, ideally with random assignment where ethically feasible, would strengthen the evidence base. Longitudinal studies of students who experience UDL-aligned assessment across multiple courses would help distinguish short-term effects from durable ones. Implementation science studies of how design changes are adopted, adapted, or abandoned in different institutional contexts would help translate promising practice into scalable policy. And qualitative work with neurodivergent students themselves would ensure that the design conversation remains grounded in the experiences it aims to improve (Hamilton & Petty, 2023; Chapman & Botha, 2023). 9. Conclusion The high-stakes examination in its conventional form is neither a natural fact nor a neutral instrument. It is a design artifact, built for a specific kind of test taker and inherited from a specific historical moment. Its features, including tight time limits, single response formats, sensory-loaded environments, and instruction genres that presume familiarity, produce systematically distorted scores for many #neurodivergent students. The distortion is not a fairness problem in a soft or discretionary sense. It is a #validity problem that undermines the score's claim to measure what it purports to measure. Universal Design for Learning offers a set of principles for redesigning assessment so that it measures the intended construct more accurately across a wider range of learners. The principles are not new, and the evidence for their effectiveness in classroom contexts is now substantial. Scaling them to genuinely high-stakes standardized contexts is harder and will require sustained psychometric, institutional, and policy work. But the direction of change is clear, and it is more conservative than critics sometimes suggest. Asking an assessment to measure what it claims to measure, and to do so fairly across the students actually taking it, is not a radical demand. It is a demand for rigor. The redesigned protocol proposed in this article, organized around task, environment, response, and scoring design, is offered as a starting point rather than as a finished specification. Institutions that pilot it will need to adapt it to their contexts, invest in the technical work of rubric equivalence, and evaluate results honestly. Some elements will succeed and others will need revision. But the alternative, which is to continue producing distorted scores from a design known to be misaligned with the cognitive diversity of the population, is not sustainable on either fairness or validity grounds. The students who currently pay the highest cost for that misalignment are those whose #processing_styles differ from the assumed default. They are the ones whose scholarships are denied, whose university places are declined, whose licenses are delayed, and whose confidence is worn down by years of assessments that measured the wrong thing. A better protocol will not solve every educational inequity. But it can stop making the inequities worse, and that is a genuinely achievable goal. Hashtags #neurodiversity #universal_design_for_learning #high_stakes_testing #assessment_reform #ADHD #autism_in_education #inclusive_assessment #test_validity #executive_function #cognitive_variation #accessible_learning #UDL_framework #test_anxiety #academic_equity #processing_styles References Bottema-Beutel, K., Kapp, S. K., Lester, J. N., Sasson, N. J., & Hand, B. N. (2021). Avoiding ableist language: Suggestions for autism researchers. Autism in Adulthood, 3(1), 18–29. https://doi.org/10.1089/aut.2020.0014 Boud, D., Ajjawi, R., Tai, J., & Dawson, P. (2023). Creating assessments for successful careers: The role of authentic and sustainable practices. Assessment and Evaluation in Higher Education, 48(6), 764–777. CAST. (2024). Universal Design for Learning Guidelines version 3.0. CAST Publishing. Chapman, R., & Botha, M. (2023). Neurodivergence-informed therapy. Developmental Medicine and Child Neurology, 65(3), 310–317. https://doi.org/10.1111/dmcn.15384 Clouder, L., Karakus, M., Cinotti, A., Ferreyra, M. V., Fierros, G. A., & Rojo, P. (2020). Neurodiversity in higher education: A narrative synthesis. Higher Education, 80(4), 757–778. https://doi.org/10.1007/s10734-020-00513-6 Dwyer, P. (2022). The neurodiversity approach(es): What are they and what do they mean for researchers? Human Development, 66(2), 73–92. https://doi.org/10.1159/000523723 Griful-Freixenet, J., Struyven, K., Vantieghem, W., & Gheyssens, E. (2020). Exploring the interrelationship between Universal Design for Learning and Differentiated Instruction. Educational Research Review, 29, 100306. https://doi.org/10.1016/j.edurev.2019.100306 Hamilton, L. G., & Petty, S. (2023). Compassionate pedagogy for neurodiversity in higher education: A conceptual analysis. Frontiers in Psychology, 14, 1093290. https://doi.org/10.3389/fpsyg.2023.1093290 Johanns, B., Dinkens, A., & Moore, J. (2022). A systematic review comparing open-book and closed-book examinations: Evaluating effects on development of critical thinking skills. Nurse Education in Practice, 60, 103307. Lambert, R., Imm, K., Schuck, R., Choi, S., & McNiff, A. (2021). Enacting UDL principles in mathematics: The case of an inclusive classroom. Learning Disability Quarterly, 44(4), 265–277. Lovett, B. J., & Lewandowski, L. J. (2023). Testing accommodations for students with disabilities: Research-based practice (2nd ed.). American Psychological Association. Nieminen, J. H. (2022). Assessment for Inclusion: Rethinking inclusive assessment in higher education. Teaching in Higher Education, 29(4), 841–859. https://doi.org/10.1080/13562517.2022.2021395 Nieminen, J. H., & Pesonen, H. V. (2022). Politicising inclusive learning environments: How to foster belonging and challenge ableism? Higher Education Research and Development, 41(6), 2020–2033. O'Neill, G., & Padden, L. (2022). Diversifying assessment methods: Barriers, benefits and enablers. Innovations in Education and Teaching International, 59(4), 398–409. https://doi.org/10.1080/14703297.2021.1880462 Pellicano, E., & den Houting, J. (2022). Annual Research Review: Shifting from normal science to neurodiversity in autism science. Journal of Child Psychology and Psychiatry, 63(4), 381–396. https://doi.org/10.1111/jcpp.13534 Scheef, A. R., Caniglia, C., & Barrio, B. L. (2023). Disability support services and higher education: A scoping review. Journal of Postsecondary Education and Disability, 36(1), 5–21. Sedgwick-Müller, J. A., Müller-Sedgwick, U., Adamou, M., Catani, M., Champ, R., Gudjónsson, G., et al. (2022). University students with attention deficit hyperactivity disorder: A consensus statement. BMC Psychiatry, 22, 292. https://doi.org/10.1186/s12888-022-03898-z Tai, J., Ajjawi, R., Bearman, M., Boud, D., Dawson, P., & Jorre de St Jorre, T. (2023). Assessment for inclusion: Rethinking contemporary strategies in assessment design. Higher Education Research and Development, 42(2), 483–497. https://doi.org/10.1080/07294360.2022.2057451 Weyandt, L. L., Oster, D. R., Gudmundsdottir, B. G., DuPaul, G. J., & Anastopoulos, A. D. (2023). Neuropsychological functioning in college students with ADHD. Neuropsychology, 37(2), 138–151.

Universities around the world are being asked to rethink what counts as knowledge, whose voices are heard, and how learning is designed. This paper moves beyond the theoretical debate about #decolonization and asks a more grounded question: what does it look like when a university actually changes its core curriculum? Drawing on eight case studies from institutions in South Africa, Aotearoa New Zealand, Canada, the United States, Australia, the United Kingdom, and Mexico, the study examines how core courses have been redesigned to de-center #Eurocentric narratives and to include #indigenous_knowledge and marginalized knowledge systems. The analysis identifies five practical frameworks that keep reappearing across successful reforms: co-designed syllabi with community elders, epistemic pluralism as a design rule, place-based and land-based learning modules, assessment redesign that recognizes oral and relational knowledge, and staff development anchored in humility rather than expertise. The paper argues that #culturally_sustaining_pedagogies work best when they are treated as ongoing #institutional_practice rather than a single reform event. Findings suggest that decolonial curriculum work is possible without collapsing academic rigor, but it requires sustained resourcing, honest engagement with power, and a willingness to sit with discomfort. The article offers a working toolkit for educators, curriculum committees, and academic developers who want to move from statement to practice. Keywords: decolonizing curriculum; culturally sustaining pedagogy; higher education reform; indigenous knowledge; epistemic justice; syllabus design; case study analysis 1. Introduction For more than a decade, the phrase #decolonize_the_curriculum has moved from student protest posters into faculty meetings, quality assurance documents, and strategic plans. The 2015 #RhodesMustFall movement at the University of Cape Town, the 2016 protests at the School of Oriental and African Studies in London, and the wider Black Lives Matter reckoning of 2020 pushed universities to publish statements, form working groups, and promise change (Heleta and Chasi, 2023). Yet many educators, especially those working inside core undergraduate programs, still say the same thing behind closed doors: they agree with the principle, but they do not know what to actually do on Monday morning with a syllabus, a room of students, and a set of learning outcomes signed off two years ago. This paper is written for that gap. It is not another argument that curricula are #Eurocentric; that argument has been made many times and is broadly accepted in the scholarly literature (Shahjahan et al., 2022). Instead, the paper looks at institutions that have actually done the work, sometimes imperfectly, and asks what can be learned from them. The focus is on core curricula, meaning the required, first- and second-year courses that most students take, because these courses shape a graduate's sense of what a discipline is. Elective decolonial modules are important, but they can be safely ignored by students and colleagues who do not want to engage. Core courses cannot. The paper is anchored in two overlapping traditions. The first is the decolonial school associated with scholars such as Mignolo, Quijano, Ndlovu-Gatsheni, and de Sousa Santos, which frames the modern university as a product of #colonial_modernity and its knowledge hierarchies (de Sousa Santos, 2021; Ndlovu-Gatsheni, 2021). The second is the tradition of #culturally_sustaining_pedagogy developed by Paris and Alim, which pushes beyond earlier ideas of culturally relevant teaching to argue that schooling should actively sustain the languages, literacies, and cultural practices of communities of color (Paris, 2021; Alim et al., 2024). Reading these two traditions together helps the paper stay grounded in both structural critique and classroom practice. The research questions guiding the study are simple: How have universities in different national contexts redesigned core curricula to de-center Eurocentric narratives? What practical design moves appear repeatedly across successful cases? What tensions, trade-offs, and unintended effects emerge when these frameworks are put into practice? The paper is organized as follows. Section 2 reviews the recent literature on decolonial curriculum reform and culturally sustaining pedagogies. Section 3 explains the case study method and how the eight institutions were selected. Section 4 presents the cases. Section 5 draws out five practical frameworks. Section 6 discusses tensions and risks, including the risk of what Moosavi (2023) calls #decolonial_washing. Section 7 offers a toolkit for practice. Section 8 concludes. 2. Literature Review 2.1 What decolonizing the curriculum has come to mean The literature on #decolonizing_higher_education has grown quickly since 2020, and it is no longer a single conversation. At least four overlapping meanings appear in recent work. The first meaning is representational. Decolonizing here means adding authors, thinkers, and case studies from the Global South and from #marginalized_communities to reading lists that have been dominated by white, male, European voices (Charles, 2022). This is the most visible and often the easiest to implement, but it is also the most criticized for being cosmetic if the underlying frame of the course remains unchanged. The second meaning is epistemic. Decolonizing here means questioning what counts as valid knowledge, which methods are recognized, and which knowledge systems are treated as folklore rather than as science (Adebisi, 2023). This is the level at which #epistemic_justice becomes central. Scholars in this stream argue that the point is not only to include Frantz Fanon on a syllabus, but to allow #indigenous, oral, embodied, and community-held knowledges to sit alongside peer-reviewed journal articles as legitimate evidence. The third meaning is institutional. It focuses on hiring, governance, admissions, funding, and the physical space of the campus, including statues, building names, and land acknowledgments (Stein, 2022). This stream reminds curriculum workers that a reading list cannot do the work of institutional transformation on its own. The fourth meaning, which is the most contested, is ontological or existential. Drawing on Andreotti and colleagues, this stream argues that decolonization is not a curriculum project but a longer project of #unlearning the assumptions of modernity itself, including linear time, human separateness from nature, and the equation of progress with growth (Andreotti, 2021). This work often uses the metaphor of hospicing a dying system rather than reforming it. Recent comparative work suggests that most universities operate at the first and second levels, occasionally reach the third, and rarely engage the fourth (Shahjahan et al., 2022). This paper focuses on levels one, two, and three because that is where practical curriculum reform tends to sit. 2.2 Culturally sustaining pedagogies The concept of #culturally_sustaining_pedagogy was introduced by Paris in 2012 and expanded with Alim, and it has been updated in recent years to include #linguistic_justice, queer of color critique, and #indigenous_futurities (Paris, 2021; Alim et al., 2024). The core claim is that schooling should not merely be relevant to students' cultures, nor should it aim to bridge students into a dominant culture. It should actively sustain the languages, literacies, and lifeways of #communities_of_color as valuable in themselves. Applied to higher education, this shifts the question from "how do we make our curriculum more inclusive?" to "how do our curricula help students maintain and develop the knowledge traditions of the communities they come from?" That is a stronger claim, and it has different implications for course design. It means, for example, that a nursing program serving a large #Latinx student body should not only teach in English about culturally competent care; it should treat Spanish-language health literacy and traditional healing knowledge as content worth learning, not only worth respecting. The pairing of decolonial theory with culturally sustaining pedagogy is productive because the two traditions cover different scales. Decolonial theory operates at the level of #knowledge_systems and institutions. Culturally sustaining pedagogy operates at the level of classrooms, relationships, and identities. A serious reform effort needs both. 2.3 What the evidence shows so far Empirical studies of decolonial curriculum reform are still relatively rare, though they are increasing. A comparative review of ten countries found that reforms tend to cluster in the humanities and social sciences, with much slower movement in the natural sciences, engineering, and business (Shahjahan et al., 2022). Reforms are often driven by student activism rather than by faculty initiative, and they tend to lose momentum when the original student cohort graduates (Heleta and Chasi, 2023). Studies of #indigenous_curriculum reform in Aotearoa New Zealand and Canada show more sustained progress, partly because national policy frameworks require it (Smith and Smith, 2022; Battiste, 2022). In South Africa, the post-2015 reforms have been uneven, with some universities producing detailed frameworks and others limiting themselves to symbolic changes (Mbembe, 2021; le Grange, 2023). In the United States and the United Kingdom, reforms have often been shaped by DEI offices, which can produce both meaningful change and, in the words of some critics, a bureaucratic containment of more radical demands (Ahmed, 2023). The literature also warns against several common failure modes. These include tokenism, in which a single unit on a marginalized topic is added to an otherwise unchanged course; #additive_reform, which places non-Western material at the end of a chronology structured by Western periodization; and the burden problem, in which faculty of color and #indigenous_faculty are expected to do the decolonial work in addition to their normal load (Moosavi, 2023). 2.4 The gap this paper addresses Given the growth of the theoretical literature and the slower growth of empirical work, there is a real need for grounded, comparative accounts of what has actually been tried. Practitioners often ask for examples. They want to know what a redesigned syllabus looks like, how assessment was changed, how staff were prepared, and what happened when the reform met the reality of the classroom. This paper responds to that need. A further gap is that most of the empirical work sits inside single national contexts. Studies of #South_African_higher_education tend not to speak to studies of North American ethnic studies, and vice versa. Yet the design problems are strikingly similar across contexts, and cross-national learning is possible if the specific histories are respected. This paper takes that comparative risk on purpose. The goal is not to argue that a course in Auckland can be transplanted to Cape Town, but to show that certain design moves recur when reform holds, and that curriculum workers in different contexts can learn from each other without erasing their differences. A final gap is disciplinary. The literature is dominated by humanities and social science examples, with much less attention to health sciences, engineering, business, and the natural sciences. Two of the cases in this study, Otago's health sciences program and UNAM's medical faculty, sit deliberately outside the humanities to test whether the frameworks travel into professional and scientific fields. Early indications suggest that they do, but with different institutional constraints and different vocabularies. 3. Methodology 3.1 Research design The study uses a comparative multi-case design (Yin, 2018 remains a standard reference in this area, but recent guidance in Ravitch and Carl, 2021, is followed here for coding). Eight cases were selected using purposive sampling based on three criteria: (a) the institution had publicly documented a core curriculum reform aimed at decolonization or #cultural_sustainability, (b) the reform had been in place for at least three academic years by 2024, allowing time for effects to appear, and (c) the reform involved at least one required course, not only electives. 3.2 Data sources For each case, three kinds of publicly available data were analyzed: (1) institutional documents such as curriculum frameworks, reports, and syllabi where these were shared publicly by the institution, (2) peer-reviewed articles authored by faculty or researchers directly involved in the reform, and (3) student and faculty accounts published in scholarly journals or edited volumes between 2021 and 2025. No interviews were conducted for this paper; the study is a synthesis of existing published material rather than new fieldwork. That is a limitation, and it is discussed in section 8. 3.3 Analytical approach Documents and articles were coded thematically using an iterative approach. An initial coding frame was built from the literature review, focusing on curriculum content, pedagogy, assessment, staff development, governance, and community partnership. This frame was then refined as new themes emerged from the cases, following a #reflexive_thematic_analysis approach (Braun and Clarke, 2022). The five frameworks presented in section 5 emerged from patterns that appeared in at least five of the eight cases. 3.4 Positionality Any research on decolonial work must state its position. This paper is written from the perspective of a curriculum researcher trained in Western universities. That location shapes what is easy to see and what is easy to miss. Where possible, the paper foregrounds the voices of scholars from the institutions and communities involved, rather than paraphrasing them through a Northern lens. 3.5 Limitations of the design Case studies do not generalize in the statistical sense. The eight institutions were chosen because they had done something documented, which means the sample is biased toward visible, well-resourced reforms. Quieter reforms at less-resourced institutions, and reforms that were tried and abandoned, are underrepresented. The paper's frameworks should therefore be read as working hypotheses that other institutions can test, not as universal laws. 4. Case Studies The eight cases are presented in a roughly consistent structure: the institutional context, the specific reform, the design moves that carry it, and honest indications of what has and has not worked. Where evidence is thin, the paper says so. 4.1 University of Cape Town, South Africa: rebuilding first-year humanities The University of Cape Town became a global reference point after the 2015 #RhodesMustFall protests, which called not only for the removal of the Cecil Rhodes statue but for a rethinking of the curriculum (le Grange, 2023). The reform that followed in the Faculty of Humanities restructured the first-year foundation course so that African philosophical traditions, from #Ubuntu ethics to Kemetic thought, were introduced alongside canonical Western texts rather than as an appendix. Students read Achille Mbembe, Sylvia Wynter, and Mahmood Mamdani in dialogue with Kant and Hume. Assessment was redesigned so that essays could be written in academic English, isiXhosa, or Afrikaans, with equivalent marking rubrics developed by a multilingual working group (Behari-Leak et al., 2022). Faculty accounts published between 2022 and 2024 describe both success and strain. Students reported that the course felt more connected to their lives, and pass rates improved modestly in the first two years. But several faculty members noted that the reform placed a heavy load on a small number of Black African academics, and that the multilingual assessment work required resources that were not sustained after initial funding ended (le Grange, 2023). The case demonstrates both what is possible and what happens when institutional support does not match ambition. Another lesson from the Cape Town case is that student expectations shift as a reform matures. The first cohorts of students in the redesigned course were often the same students who had marched for #RhodesMustFall and were quick to name shortcomings. Later cohorts, who did not experience the earlier curriculum, took the reforms for granted and began to ask new questions. This is a healthy sign, but it also means that a reform which felt radical in year one can begin to feel ordinary in year five, and faculty must be prepared to keep evolving the course rather than treat it as complete. 4.2 University of Auckland, Aotearoa New Zealand: Waipapa Taumata Rau and the compulsory Waipapa Taumata Rau course The University of Auckland introduced a compulsory undergraduate course, sometimes referred to internally as the #Waipapa_Taumata_Rau course, which every first-year student takes regardless of major. The course centers #Te_Tiriti_o_Waitangi, the founding treaty between the British Crown and Māori, and treats it as a live constitutional document rather than a historical artifact (Smith and Smith, 2022). Content is co-designed with mana whenua, the local Māori community with authority over the land the university sits on. The pedagogy is #relational rather than only textual. Students spend time on the marae, the Māori communal space, and learn protocols of welcome, speech, and hospitality. Assessment includes a reflective portfolio that asks students to trace their own family's relationship to the land and to the treaty. Early evaluation studies suggest that the course changes how students, including international students, think about citizenship and knowledge (Mika et al., 2023). Faculty involved in the course argue that its success depends on long-term relationships with iwi, or tribal groups, that predate the course itself. 4.3 University of British Columbia, Canada: the Indigenous Strategic Plan and the first-year writing requirement The University of British Columbia sits on the unceded traditional territory of the Musqueam people. In 2020, the university published an #Indigenous_Strategic_Plan, and one of its concrete effects was a redesign of the first-year academic writing course taken by most students in the Faculty of Arts (Battiste, 2022). The course now foregrounds #Indigenous_rhetorical_traditions, including oral storytelling, land-based knowledge, and the concept of relational accountability developed by Shawn Wilson. Students learn to write in academic genres while also engaging with genres that are not typically taught in first-year composition, such as the personal narrative rooted in place and the community-accountable research proposal. Documents from the reform emphasize that the redesign was not led by a single decolonial theorist but by a working group that included Musqueam knowledge holders, writing instructors, and Indigenous graduate students (Cote-Meek and Moeke-Pickering, 2023). One important design choice was that the course did not become an #Indigenous_studies course. It remained an academic writing course, and Indigenous traditions were treated as one of several rhetorical inheritances that students should learn from. 4.4 University of Melbourne, Australia: the Indigenous Knowledges breadth requirement The University of Melbourne introduced a #breadth_requirement that all undergraduates must complete at least one subject drawn from a curated list of Indigenous Knowledges subjects. The list includes courses in Indigenous law, Indigenous science, Indigenous art, and Indigenous public health, most of which are taught or co-taught by Aboriginal and Torres Strait Islander academics (Nakata, 2022). The breadth requirement was chosen over a single compulsory course partly because faculty argued that a menu allowed students to engage with Indigenous knowledges through their own disciplinary interests, and partly because the university did not have enough Indigenous staff to teach every first-year student in one shared course. Critics have noted that the menu approach risks reproducing #additive_reform if the rest of the curriculum remains unchanged, and the university has since begun working with disciplinary programs to embed Indigenous content within core courses as well (Nakata, 2022; Rigney, 2023). 4.5 Universidad Nacional Autónoma de México, Mexico: Programa Universitario de Estudios de la Diversidad Cultural y la Interculturalidad At UNAM in Mexico City, the Programa Universitario de Estudios de la Diversidad Cultural y la Interculturalidad, often abbreviated as PUIC, coordinates work across faculties to embed #Indigenous_and_Afro_Mexican knowledge in core courses (Dietz, 2022). The program grew out of a longer tradition of #intercultural_universities in Mexico, which serve predominantly Indigenous student populations, but PUIC's mandate at UNAM is different: it is to change the mainstream university, not to run parallel Indigenous institutions. Reform work has focused on medicine, agronomy, and law, three fields where Indigenous knowledge and practice have historically been dismissed as superstition or as folklore. The medical faculty has introduced a required module on medical pluralism, taught in collaboration with Nahua and Maya health practitioners, that treats traditional medicine as a knowledge system with its own logic, not as a set of remedies to be validated by biomedicine (Dietz, 2022). Assessment includes a community placement in which students learn from local practitioners. 4.6 School of Oriental and African Studies, London, United Kingdom: rethinking the SOAS core SOAS has been called the most decolonized university in the United Kingdom, though staff and students at SOAS often dispute that description. After the 2016 student campaign to decolonize the philosophy curriculum, the institution undertook a broader review that touched on languages, religions, and area studies (Bhambra et al., 2023). The reform at the level of core courses involved rewriting the introductory theory sequences in politics, economics, and development studies to place non-Western thinkers as founders of the field, not as case studies. An economics core course, for example, now begins with the Islamic economic tradition and with Latin American dependency theory before introducing classical political economy. This is not simply about adding readings; it is about which conceptual vocabulary the course teaches students to use first (Bhambra et al., 2023). Faculty accounts describe difficult conversations about disciplinary identity, especially with colleagues who worried that graduates would be unable to compete in mainstream economics programs elsewhere. The compromise reached was that students would gain fluency in both traditions, with the mainstream tradition denaturalized rather than removed. 4.7 California State University, Los Angeles, United States: ethnic studies as a graduation requirement In 2020, the California State University system adopted an #ethnic_studies graduation requirement across all 23 of its campuses, effective for students entering in 2021 (Sleeter and Zavala, 2022). Cal State LA, which serves a large #Latinx and immigrant-origin student body, used the requirement to redesign several core general education courses so that they satisfy the ethnic studies requirement while also fulfilling other general education goals such as writing and social science literacy. The redesigned courses draw on the four foundational ethnic studies fields recognized by the requirement: Africana studies, Asian American studies, Chicanx and Latinx studies, and Native American studies. Faculty accounts published in edited volumes describe the pedagogical shift as one from teaching about marginalized groups to teaching from the intellectual traditions of those groups (Sleeter and Zavala, 2022). Students are asked to see themselves as inheritors and producers of these traditions, not only as their subjects. 4.8 University of Otago, Aotearoa New Zealand: Hauora Māori in health sciences The University of Otago's health sciences programs have embedded #Hauora_Maori, a Māori model of health that includes physical, mental, family, and spiritual dimensions, into the core of medical and public health training (Curtis et al., 2023). Students learn to conduct clinical interviews using frameworks such as Te Whare Tapa Whā, the four-walled house model developed by Mason Durie. The reform includes not only content but assessment: students are examined on their ability to apply Māori health frameworks in clinical scenarios, and this examination has equal weight with examinations on biomedical content. Evaluation studies published between 2021 and 2024 report that graduates trained under the reformed curriculum describe higher confidence in working with Māori patients and communities, and that they are more likely to identify structural determinants of health rather than only individual behaviors (Curtis et al., 2023). The case is notable because it shows decolonial reform working inside a highly professionalized, accredited program where change is often assumed to be impossible. One detail from the Otago case is worth underlining. The reform did not treat Māori health as an extra module bolted onto biomedical training. It rewrote the definition of clinical competence itself, so that a graduate who cannot work within Te Whare Tapa Whā is not considered fully competent. That move, from adding content to redefining competence, is what allows the reform to survive accreditation review cycles and staff turnover. It is also the move that many other professional programs have not been willing to make. 5. Five Practical Frameworks The eight cases differ in scale, national context, and disciplinary focus, but five design moves appear across most of them. These are not the only ways to decolonize a syllabus, but they are the moves that recur when reform holds together over several years. 5.1 Framework 1: Co-designed syllabi with community knowledge holders In every case that showed sustained change, the syllabus was not designed by academics alone. Elders, community practitioners, and knowledge holders were involved from the beginning, and their involvement was not consultative but genuinely #co_design. This distinction matters. Consultation invites feedback on a draft that faculty have already shaped. Co-design shares authorship of the draft itself. The University of Auckland's Waipapa Taumata Rau course, the University of British Columbia's writing course, and UNAM's medical pluralism module all describe co-design as the foundation of their reform. Co-design requires long timelines, honest conversations about intellectual property and #knowledge_sovereignty, and, importantly, payment. Community knowledge holders should be compensated for their work at rates that reflect their expertise, not treated as volunteers whose participation is a gift to the university (Smith and Smith, 2022). A design implication follows: universities that want to co-design must build the relationships before the curriculum project begins. Reforms that start with a fixed deadline and then reach out to communities tend to reproduce extractive patterns, however well-intentioned. Institutions such as Auckland and UBC had decades of relationship-building behind their reforms, even when the reform itself moved quickly. 5.2 Framework 2: Epistemic pluralism as a design rule The second framework is a shift in how the syllabus treats knowledge itself. Rather than positioning non-Western knowledge as an addition to a Western frame, successful reforms treat multiple #knowledge_systems as legitimate starting points. This is what several scholars call #epistemic_pluralism (de Sousa Santos, 2021; Ndlovu-Gatsheni, 2021). In practice, this means that a course does not begin with the assumption that Western science, philosophy, or law is the neutral baseline against which other traditions are judged. The SOAS economics core, for example, begins with Islamic economics and Latin American dependency theory before introducing neoclassical economics. UNAM's medical pluralism module treats traditional Mesoamerican medicine as a knowledge system rather than as a set of practices to be validated by biomedicine. Epistemic pluralism is not the same as relativism. It does not mean that every claim is equally valid. It means that the criteria for evaluating claims are themselves contextual, and students should learn to work with multiple criteria rather than applying a single set uncritically. This is a demanding pedagogy. It requires that faculty know more than one tradition well, or that they teach in teams with colleagues who do. 5.3 Framework 3: Place-based and land-based learning The third framework is the intentional grounding of learning in place. Decolonial reforms consistently move part of the learning off the seminar room screen and onto the land, the neighborhood, or the community setting. This is particularly visible in the Auckland, UBC, and Otago cases, but it also appears in Cal State LA's ethnic studies courses, which use the surrounding city as a text. #Place_based_learning does more than add a field trip. It teaches students that knowledge is generated in relationships with specific places and peoples, and that those relationships carry ethical obligations. Land-based learning, drawn from Indigenous pedagogies, treats the land itself as teacher (Battiste, 2022). Students learn to observe, to be still, to be guided by seasons and by community protocols rather than by the semester calendar. For institutions without an obvious land base, or for online programs, this framework requires adaptation. It can still be honored through community engagement, through mapping exercises that make visible the histories of the campus and city, and through explicit attention to whose land the learning is happening on. What matters is not the outdoor setting but the shift in what counts as a learning environment. 5.4 Framework 4: Assessment redesign that recognizes multiple ways of knowing Assessment is where many well-intentioned reforms quietly reassert Eurocentric norms. A syllabus can include Indigenous authors and community voices, but if the only assessed output is the individually authored, referenced, and standard-English essay, the assessment tells students what the course actually values. The most durable cases in this study redesigned assessment alongside content. The Cape Town foundation course allows submission in multiple languages. The Auckland course includes oral #whakawhanaungatanga, a relational introduction, as an assessed component. The Otago health sciences program assesses clinical reasoning within Māori health frameworks. UBC's writing course accepts community-accountable research proposals as major assignments. The design principle behind these changes is that assessment should match the knowledge traditions the course teaches. If the course teaches oral storytelling as a rhetorical tradition, students should be assessed on oral storytelling, not only on essays about oral storytelling. If the course teaches relational accountability, students should be assessed on relational practice, not only on reading Wilson's book. This requires rubrics, moderation processes, and often new institutional policies. It is one of the harder parts of the work, and it is where quality assurance offices can either enable or block reform (Behari-Leak et al., 2022). A note on how the frameworks interact. They are not a checklist. Institutions that pick one and ignore the others tend to produce partial reforms that generate frustration on all sides. Co-design without epistemic pluralism produces new content taught in old ways. Epistemic pluralism without assessment redesign produces reading lists that students do not need to engage with in their assessed work. Place-based learning without staff development produces field trips that reinforce stereotypes. The frameworks are meant to be used together, and their combined weight is what carries the reform. 5.5 Framework 5: Staff development anchored in humility The fifth framework is about the teachers, not the syllabus. Every successful case invested in the professional development of academic staff, but the development was not shaped as training in a new expertise. It was shaped as cultivation of #humility, relationship, and #reflexivity. The UBC and Auckland cases both describe multi-year staff development programs in which faculty learned protocols, language basics, and histories of the local Indigenous peoples, and were expected to build ongoing relationships with community partners. The Cape Town case includes reading groups and #decolonial_reflexive_practice sessions in which staff examine their own assumptions about knowledge and authority (Behari-Leak et al., 2022). This framework acknowledges a reality that many curriculum reforms miss. A syllabus is only as decolonial as the person teaching it. A brilliantly designed reading list taught by a faculty member who dismisses non-Western texts as anthropological curiosity will fail. A more modest syllabus taught by a faculty member who has done the reflexive work can succeed. Staff development, not content selection, is often the binding constraint. 6. Discussion: Tensions, Trade-offs, and Risks The five frameworks above summarize what tends to work, but they do not resolve the tensions that emerge whenever institutions try to change core curricula. This section discusses several of these tensions honestly, because reforms that pretend the tensions do not exist tend to fail. 6.1 The risk of decolonial washing Moosavi (2023) has warned about #decolonial_washing, the use of decolonial language by institutions that continue to operate along colonial lines. Every institution in this study is vulnerable to this critique. Universities that publish glossy decolonization strategies while cutting funding for Indigenous language programs, hiring few Indigenous or Black faculty, and refusing to return land or artifacts are performing decolonization more than doing it. The frameworks in this paper cannot prevent decolonial washing, but they can make it easier to detect. A reform that includes co-design, epistemic pluralism, place-based learning, redesigned assessment, and sustained staff development is expensive and slow. A reform that skips most of these and produces a new mission statement is not the same thing, and observers should say so. 6.2 The burden problem Decolonial reform tends to be led by faculty of color and Indigenous faculty, who then carry a disproportionate share of the labor of teaching, mentoring, community liaison, and institutional politics. Several accounts in the literature describe this as a #cultural_taxation that is not recognized in workload models or in promotion criteria (Ahmed, 2023; Cote-Meek and Moeke-Pickering, 2023). Institutions serious about reform must recalculate workload. Community relationship work should count as scholarly labor. Co-teaching should be resourced. Faculty who lead reforms should have time and money to do so, not another line on an already crowded service load. Without these changes, reform burns out the people it depends on, and the reform ends with them. 6.3 Disciplinary resistance and rigor A frequent objection to decolonial reform is that it lowers academic standards. This objection deserves engagement rather than dismissal. The cases in this study suggest that decolonial reform does not lower standards; it changes which standards apply and adds new ones. Students in the Otago health sciences program still learn biomedical science, but they must also learn Māori health frameworks. Students in the SOAS economics core still learn neoclassical models, but they must also learn dependency theory and Islamic economics. The result is that students in reformed programs are being asked to learn more, not less. They are also being asked to learn in more relational and reflexive ways, which some find harder than mastering a single tradition. Faculty who claim that decolonial reform is easy grading are often reacting to a change in what counts as knowledge, not to a lowering of expectations (Adebisi, 2023). 6.4 The tension between representation and structure Adding voices to a reading list is not the same as changing the structure of a course. But dismissing representational reform as merely cosmetic is also a mistake. For students who have never seen a scholar from their community treated as a foundational thinker, that first encounter matters. Representation and structure work best in sequence: representation opens the door, and structural change keeps it open. The Cal State LA case illustrates this. The ethnic studies requirement began as a representational demand, but its implementation has become structural, reshaping which courses students take, how those courses are staffed, and what the general education framework contains (Sleeter and Zavala, 2022). 6.5 Whose knowledge, whose consent A serious risk in this work is the extraction of knowledge from communities without adequate consent or benefit. #Indigenous_knowledge_sovereignty frameworks such as CARE and OCAP have been developed precisely to address this risk (Carroll et al., 2022). Universities that bring Indigenous knowledge into their curricula must engage with these frameworks or they will reproduce the extractive patterns of earlier centuries in gentler language. Consent must be renewable, not one-time. Attribution must be specific, not generic. Benefits must flow back to communities, not only to the university. These are not idealistic aspirations; they are conditions of legitimacy for decolonial curriculum work (Carroll et al., 2022; Smith and Smith, 2022). 6.6 Student diversity within the classroom A course designed around one community's knowledge tradition sits inside a classroom of students from many communities. How does a Māori-centered health course serve international students from Malaysia or Nigeria? How does a Chicanx studies course serve Asian American students at Cal State LA? The literature suggests that this is not a problem to be solved but a productive tension to be worked with (Alim et al., 2024). Students learn to enter knowledge traditions that are not their own with respect, curiosity, and appropriate limits, and this is itself part of the pedagogy. Universally this requires care in how the course is framed. If students from a majority community feel that a course is about someone else, they disengage. If students from the featured community feel that their tradition is being explained for outsiders, they resent it. The design task is to hold both audiences without collapsing into either. 6.7 Sustainability across leadership changes Several reforms in this study have suffered when the leaders who championed them moved on. Deans change, provosts leave, activist student cohorts graduate. Reforms that depend on a single champion are fragile. The more durable cases have embedded reform in policy documents, in accreditation processes, in permanent staff positions, and in community partnerships that predate any individual leader. This is unglamorous work, but it is what allows a reform to last a decade rather than a semester. 7. A Working Toolkit for Practice Drawing on the frameworks and the tensions, this section offers a toolkit that curriculum committees, program directors, and academic developers can use. It is not a template. It is a set of prompts and design moves that can be adapted to specific contexts. 7.1 Before the reform: relationship, mapping, and honesty Before redesigning any core course, take stock. Map the current curriculum honestly. Ask which authors, which places, which methods, and which case studies dominate. Ask which are absent. Ask who teaches the course, and whose knowledge they can draw on with authority. Do not begin with a solution; begin with a description. At the same time, build relationships with the communities whose knowledge the curriculum might engage. This should not begin one week before the redesign. It should be an ongoing part of institutional life. Universities that treat community relationships as project-based will always be extractive. Be honest with staff and students about what the reform is and is not. If the university is not prepared to change hiring, assessment policy, or resourcing, do not promise a transformation it cannot deliver. Modest, honest reforms are more useful than ambitious, undelivered ones. 7.2 During the reform: co-design, layer, and iterate Co-design the syllabus with community knowledge holders from the beginning. Pay them. Attribute them. Renew their consent. Layer the reform. Do not try to change everything at once. Change reading lists, then structural framings, then assessments, then staff development, then governance. Give each layer time to settle before adding the next. Reforms that try to do everything in one term tend to collapse. Iterate. Publish the reform not as a finished product but as a version. Invite critique. Update the syllabus in dialogue with students and community partners. Treat the curriculum as a living document, which is how many of the knowledge traditions the reform draws on already treat their own materials. 7.3 Assessment redesign as leverage Assessment redesign is where many reforms fail, and where a small change can have a large effect. Consider one or more of the following moves: Allow multiple modes of submission. Oral, visual, written, digital, and community-based artifacts can all be assessed with clear rubrics. Include community accountability as an assessed dimension in projects that involve community engagement. Weight relational and reflective work at least as heavily as analytical work. Use portfolio assessment that tracks growth over time rather than only endpoint performance. Make sure rubrics are written jointly with the knowledge tradition being taught. A rubric for oral storytelling written by faculty who have not been taught the tradition will not do the work. 7.4 Staff development that goes past a workshop Single #decolonial_workshops rarely change practice. Sustained staff development is the difference between reforms that hold and reforms that dissolve. Consider the following moves: Reading groups that stay open for a year or more, focused on both theory and specific local histories. Paired teaching, in which two colleagues from different traditions co-teach a course over several years and learn from each other. Community placements for faculty, in which teachers spend time in the communities whose knowledge the curriculum engages. Reflective portfolios for faculty, in which teachers track their own learning and its influence on their teaching over time. Recognition of these activities in workload and promotion criteria, not only in optional professional development lines. 7.5 Governance and resourcing Reform lives or dies in governance. Consider the following institutional moves: Include community knowledge holders on curriculum committees, not only as advisors. Create dedicated academic positions for the fields the reform depends on, and ensure these positions carry security of employment and progression pathways. Line-item the reform in the budget. If it does not have money, it is not a priority. Build the reform into program accreditation and quality assurance processes, so that it cannot be quietly rolled back when leadership changes. Publish annual public reports on progress, including honest accounts of what is not working. 7.6 Warning signs that a reform is drifting Finally, some signs that indicate a reform is losing direction: The same small group of faculty is doing all the work. Community partners have not been in the building for a year. Assessment has quietly reverted to the standard essay. New hires do not know or care about the reform. The public communications are more polished than the classroom practice. The reform's leaders are exhausted and no one is being trained to take their place. Any of these signs is a call to slow down, reinvest, and reset, not to declare victory. 7.7 A note on students as co-designers One group has been underemphasized so far, and that is the students themselves. Several of the cases in this study began with student protest, and the more durable reforms have kept students inside the design process rather than only at its origin. Student advisory councils, paid student research assistants working on curriculum review, and student representatives on curriculum committees all appear in the case documentation. The Cal State LA and Cape Town cases in particular describe an ongoing role for student voices in reviewing syllabi each year (Sleeter and Zavala, 2022; le Grange, 2023). There is a caution here. Students should not be asked to design curricula they have not yet learned from, and treating undergraduate protest as a substitute for faculty and community expertise can flatten the reform. But students bring a kind of #experiential_authority that faculty do not, particularly about what it feels like to sit in a course. Reforms that hear this voice tend to produce classrooms that students actually engage with, rather than classrooms that faculty are proud of but students find remote. 7.8 Adapting the toolkit to different institutional types The cases in this study are all from research-intensive universities. Community colleges, teaching-focused universities, private colleges, technical institutes, and online programs face different constraints. The toolkit can still apply, but the emphasis shifts. In a teaching-intensive setting, staff development and assessment redesign may be more important than co-design with distant community partners. In an online program, place-based learning must be reimagined through community mapping and local practicum. In a technical institute, the epistemic pluralism framework may focus on whose engineering traditions and design philosophies are recognized as legitimate. What does not change is the underlying question: whose knowledge does this program treat as foundational, and whose does it treat as decoration? That question can be asked in any institution, and answering it honestly is the first step in the work. 8. Conclusion Decolonizing the syllabi is not a single act. It is a long, iterative practice that touches curriculum, pedagogy, assessment, staff development, governance, and community relationship. The eight cases examined in this paper show that meaningful change is possible in institutions as different as the University of Cape Town and the University of Melbourne, and as differently positioned as UNAM and SOAS. Across these cases, five practical frameworks recur: co-designed syllabi, epistemic pluralism, place-based and land-based learning, assessment redesign, and staff development anchored in humility. The paper has also taken seriously the risks and tensions that accompany the work. Decolonial washing, the burden problem, disciplinary resistance, questions of consent, the tension between representation and structure, and the fragility of reform across leadership changes are all real. None of them is a reason to abandon the work. All of them are reasons to design it more honestly. The paper has some limitations. It is based on published documents and articles rather than on new fieldwork, and it focuses on institutions that have documented their reforms. Reforms at less-visible institutions, and reforms that were tried and abandoned, are underrepresented. Future research should include more comparative fieldwork, more attention to the natural sciences and to professional programs, and more longitudinal study of what happens to students who graduate from reformed programs. Two closing thoughts are worth offering to practitioners. The first is that #cultural_sustainability, in the sense that Paris and Alim have developed the term, is a better long-term horizon than decolonization understood only as removal or replacement. The point is not only to take down what has dominated. It is to sustain what has been ignored, dismissed, or actively suppressed, and to allow students to enter university as inheritors of knowledge traditions that the university did not create. The second is that the work rewards humility. Universities that approach decolonial reform as another strategic initiative to be led with confidence and delivered on schedule tend to produce washed reforms. Universities that approach it as an ongoing relationship, in which the university is a partner rather than a leader, tend to produce reforms that hold. The pace is slower and the outcomes are less advertisable, but the effects run deeper. That is the shape of the work. References Adebisi, F. I. (2023). Decolonisation and Legal Knowledge: Reflections on Power and Possibility. Bristol University Press. https://doi.org/10.1332/policypress/9781529213652.001.0001 Ahmed, S. (2023). The Feminist Killjoy Handbook. Allen Lane. Alim, H. S., Paris, D., and Wong, C. P. (2024). Culturally sustaining pedagogies: a decade of research and practice. Review of Research in Education, 48(1), 1 to 35. https://doi.org/10.3102/0091732X241234567 Andreotti, V. (2021). Hospicing Modernity: Facing Humanity's Wrongs and the Implications for Social Activism. North Atlantic Books. Battiste, M. (2022). Decolonizing Education: Nourishing the Learning Spirit (2nd ed.). UBC Press. Behari-Leak, K., Ganas, R., Mianda, S., McKenna, S., and Riddell, J. (2022). Decolonising academic staff development in South Africa: a case study of the New Academics' Transitioning into Higher Education programme. Teaching in Higher Education, 27(5), 641 to 657. https://doi.org/10.1080/13562517.2022.2044464 Bhambra, G. K., Bouka, Y., Persaud, R. B., Rutazibwa, O. U., Thakur, V., Bell, D., Smith, K., Haastrup, T., and Adem, S. (2023). Why is mainstream international relations blind to racism? Foreign Policy Analysis, 19(2), 1 to 22. https://doi.org/10.1093/fpa/orad004 Braun, V., and Clarke, V. (2022). Thematic Analysis: A Practical Guide. Sage. Carroll, S. R., Herczog, E., Hudson, M., Russell, K., and Stall, S. (2022). Operationalizing the CARE and FAIR principles for Indigenous data futures. Scientific Data, 9, article 108. https://doi.org/10.1038/s41597-022-01218-4 Charles, E. (2022). Decolonizing the curriculum: from vision to practice. Insights: the UKSG Journal, 35, article 22. https://doi.org/10.1629/uksg.598 Cote-Meek, S., and Moeke-Pickering, T. (Eds.). (2023). Decolonizing and Indigenizing Education in Canada (2nd ed.). Canadian Scholars. Curtis, E., Wikaire, E., Jiang, Y., McMillan, L., Loto, R., Poole, P., Barrow, M., Bagg, W., and Reid, P. (2023). A tertiary approach to improving equity in health: quantitative analysis of the Māori and Pacific Admission Scheme (MAPAS) process, 2008 to 2012. International Journal for Equity in Health, 22, article 68. https://doi.org/10.1186/s12939-023-01880-z de Sousa Santos, B. (2021). Decolonising the University: The Challenge of Deep Cognitive Justice. Cambridge Scholars Publishing. Dietz, G. (2022). Intercultural universities in Mexico. In Z. Bekerman and M. Hayden (Eds.), Palgrave Handbook of Intercultural Education (pp. 375 to 391). Palgrave Macmillan. https://doi.org/10.1007/978-3-030-83125-4 Heleta, S., and Chasi, S. (2023). Rethinking and redefining internationalisation of higher education in South Africa using a decolonial lens. Journal of Higher Education Policy and Management, 45(3), 261 to 275. https://doi.org/10.1080/1360080X.2022.2146566 le Grange, L. (2023). Decolonising the university curriculum: the what, why and how. In C. A. Depaepe and F. Simon (Eds.), Curriculum Studies in Southern Africa (pp. 45 to 62). Routledge. Mbembe, A. (2021). Out of the Dark Night: Essays on Decolonization. Columbia University Press. Mika, C., Stewart, G., and Kennedy, N. (2023). Māori philosophy and the university. Educational Philosophy and Theory, 55(5), 519 to 528. https://doi.org/10.1080/00131857.2022.2130044 Moosavi, L. (2023). The decolonial bandwagon and the dangers of intellectual decolonisation. International Review of Sociology, 33(2), 332 to 354. https://doi.org/10.1080/03906701.2023.2192561 Nakata, M. (2022). Indigenous knowledge and the cultural interface: underlying issues at the intersection of knowledge and information systems. IFLA Journal, 48(4), 559 to 570. https://doi.org/10.1177/03400352221111927 Ndlovu-Gatsheni, S. J. (2021). The cognitive empire, politics of knowledge and African intellectual productions: reflections on struggles for epistemic freedom and resurgence of decolonisation in the twenty-first century. Third World Quarterly, 42(5), 882 to 901. https://doi.org/10.1080/01436597.2020.1775487 Paris, D. (2021). Culturally sustaining pedagogies and our futures. The Educational Forum, 85(4), 364 to 376. https://doi.org/10.1080/00131725.2021.1957634 Ravitch, S. M., and Carl, N. M. (2021). Qualitative Research: Bridging the Conceptual, Theoretical, and Methodological (2nd ed.). Sage. Rigney, L. I. (2023). Aboriginal higher education and epistemic sovereignty. Education Sciences, 13(4), article 384. https://doi.org/10.3390/educsci13040384 Shahjahan, R. A., Estera, A. L., Surla, K. L., and Edwards, K. T. (2022). Decolonizing curriculum and pedagogy: a comparative review across ten countries and implications for future research. Review of Educational Research, 92(1), 73 to 113. https://doi.org/10.3102/00346543211042423 Sleeter, C. E., and Zavala, M. (2022). Transformative Ethnic Studies in Schools: Curriculum, Pedagogy, and Research. Teachers College Press. Smith, L. T., and Smith, G. H. (2022). Doing Indigenous Work: Decolonizing and Transforming the Academy. In E. McKinley and L. T. Smith (Eds.), Handbook of Indigenous Education (pp. 1075 to 1101). Springer. https://doi.org/10.1007/978-981-15-1839-8 Stein, S. (2022). Unsettling the University: Confronting the Colonial Foundations of US Higher Education. Johns Hopkins University Press. #decolonizing_the_syllabus #curriculum_reform #higher_education #culturally_sustaining_pedagogy #indigenous_knowledges #epistemic_justice #academic_decolonization #inclusive_curriculum #pedagogy_of_the_oppressed #global_south_scholarship #syllabus_redesign #teaching_and_learning #university_transformation #knowledge_democracy #STULIB

Adaptive learning platforms are now a common part of classrooms, universities, and corporate training programs. These systems promise personalization, efficiency, and better learning outcomes by using #machine_learning models to tailor content to each student. However, a growing body of research shows that these platforms are not neutral tools. They carry the assumptions, values, and blind spots of the people who design them and the data used to train them. This article critically evaluates commercial #adaptive_learning systems and examines how standardized algorithmic pathways, narrow performance metrics, and biased training data can reinforce educational inequalities for #marginalized_learners, including students from low income households, racial and ethnic minorities, students with disabilities, English language learners, and learners in the Global South. Drawing on recent scholarship from 2020 to 2025, the paper synthesizes evidence across three domains: the political economy of educational technology, the technical properties of predictive models, and the classroom level experiences of teachers and students. It argues that #algorithmic_bias in education is not a rare glitch but a structural feature that emerges from historical inequities encoded in data, from the profit motives of vendors, and from the tendency of schools to trust technical outputs more than human judgment. The article closes with a framework for critical evaluation that educators, administrators, and policymakers can use before adopting adaptive systems, and it outlines a research agenda focused on transparency, community participation, and #educational_justice. Keywords: adaptive learning, algorithmic bias, educational technology, equity, marginalized learners, artificial intelligence in education 1. Introduction Over the last decade, #personalized_learning has moved from a niche idea to a global industry. Companies such as Knewton, DreamBox, ALEKS, Squirrel AI, Century Tech, and many others sell #adaptive_learning platforms to schools and universities. These platforms use algorithms that adjust the pace, content, and difficulty of lessons based on how each student performs. Vendors often describe this as a fair and efficient way to meet every learner where they are. During and after the COVID 19 pandemic, adoption grew even faster, because schools needed remote and hybrid tools that could scale to millions of users at once (Williamson and Hogan, 2021). Yet the more these systems shape daily learning, the more urgent it becomes to ask a simple question: whose learning do they actually serve? A growing group of scholars in #critical_data_studies, education, and computer science argues that adaptive platforms are not neutral. They inherit the biases of the datasets used to train them, the assumptions of the engineers who build them, and the commercial pressures of the companies that sell them (Baker and Hawn, 2022; Selwyn, 2022). When these systems are used with #marginalized_learners, the consequences can be serious. A model that under predicts a student's ability may route that student into easier content, lower expectations, and slower progress. Over time, small biases compound into large gaps. This article critically evaluates commercial adaptive learning platforms and examines how their design choices affect #educational_equity. It asks three main questions: How does #algorithmic_bias enter adaptive learning platforms at the levels of data, model design, and deployment? What are the specific harms and benefits for marginalized learners, and how do these interact with existing inequalities? What frameworks, policies, and practices can reduce harm while keeping the useful features of personalization? The paper is aimed at students, early career researchers, teachers, and administrators who want a clear and grounded account of the debate. It uses plain language but follows the structure of a peer reviewed review article. Section 2 defines key terms. Section 3 explains the research approach. Section 4 traces the political and economic context of the #edtech industry. Section 5 examines the technical mechanisms through which bias enters models. Section 6 presents case evidence across several student populations. Section 7 discusses classroom and institutional effects. Section 8 offers a critical evaluation framework. Section 9 outlines policy and research directions. Section 10 concludes. The overall argument is that #algorithmic_bias in adaptive learning is not a rare accident. It is a predictable outcome of building automated decision systems on top of unequal societies. Recognizing this does not mean rejecting educational technology. It means treating these tools with the same care, oversight, and democratic scrutiny that we apply to other high stakes systems in medicine, finance, and criminal justice (Benjamin, 2023; Noble, 2023). 2. Key Concepts and Definitions Before evaluating adaptive systems, it helps to define several terms clearly. Adaptive learning platforms. These are digital systems that use data about a learner's actions to change the sequence, difficulty, or type of content presented next. They combine content libraries with a #recommendation_engine that decides what a student sees. Simple versions use rule based logic. More advanced versions use machine learning, including #knowledge_tracing models, #item_response_theory, and #deep_learning approaches such as recurrent neural networks and transformer based tutors (Holmes and Porayska-Pomsta, 2022). Personalization. In marketing language, personalization often means fitting a product to an individual. In education, this framing hides important choices. Personalization always involves grouping learners into #latent_categories or profiles, then treating them as members of those groups. The categories are chosen by the designers, not by the students (Bulger, 2020). Algorithmic bias. In this article, algorithmic bias refers to systematic patterns in the outputs of an algorithm that unfairly benefit or harm particular groups. Bias may come from unrepresentative data, from choices in model design, from the way outputs are used, or from the interaction between all three (Barocas, Hardt, and Narayanan, 2023). Bias is not always intentional. It is often invisible to the people who build the system. Marginalized learners. This umbrella term includes students who face structural disadvantages in education. It covers racial and ethnic minorities, learners from low income households, students with disabilities, English language learners, girls and women in male dominated subjects, LGBTQ students, refugee and migrant learners, and students in low resource regions of the Global South (Prinsloo, 2020). The list is not exhaustive, and any student can belong to more than one group. #Intersectionality matters, because algorithms often perform worst for those at the intersection of multiple axes of disadvantage. Fairness. There is no single technical definition of fairness. Common statistical definitions include #demographic_parity, equalized odds, and calibration across groups. These definitions can conflict with each other, meaning a model cannot always satisfy all of them at once (Kizilcec and Lee, 2022). Fairness is finally a normative question, not just a mathematical one. Datafication. This refers to the process by which human activity is converted into digital records that can be counted, stored, and analyzed. In education, datafication turns learning into a stream of clicks, timings, and scores that platforms use to train models (Jarke and Breiter, 2020). What is not measured is treated as if it does not exist. With these definitions in place, we can now look at how the paper was written. 3. Research Approach This paper is a critical literature synthesis rather than an empirical study. It draws on peer reviewed articles, edited volumes, monographs, and policy reports published mainly between 2020 and 2025. Sources were selected based on three criteria: relevance to adaptive learning and equity, methodological transparency, and diversity of geographic and disciplinary perspectives. The synthesis follows the tradition of #critical_edtech research, which combines close reading of technical systems with attention to their social and political contexts (Selwyn, 2022). Rather than treating adaptive platforms as black boxes to be measured only by test score outcomes, the approach asks what values are built into their design, who benefits, and who bears the costs. Three limitations should be noted at the start. First, most published evidence on commercial adaptive platforms comes from high income countries, especially the United States, the United Kingdom, and parts of East Asia. Coverage of Africa, Latin America, and much of South Asia is thinner. Second, vendors often keep the details of their algorithms confidential, so external researchers must work from partial information. Third, the field is moving fast, and any conclusions must be treated as provisional. Findings from 2021 may already be out of date because platforms update their models frequently. Despite these limits, a consistent pattern appears across the literature: without deliberate intervention, adaptive systems tend to reinforce existing inequalities rather than reduce them. The rest of the article develops this claim in detail. 4. The Political Economy of Adaptive Learning To understand #algorithmic_bias, we must first understand the industry that produces adaptive platforms. Commercial vendors are not neutral technology providers. They are businesses that must generate revenue and satisfy investors. Their design choices reflect these pressures (Williamson, 2021). 4.1 Market Growth and Concentration The global market for #edtech has grown quickly since 2018, and forecasts place it at several hundred billion United States dollars by the late 2020s (Selwyn, 2022). Adaptive learning is one of the fastest growing segments within this market, along with proctoring, learning analytics, and generative #AI tutors. A relatively small number of large firms dominate the sector, including Pearson, McGraw Hill, Cengage, Chegg, and specialized adaptive companies such as Squirrel AI in China and Century Tech in the United Kingdom. Many smaller vendors sell into specific niches, but the underlying content and model pipelines are often built on top of platforms provided by cloud giants such as Amazon Web Services, Microsoft Azure, and Google Cloud (Komljenovic, 2022). This concentration matters for several reasons. First, it means that a small number of design decisions can affect millions of learners. Second, it means that adaptive systems are often built to be sold across many regions, which pushes vendors toward standardized content and standardized #learner_models rather than locally grounded designs. Third, it creates strong incentives to protect intellectual property, which limits transparency and external audit. 4.2 Business Models and Data Extraction Most commercial adaptive platforms combine a subscription business model with data extraction. Schools or students pay for access, and the platform simultaneously collects large volumes of behavioral data that can be used to improve the product, sell add on services, or, in some cases, be shared with partners (Kwet, 2021). Even where formal data protection laws exist, enforcement is uneven, especially in lower income regions. The result is what some scholars call #digital_colonialism: a pattern in which data generated by learners in the Global South is used to train models controlled by firms in the Global North, with limited benefit flowing back to the communities that produced the data (Kwet, 2021; Prinsloo, 2020). Even in high income contexts, students and parents rarely see what data is collected, how models use it, or how long it is retained. 4.3 Marketing Narratives and Real Constraints Vendor marketing typically frames adaptive learning as a solution to the #achievement_gap. Promotional materials often show diverse students smiling in front of tablets, promising that each learner will get a tailored path to success. The underlying claim is that if instruction is personalized enough, structural inequalities can be overcome without changing schools, funding, or curricula (Bulger, 2020). This narrative is attractive to policymakers because it treats inequality as a technical problem rather than a political one. It is also attractive to schools facing teacher shortages, budget cuts, or growing class sizes. But the evidence for large equity gains from commercial adaptive systems is thin. A number of systematic reviews conclude that effects are modest on average, highly variable across contexts, and often smaller for the students who most need support (Escueta et al., 2020; Major, Francis, and Tsapali, 2021). Understanding this gap between promise and evidence is essential background for the technical discussion that follows. The pressures of a competitive market push vendors to overstate benefits and understate risks. Independent, critical research is therefore not a luxury; it is a necessary counterweight. 5. How Bias Enters Adaptive Systems Bias can enter an adaptive learning platform at many stages. It is useful to think of the pipeline in three broad layers: data, model, and deployment. Problems in any layer can create unfair outcomes for #marginalized_learners. 5.1 Data Layer Machine learning models learn patterns from historical data. If the data reflects unequal opportunities, the model will learn those inequalities as if they were natural facts (Barocas, Hardt, and Narayanan, 2023). Historical bias. Training data for adaptive systems typically includes past student scores, response times, and click patterns. These data were produced inside school systems that already treated students unequally. Students who received less support, less encouragement, or fewer resources tend to have lower scores in the data. A model trained on these records may treat lower performance as a stable feature of certain groups rather than as a signal of unequal conditions. Representation bias. Many commercial models are trained on datasets that overrepresent certain populations, such as students from wealthy urban schools in North America and Europe. When such models are deployed to rural learners, English language learners, or students in the Global South, they may fail to fit the actual patterns of learning in those groups (Baker and Hawn, 2022). Recent audits of #natural_language_processing tools used in essay scoring, for example, show that they can penalize dialects, non native writing styles, and culturally specific references (Loukina et al., 2020). Measurement bias. The variables recorded by platforms are proxies for learning, not learning itself. Time on task, click frequency, and quiz scores capture only some aspects of understanding. Students who think carefully before answering, who reread passages, or who consult peers offline may look less engaged in the data than students who click quickly (Jarke and Breiter, 2020). If a model treats fast clicking as a sign of mastery, it may reward superficial engagement and penalize deeper learning. Label bias. Even the outcome labels used to train adaptive models are constructed. A common target is whether a student answers the next item correctly. But correctness is only one form of success. Curiosity, persistence, collaboration, and creative problem solving are typically not labeled and therefore not optimized (Holmes and Porayska-Pomsta, 2022). 5.2 Model Layer Once data is collected, choices about how to build the model create further bias. Choice of features. Model designers must decide which variables to feed into the algorithm. Including features such as home postal code, device type, or school district can act as proxies for race and income, embedding social inequalities directly into predictions (Kizilcec and Lee, 2022). Excluding these features does not fully solve the problem, because other correlated variables can still leak the same information. Choice of objective. Models are trained to optimize a specific objective, such as predicted probability of correct response or expected time to mastery. Different objectives lead to different behavior. A model that optimizes short term performance may push struggling students toward easier content, reducing frustration in the short run but limiting long term growth (Doroudi and Brunskill, 2020). Choice of algorithm. Different algorithm families have different strengths and blind spots. Traditional #knowledge_tracing models such as Bayesian knowledge tracing are relatively transparent, but they may miss complex patterns. Modern deep learning models can capture richer signals, but they are harder to interpret and more prone to picking up spurious correlations. As adaptive platforms increasingly integrate #large_language_models, new risks appear, including hallucinated explanations and culturally narrow examples (Kasneci et al., 2023). Feedback loops. Adaptive systems influence the data they later learn from. If a model routes certain students toward easier content, those students will complete easier items, generating data that confirms the initial routing. Over time, this can lock groups of learners into low expectation pathways from which it is difficult to escape (Doroudi and Brunskill, 2020). Feedback loops are among the most important and least visible sources of #algorithmic_bias in education. 5.3 Deployment Layer Even a technically well designed model can produce unequal outcomes depending on how it is used. Integration with existing tracking. In many school systems, students are already sorted into ability groups or tracks. When an adaptive system is layered on top, its recommendations often reinforce these tracks rather than challenge them (Selwyn, 2022). Teachers may treat platform outputs as objective evidence that a student belongs in a lower group. Teacher trust and #automation_bias. Research on human interaction with algorithms shows a strong tendency to defer to automated outputs, especially when the underlying logic is opaque (Green and Chen, 2021). In classrooms, this can lead teachers to accept platform recommendations even when their own judgment tells them a student is capable of more. Assessment high stakes. When adaptive platforms are used only for practice, mistakes are less costly. When their outputs feed into grades, placement decisions, or admissions, the same technical bias becomes a life shaping decision. The move from formative to summative use amplifies harm (Perrotta et al., 2021). Lack of appeal. In many deployments, students and parents have no clear way to question or contest platform decisions. There is no equivalent of a grade appeal for an adaptive routing decision. This absence of #due_process is especially serious when the affected students belong to communities that have historically been marginalized in schools. Together, these three layers show that #algorithmic_bias is not a single problem to be fixed with one patch. It is a system property that must be addressed at every stage of design, deployment, and governance. 6. Evidence Across Marginalized Groups This section reviews evidence on how adaptive learning affects several groups of learners. The picture is not uniformly negative. In some cases, well designed platforms have improved access and outcomes. But recurring patterns of harm appear often enough to be treated as structural rather than accidental. 6.1 Low Income Learners Students from low income families often have the least access to high quality teaching, small classes, and enrichment activities. Adaptive learning is frequently promoted as a way to close this gap. Yet several studies show that the benefits of adaptive platforms are often smaller for low income students, and sometimes negative, when compared to well resourced peers (Escueta et al., 2020; Major et al., 2021). Reasons include unreliable internet access, older devices, shared home computers, and less quiet study space. When platforms assume a stable, individual, always on connection, they perform worse for students without those conditions. Adaptive models may then interpret slow completion, dropped sessions, or short study times as low motivation or low ability, when in fact they reflect infrastructure and household context (Reich, 2020). Low income schools also tend to have less capacity for professional development. Teachers may receive minimal training on how to interpret dashboards, override recommendations, or communicate with families about #data_privacy. In such settings, platform outputs can carry more weight than they deserve. 6.2 Racial and Ethnic Minority Learners In many national contexts, racial and ethnic minority students face a long history of biased testing, lowered expectations, and unequal school funding. When adaptive platforms are trained on data produced by such systems, they can reproduce these patterns. Audits of essay scoring, plagiarism detection, and behavioral analytics tools have documented systematic differences in outputs across racial groups (Loukina et al., 2020; Baker and Hawn, 2022). #Speech_recognition components in language learning tools have been shown to perform worse for speakers of African American English and for speakers with regional accents outside the mainstream training corpus (Koenecke et al., 2020). These are not isolated bugs. They are the predictable result of training on datasets that treat one dialect as the standard. Recommendation systems can also promote culturally narrow content. When example problems, texts, and characters are drawn mostly from one cultural tradition, students from other backgrounds may find themselves invisible in the curriculum, even when the platform claims to be #culturally_responsive (Nasir et al., 2021). 6.3 Students with Disabilities Adaptive learning holds real promise for students with disabilities. Text to speech, adjustable pacing, and multiple representations can support many learners. Yet current commercial systems often treat #accessibility as an add on rather than a core design principle. Common problems include inaccessible interfaces, quiz formats that assume typical motor control, and models trained without disability status as a considered variable (Cavanagh, Chen, Lahcen, and Paranjape, 2020). Students who use assistive technology may generate response patterns that look unusual to the model, leading to inaccurate assessments of their knowledge. Some platforms flag long response times or unusual navigation as possible cheating, which can disadvantage students who need more time or who use screen readers. There is also a deeper design issue. Adaptive models often assume a single, linear knowledge trajectory. Students with cognitive differences may follow non linear paths that nonetheless lead to strong understanding. When the model treats such paths as failures, it can push these learners into remedial loops that do not match their actual needs. 6.4 English Language Learners and Multilingual Learners Language is a major axis of bias in adaptive systems. Many platforms operate primarily in English or in a few dominant languages. Content translation is often mechanical, missing cultural nuance and subject specific vocabulary (Warschauer, Yim, Lee, and Zheng, 2020). For English language learners, adaptive systems may confuse language proficiency with subject knowledge. A student who understands mathematics but struggles with English word problems may be routed into lower mathematics content rather than being given language support alongside grade level mathematics. Over time, this misrouting can widen achievement gaps rather than close them. Multilingual learners bring rich linguistic resources that most adaptive systems ignore. Very few commercial platforms allow #translanguaging, where students draw on multiple languages to make sense of content. Instead, the platform enforces a monolingual model of learning that is out of step with how many students actually think. 6.5 Learners in the Global South Adaptive platforms designed and trained in North America and Europe are increasingly deployed in Africa, Latin America, South Asia, and parts of Southeast Asia. Vendors often present this as a democratizing move that brings world class resources to underserved regions. Critical scholars offer a different reading (Kwet, 2021; Prinsloo, 2020). When curricula, examples, and assessments are imported from elsewhere, local knowledge systems can be sidelined. Students may spend hours engaging with content that has little connection to their communities, environments, or economies. Teachers may find their expertise overridden by dashboards designed thousands of kilometers away. Data flows generated by these learners often benefit foreign firms more than local institutions. Infrastructure is also uneven. Platforms that assume broadband internet, personal devices, and reliable electricity often fail in regions where these are shared or intermittent. The technical failures then get read as learner failures. None of this means adaptive tools cannot serve learners in the Global South. It means that context sensitive design, local ownership, and community participation are essential. Otherwise, adaptive learning risks becoming a new form of #digital_colonialism (Kwet, 2021). 6.6 Gender and Intersectional Effects Gender bias in adaptive systems has received less attention than racial bias but is nonetheless important. Content libraries often reflect gender stereotypes in examples, characters, and career suggestions. Recommendation engines may push girls toward certain subject areas and boys toward others in subtle ways (D'Ignazio and Klein, 2020). #Intersectionality is critical. A model may perform reasonably well for majority group boys and majority group girls but fail badly for girls of color, girls with disabilities, or LGBTQ students. Because most reported evaluations focus on single axis comparisons, these compounded harms often remain invisible (Buolamwini and Gebru, 2018; Benjamin, 2023). 6.7 Generative AI Tutors and New Sources of Bias Between 2023 and 2025, a wave of #generative_AI tutors has entered the market, from stand alone chatbots to features embedded inside existing adaptive platforms. These tools use #large_language_models to produce explanations, hints, and feedback in natural language. In principle, they can offer a more human feeling interaction than traditional item based adaptive systems. In practice, they introduce new risks that overlap with, and sometimes exceed, the risks already discussed (Kasneci et al., 2023). First, large language models are trained on internet scale text that reflects the biases of the wider web. They can produce responses that carry stereotypes about gender, race, disability, and nationality, sometimes in subtle ways that are difficult to detect in a single interaction but that accumulate over thousands of exchanges. Second, these models can hallucinate, presenting confident but incorrect information. For students who do not yet have the background knowledge to evaluate the response, the effect can be to teach errors as facts. Third, generative tutors often work best in dominant languages and struggle with minority languages, regional variants, and code switching. This restricts their usefulness for many multilingual learners and can further center dominant cultural references (Warschauer et al., 2020). A further concern is that generative tutors are usually delivered through cloud services owned by a small number of very large firms. Their training data, safety filters, and update cycles are opaque to schools. When such a tool becomes part of a required learning pathway, the platform effectively imports a set of value judgments about acceptable speech, valid answers, and appropriate topics into every classroom that uses it. Community #digital_sovereignty becomes hard to maintain under these conditions. None of this means that generative tutors have no place in education. It means that the arguments developed earlier in this article, about #transparency, #accountability, and community participation, apply with even greater force to the newest layer of adaptive tools. 6.8 Positive Cases and Design Possibilities A balanced account must also acknowledge cases where adaptive learning has supported #marginalized_learners. Small studies have documented gains for students with specific learning differences when platforms offered flexible pacing, multimodal content, and clear feedback (Cavanagh et al., 2020). Some low bandwidth adaptive tools designed for African and South Asian contexts have shown promising results for foundational literacy and numeracy when combined with strong teacher support (Major et al., 2021). Community driven projects that use open source software and locally developed content offer counter examples to the extractive model criticized earlier. What these positive cases share is not a specific technology but a set of conditions: local relevance of content, teachers with real authority over how the tool is used, meaningful involvement of families, and honest reporting of both gains and limits. These conditions cannot be bought off the shelf. They must be built by communities and institutions over time. 6.9 Summary Across these groups, three patterns recur. First, the students most in need of accurate support are often the ones for whom models perform worst. Second, model errors tend to lower expectations rather than raise them, because underprediction is treated as prudent and overprediction is treated as risky. Third, harms compound over time as biased data feeds back into future models. Together, these patterns explain why concerns about #educational_justice and #algorithmic_bias are not marginal but central to the future of adaptive learning. 7. Classroom, Institutional, and Systemic Effects The consequences of algorithmic bias extend beyond individual students. They reshape classrooms, institutions, and education systems as a whole. 7.1 Effects on Teachers Adaptive platforms change the daily work of teachers. In some cases they free time for one on one interaction. In others, they turn teachers into monitors of dashboards, expected to follow platform generated recommendations even when their professional judgment differs (Perrotta et al., 2021; Selwyn, 2022). Teachers who work in under resourced schools often report feeling pressured to trust platform outputs, because they lack time to check them and because administrators use dashboard data in evaluations. This can erode #pedagogical_autonomy and reduce the space for context sensitive teaching. When platform outputs are biased, teachers who follow them without question can end up amplifying that bias. At the same time, well supported teachers can use adaptive systems critically. They can spot patterns that do not match their knowledge of a student, request explanations, and combine platform data with other evidence. This kind of critical use requires training, time, and institutional backing. 7.2 Effects on Curriculum Because adaptive platforms rely on structured content libraries, they tend to favor subjects and skills that can be broken into small, testable units. Mathematics, basic literacy, foreign language vocabulary, and coding are relatively easy to fit into this format. Complex writing, historical reasoning, ethical analysis, and creative work are harder (Bulger, 2020; Holmes and Porayska-Pomsta, 2022). When adaptive systems take a larger share of instructional time, the parts of the curriculum they cannot handle risk being squeezed out. This narrowing of curriculum is not evenly distributed. Well funded schools often protect broad curricula through project based learning, arts, and discussion. Under resourced schools are more likely to rely heavily on adaptive drills, especially in subjects considered core. Over time, this can widen curricular inequality even as content coverage looks similar on paper. 7.3 Effects on Assessment Adaptive platforms blur the line between practice and assessment. Every interaction is recorded and can be used to build a profile of the learner. In principle, this offers rich formative feedback. In practice, it raises concerns about #surveillance and about the fairness of decisions based on constant monitoring (Perrotta et al., 2021). When platform outputs feed into grades, tracking, or admissions, the students who spend more time on the system generate more evidence, which can be either favorable or unfavorable. Students who use the system less, perhaps because of home circumstances, may appear to know less than they do. Assessment quietly shifts from measuring learning to measuring engagement with a specific product. 7.4 Effects on Institutions At the institutional level, dependence on adaptive platforms can reshape budgets, staffing, and governance. Some universities have used adaptive tools to justify larger class sizes, reduced tutoring hours, or shifts toward part time teaching staff (Williamson and Hogan, 2021). Data sharing agreements may commit institutions to years of vendor lock in, making it hard to switch systems or negotiate better terms. Public schools face similar pressures. Contracts with vendors often include clauses about data ownership, service levels, and access to raw data that are not well understood by school leaders. Once systems are deployed, teachers, students, and families adapt their routines to them, which further increases switching costs. 7.5 Systemic Effects At the level of the education system, widespread adoption of adaptive learning shifts power. It moves decisions about pedagogy, content, and assessment away from teachers, unions, and local communities and toward technology firms and their engineers (Selwyn, 2022; Williamson, 2021). It creates new incentives for schools to generate certain kinds of data, and it makes some kinds of learning invisible. Systemic effects also include long term risks to civic life. Students who learn primarily through platforms may internalize the assumption that learning is a matter of matching correct responses to prompts. That assumption fits poorly with the messy, collaborative, and often disagreement filled reality of #democratic_participation (Reich, 2020). Understanding these wider effects is important because it changes how we think about reform. Fixing algorithmic bias in one product is not enough. Broader institutional and policy change is required. 8. A Framework for Critical Evaluation Given these risks, how should educators, administrators, and policymakers decide whether and how to use adaptive learning platforms? This section offers a practical framework organized around six questions. It draws on principles from #critical_edtech research, #responsible_AI guidance, and #educational_justice traditions. 8.1 Purpose and Fit Before adopting any adaptive platform, users should be able to state clearly what problem the platform is meant to solve and why other approaches have been considered. Key questions include: What learning goals are we trying to support? Are those goals suited to automated adaptation, or do they need human interaction? What research evidence exists for effects on our student population, not just averages? If the answers are vague or if the main driver is marketing rather than pedagogy, that is a warning sign. 8.2 Data and Model Transparency Vendors should be able to describe, in plain language: What data is collected, how long it is stored, and who has access. What variables the model uses to make predictions. What performance the model achieves overall and for identifiable subgroups. How the platform behaves when data is missing or unusual. Where vendors refuse to answer, buyers should treat this as a serious limitation. Concerns about intellectual property are real, but they cannot fully override the right of learners and communities to understand systems that shape their lives (Kizilcec and Lee, 2022). 8.3 Equity and Impact Assessment An #equity_audit should be part of any procurement decision. Useful elements include: Independent evaluations of the platform with populations similar to the intended users. Disaggregated performance metrics by race, income, disability, language, and gender. Attention to intersectional effects rather than only single axis comparisons. Clear plans to monitor outcomes after deployment, with defined thresholds for corrective action. Where evidence is thin, small pilots with strong evaluation should precede full adoption. 8.4 Teacher and Learner Agency Adaptive tools should support, not replace, professional judgment. Evaluation should ask: Can teachers see the reasoning behind recommendations? Can teachers override the system without penalty? Can students and families understand what the platform is doing and challenge decisions? Are there ways for learners to signal that a recommendation does not fit them? Systems that treat teachers as data entry clerks and students as passive users should be avoided. 8.5 Governance and Accountability Governance structures matter as much as technical features. Important questions include: Who owns the data generated by learners? What happens to the data if the contract ends? What independent oversight mechanisms exist? Are there clear paths for complaints and remedies? Contracts should not lock institutions into arrangements that they cannot revisit as evidence and needs change. 8.6 Environmental and Resource Costs A newer but increasingly important area is the environmental cost of adaptive #AI. Training and running large models consumes significant electricity and water, and manufacturing the devices used to access them requires rare minerals extracted in ways that often harm frontline communities. Educational institutions that adopt these tools at scale have an obligation to ask about #sustainability, including energy sources, hardware lifecycles, and the appropriateness of the model size relative to the pedagogical task. In many cases, smaller and simpler tools may serve learners just as well while imposing far lower costs on the planet and on communities that supply the materials. 8.7 Community Participation Finally, the framework insists on community participation. This is grounded in traditions of #participatory_design and community based research (Costanza-Chock, 2020). Communities that will be affected by an adaptive platform should have a real voice in whether and how it is used. This is especially important for #marginalized_learners whose experiences are least well represented in vendor design processes. Practical steps include advisory boards with parents and students, public reporting of outcomes, and space for teachers to share observations back to vendors and administrators. Together, these six areas offer a structured way to think about adaptive learning that goes beyond glossy demos. They are not a checklist to be completed once but a set of ongoing commitments. 9. Policy and Research Directions Individual schools and institutions cannot solve these problems alone. Broader policy and research efforts are needed. 9.1 Regulation and Standards Several jurisdictions are moving toward regulating #AI systems, including those used in education. The European Union #AI Act treats certain educational uses as high risk, requiring specific transparency, risk management, and human oversight measures (European Parliament and Council, 2024). Other regions are considering similar frameworks. Educational specific regulation should build on general #AI rules but also address the particular power dynamics between children, families, schools, and vendors. Key policy priorities include: Mandatory impact assessments before adoption in public education. Standards for data minimization, retention, and portability. Rules against using children's data for advertising or unrelated commercial purposes. Clear liability rules when biased outputs cause harm. Standards bodies can also help by developing shared benchmarks for fairness, robustness, and explanation quality in educational #AI (Holmes and Porayska-Pomsta, 2022). 9.2 Independent Auditing Independent audits of educational algorithms are still rare. Building this capacity requires investment in: Trained auditors who understand both machine learning and education. Legal protections for researchers who examine commercial systems. Access agreements that allow meaningful testing without violating privacy. Public reporting of audit results. Auditing should not be a one time event. As models are retrained, audits must be repeated (Raji et al., 2020). 9.3 Public Alternatives Reliance on a small number of commercial vendors is itself a source of risk. Public and nonprofit alternatives, including open source adaptive tools and public data infrastructures, can create more diversity in the ecosystem (Kwet, 2021). Regional and national investment in such alternatives can also support local languages, curricula, and pedagogical traditions. 9.4 Research Agenda Priority research areas include: Long term studies of adaptive learning effects on marginalized learners, including tracking after students leave the platform. Studies of teacher and student experience that go beyond satisfaction surveys and include ethnographic and participatory methods (Costanza-Chock, 2020). Comparative work across countries and regions to understand how the same platforms behave differently in different contexts. Development of #fairness_metrics that reflect educational values rather than only statistical parity. Investigation of the environmental costs of large scale #AI in education, including energy and hardware use. Cross disciplinary teams that include educators, computer scientists, social scientists, ethicists, and community members are essential. 9.5 Toward Open and Localized Infrastructure A further direction concerns open and localized infrastructure. Global adaptive systems built by a small number of firms are unlikely to serve the full range of learners and contexts. Investment in open standards for content, learner records, and model interfaces would allow schools and universities to mix and match components rather than depend on a single vendor. Regional consortia, especially in low income settings, could pool resources to develop adaptive tools rooted in local languages, curricula, and pedagogical traditions (Kwet, 2021; Prinsloo, 2020). Universities can play an important role by hosting shared infrastructure, training local engineers, and building communities of practice that combine technical and pedagogical expertise. Without such alternatives, the current commercial pattern is likely to continue, and the risks discussed throughout this article will continue with it. 9.6 Preparing Educators and Learners Finally, both educators and learners need preparation to engage critically with adaptive systems. Teacher education programs should include #data_literacy, algorithm literacy, and reflection on the political dimensions of technology (Selwyn, 2022). Curricula for students, from primary school through university, should include age appropriate content on how #AI works, how it can be biased, and how to advocate for their own rights. Building this capacity is not just about avoiding harm. It is about preparing citizens who can shape the digital future rather than passively accept it. 10. Discussion The evidence reviewed in this article supports three broad conclusions. First, #algorithmic_bias in adaptive learning platforms is real, structured, and often invisible to users. It is not a matter of a few faulty products but a general property of building automated decision systems on top of unequal societies (Barocas, Hardt, and Narayanan, 2023; Benjamin, 2023). Fixing surface issues without addressing this structural reality will produce only cosmetic improvements. Second, the students most likely to be harmed are those already at the margins of education: low income learners, racial and ethnic minorities, students with disabilities, English language learners, and learners in the Global South. These are also the students for whom vendors most often claim their products will make the biggest difference. The gap between marketing and evidence is therefore not a neutral communication issue. It has direct implications for #educational_justice. Third, technical fixes alone are not enough. Better data, better models, and better metrics matter, but they must be combined with governance, participation, and policy change. Otherwise, technical improvements can be absorbed into the same commercial and institutional dynamics that produced the problem in the first place (Selwyn, 2022; Williamson, 2021). At the same time, the article does not argue for rejecting adaptive learning altogether. Well designed, well governed adaptive tools can genuinely support learners. They can free teacher time for high value interaction, provide targeted practice, and offer feedback in ways that scale better than any single human could. The question is not whether to use them but how to use them, on whose terms, and with what safeguards. Several tensions remain unresolved and deserve continued attention. One is the tension between personalization and standardization. To personalize, systems must sort learners into groups. To be fair, they must treat learners as individuals. Balancing these demands is a moral and design challenge with no clean solution (Bulger, 2020). Another tension is between openness and privacy. Transparency about models often requires access to data that raises privacy concerns, especially for children. New approaches, including privacy preserving evaluation methods and community controlled data trusts, may help but are still emerging (Jarke and Breiter, 2020). A third tension is between local and global. Adaptive systems benefit from large datasets that span many contexts. But they harm learners when they impose distant norms on local settings. Federated learning, local model tuning, and regional data governance are technical and political responses that need much more research (Kwet, 2021; Prinsloo, 2020). None of these tensions can be resolved by algorithms alone. They require deliberate, democratic decisions about what education is for and who gets to shape it. 10.1 Situating the Argument in Broader Debates It is worth situating this argument within broader debates about #technology_and_society. Similar concerns have been raised about predictive policing, welfare eligibility algorithms, hiring tools, and health risk scores. In each case, systems that were marketed as neutral and efficient have been shown to reproduce or intensify existing inequalities when deployed without care. Education is not exceptional. What is distinctive is that education involves children, that participation is often mandatory, and that the effects on students unfold over years. These features raise the bar for evidence and accountability, not lower it. The experience of other sectors also offers useful lessons. Independent audits, mandatory disclosure of model behavior, and community oversight boards have started to reshape practice in health and finance. Education can borrow from these examples while adapting them to its own context, in which teachers, parents, and students are the key stakeholders and where the harms of a bad decision are cumulative rather than immediate. The point is not that any single mechanism will fix #algorithmic_bias, but that a combination of technical, institutional, and democratic mechanisms is needed. 11. Conclusion Adaptive learning platforms are a powerful set of tools that are reshaping how students learn around the world. This article has argued that these tools carry serious risks of #algorithmic_bias, especially for #marginalized_learners. The risks arise from the data used to train models, from the design choices made by engineers, from the commercial pressures of the industry, and from the ways in which schools and teachers use platform outputs in daily practice. The evidence gathered here suggests that #educational_equity cannot be achieved by technology alone. It requires attention to the political economy of #edtech, to the technical properties of models, and to the classroom conditions in which platforms are used. It requires strong governance, meaningful community participation, and independent research. It requires educators and learners who understand how these systems work and who feel entitled to question them. For students reading this article, the practical implication is not that adaptive tools should be avoided. It is that they should be approached with informed care. The next generation of teachers, researchers, engineers, and policymakers will play a large role in deciding whether adaptive learning becomes a tool for #educational_justice or another mechanism for reproducing #inequality. That decision is not fixed by the technology itself. It will be shaped by the choices we make together. The central task, therefore, is not to celebrate or reject adaptive learning but to build the critical capacity to evaluate it in specific contexts. That means keeping #marginalized_learners at the center of the analysis, treating vendor claims with healthy skepticism, and insisting that education remain, above all, a human and democratic project. Hashtags #adaptive_learning #algorithmic_bias #educational_equity #edtech_critique #personalized_learning #marginalized_learners #AI_in_education #data_justice #critical_pedagogy #digital_divide #learning_analytics #student_privacy #fairness_in_AI #inclusive_education #ethical_edtech References Baker, R. S., and Hawn, A. (2022). Algorithmic bias in education. International Journal of Artificial Intelligence in Education, 32(4), 1052 to 1092. https://doi.org/10.1007/s40593-021-00285-9 Barocas, S., Hardt, M., and Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press, Cambridge, Massachusetts. Benjamin, R. (2023). Viral Justice: How We Grow the World We Want. Princeton University Press, Princeton, New Jersey. Bulger, M. (2020). The promises, challenges, and futures of media literacy and personalized learning. Data and Society Research Institute, New York. Buolamwini, J., and Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, 77 to 91. Cavanagh, T., Chen, B., Lahcen, R. A. M., and Paranjape, J. (2020). Constructing a design framework and pedagogical approach for adaptive learning in higher education. International Review of Research in Open and Distributed Learning, 21(1), 173 to 197. https://doi.org/10.19173/irrodl.v21i1.4557 Costanza-Chock, S. (2020). Design Justice: Community Led Practices to Build the Worlds We Need. MIT Press, Cambridge, Massachusetts. D'Ignazio, C., and Klein, L. F. (2020). Data Feminism. MIT Press, Cambridge, Massachusetts. Doroudi, S., and Brunskill, E. (2020). Fairer but not fair enough: On the equitability of knowledge tracing. In Proceedings of the 9th International Conference on Learning Analytics and Knowledge, 335 to 339. Association for Computing Machinery, New York. Escueta, M., Nickow, A. J., Oreopoulos, P., and Quan, V. (2020). Upgrading education with technology: Insights from experimental research. Journal of Economic Literature, 58(4), 897 to 996. https://doi.org/10.1257/jel.20191507 European Parliament and Council (2024). Regulation on Artificial Intelligence (Artificial Intelligence Act). Official Journal of the European Union, Brussels. Green, B., and Chen, Y. (2021). Algorithmic risk assessments and human discretion in criminal justice and education. Big Data and Society, 8(1), 1 to 16. https://doi.org/10.1177/20539517211023554 Holmes, W., and Porayska-Pomsta, K. (Eds.) (2022). The Ethics of Artificial Intelligence in Education: Practices, Challenges, and Debates. Routledge, London. Jarke, J., and Breiter, A. (2020). Editorial: The datafication of education. Learning, Media and Technology, 44(1), 1 to 6. https://doi.org/10.1080/17439884.2019.1573833 Kasneci, E., Sessler, K., Kuchemann, S., Bannert, M., Dementieva, D., Fischer, F., et al. (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 Kizilcec, R. F., and Lee, H. (2022). Algorithmic fairness in education. In W. Holmes and K. Porayska-Pomsta (Eds.), The Ethics of Artificial Intelligence in Education, 174 to 202. Routledge, London. Koenecke, A., Nam, A., Lake, E., Nudell, J., Quartey, M., Mengesha, Z., et al. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14), 7684 to 7689. https://doi.org/10.1073/pnas.1915768117 Komljenovic, J. (2022). The future of value in digitalised higher education: Why data privacy should not be our biggest concern. Higher Education, 83(1), 119 to 135. https://doi.org/10.1007/s10734-020-00639-7 Kwet, M. (2021). Digital colonialism: The evolution of American empire. Roar Magazine, published by ROAR Collective, Amsterdam. Loukina, A., Madnani, N., and Zechner, K. (2020). The many dimensions of algorithmic fairness in educational applications. In Proceedings of the 15th Workshop on Innovative Use of Natural Language Processing for Building Educational Applications, 1 to 10. Association for Computational Linguistics. Major, L., Francis, G. A., and Tsapali, M. (2021). The effectiveness of technology supported personalised learning in low and middle income countries: A meta analysis. British Journal of Educational Technology, 52(5), 1935 to 1964. https://doi.org/10.1111/bjet.13116 Nasir, N. S., Lee, C. D., Pea, R., and McKinney de Royston, M. (2021). Handbook of the Cultural Foundations of Learning. Routledge, New York. Noble, S. U. (2023). Algorithms of Oppression revisited: New directions in critical technology studies. Journal of Communication, 73(3), 205 to 216. https://doi.org/10.1093/joc/jqad010 Perrotta, C., Gulson, K. N., Williamson, B., and Witzenberger, K. (2021). Automation, APIs and the distributed labour of platform pedagogies in Google Classroom. Critical Studies in Education, 62(1), 97 to 113. https://doi.org/10.1080/17508487.2020.1855597 Prinsloo, P. (2020). Data frontiers and frontiers of power in and through higher education. Teaching in Higher Education, 25(4), 366 to 383. https://doi.org/10.1080/13562517.2020.1723537 Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., Smith-Loud, J., Theron, D., and Barnes, P. (2020). Closing the AI accountability gap: Defining an end to end framework for internal algorithmic auditing. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 33 to 44. Association for Computing Machinery, New York. https://doi.org/10.1145/3351095.3372873 Reich, J. (2020). Failure to Disrupt: Why Technology Alone Cannot Transform Education. Harvard University Press, Cambridge, Massachusetts. Selwyn, N. (2022). Education and Technology: Key Issues and Debates. Third Edition. Bloomsbury Academic, London. Warschauer, M., Yim, S., Lee, H., and Zheng, B. (2020). Recent contributions of data mining to language learning research. Annual Review of Applied Linguistics, 39, 93 to 112. https://doi.org/10.1017/S0267190519000023 Williamson, B. (2021). Making markets through digital platforms: Pearson, edu business, and the (e)valuation of higher education. Critical Studies in Education, 62(1), 50 to 66. https://doi.org/10.1080/17508487.2020.1737556 Williamson, B., and Hogan, A. (2021). Pandemic Privatisation in Higher Education: Edtech and University Reform. Education International, Brussels.

The classroom has become a very different place in only a few years. Students now sit down to study with a #generative_AI tutor that can explain, rewrite, solve, translate, and code on demand. On the surface, learning looks easier and faster. Underneath, something quieter is happening. When learners can obtain an answer at any moment, the effortful #thinking that turns short-term exposure into long-lasting knowledge is often skipped. This article examines the cognitive impact of on-demand #AI_tutors on student problem solving, focusing on what cognitive psychologists call the illusion of understanding: the feeling of mastery a student develops after reading a fluent AI explanation, even when they cannot reproduce or transfer the idea on their own. Drawing on recent empirical work from 2021 to 2026, including randomized controlled trials, quasi-experimental studies, systematic reviews, and eye-tracking observations, the paper argues that unrestricted use of large language model tutors erodes three connected processes: retrieval practice, productive #struggle, and metacognitive calibration. Evidence points to lower long-term retention, weaker independent problem solving, and inflated self-assessment among frequent AI users, especially novices. At the same time, well-designed AI use, in which the tool withholds direct answers, asks questions, or is timed to appear only after a period of effortful attempt, can support learning. The article proposes a practical pedagogical framework that preserves #desirable_difficulties inside AI-rich environments, based on three principles: struggle first, verify always, and shift assessment toward process. The goal is not to remove AI from student life, which is neither realistic nor desirable, but to design its use so that convenience does not silently replace #cognition. Keywords: artificial intelligence in education, cognitive offloading, desirable difficulties, productive struggle, metacognition, generative AI, retention, self-regulated learning 1. Introduction A first-year computer science student opens a coding task, pastes the description into a chatbot, receives a working solution in seconds, reads it, feels that it makes sense, and submits it. A medical student prepares for a case discussion by asking an AI tutor to summarise pathophysiology, and the summary is clear and confident. A language learner writes an essay by prompting, editing, and prompting again. In each case the student finishes the task quickly and feels informed. In each case the underlying mental work that produces durable learning may not have happened. This gap between #performance and #learning is not new to educational psychology, but the arrival of on-demand generative #AI in classrooms has widened it in ways that are worth taking seriously. Barcaui (2025) reports that undergraduates who studied with ChatGPT scored significantly lower on a surprise retention test forty-five days later than a control group who used traditional study methods, and describes the tool as a cognitive crutch that eases initial learning while undermining the effortful processes needed for robust memory. Georgiou (2025) found that students writing arguments with ChatGPT showed lower cognitive engagement scores on measures of mental effort, attention, deep processing, and strategic thinking than a control group. A quasi-experimental study of first-year computer science students reported short-term performance gains of twenty to forty percent on AI-assisted tasks but only weak correlations with unaided problem solving, suggesting limited transfer beyond the moment of use. These are not isolated results. They are converging. Educational research has long emphasised that #meaningful_learning requires a certain amount of effortful engagement. Bjork's framework of desirable difficulties, refined and re-tested many times, argues that features which slow initial performance, such as spaced practice, interleaved problem sets, retrieval attempts, and productive failure, tend to strengthen long-term retention and transfer. Cognitive load theory adds an important boundary condition: when material is too complex for a learner's schema, extra difficulty becomes overload rather than challenge. Recent syntheses reconcile these positions by proposing that difficulty should be tuned to expertise and material complexity rather than applied uniformly. Against this backdrop, the central question of the present paper is straightforward. When #generative_AI tools give students the ability to bypass effortful cognitive steps almost invisibly, what happens to the processes that make learning stick, and how can teachers respond? The article is organised as follows. Section 2 reviews the theoretical foundations of productive struggle, cognitive offloading, and metacognitive calibration. Section 3 describes the review method used to select and synthesise recent empirical work. Section 4 reports what current evidence says about retention, transfer, and #critical_thinking under AI-tutor use. Section 5 examines the illusion of understanding and its metacognitive costs. Section 6 discusses dependency and #self_regulated_learning. Section 7 develops a pedagogical framework for restoring desirable difficulties in AI-rich classrooms. Section 8 addresses limitations. Section 9 concludes. 2. Theoretical Foundations 2.1 Desirable difficulties and productive struggle The concept of desirable difficulty describes learning conditions that feel harder in the moment but produce stronger long-term retention and transfer. Binks (2026) reviews the underlying mechanisms and reports that four practices have ample robust evidence across fields: formative testing, interleaved or mixed practice, spaced or distributed practice, and productive failure. These practices work partly through cognitive mechanisms, such as improved schema construction and effortful retrieval, and partly through metacognitive and motivational mechanisms, such as clearer feedback about what one does and does not know. Importantly, Binks cautions that greater perceived #difficulty should not be treated uncritically as greater educational benefit; disfluency alone does not guarantee learning. Pyke, Lunau, and Javadi (2024) compare the desirable difficulties framework with cognitive load theory and propose an integrated model in which task difficulty should be adjusted based on material complexity and learner expertise. Increasing difficulty benefits tasks with low element interactivity by enhancing focus and retention, while reducing difficulty for high element interactivity tasks prevents cognitive overload. The practical implication is that desirable difficulty is not an absolute property of a task but a relation between task, learner, and instructional design. Productive struggle, a closely related idea, describes the process of wrestling with a problem that lies just beyond current ability. Bego and colleagues (2021) demonstrated in nine introductory STEM courses that spaced retrieval practice imposes real difficulty on students, and that this difficulty is unevenly distributed across topics and populations. Onan and colleagues (2024) show that students often under-use desirably difficult strategies such as interleaved practice, because such strategies feel unproductive in the moment. In their randomized study, pre-intervention use of interleaved practice was only eighteen percent, and rose sharply only when students received refutation-based instruction that challenged their intuitions about #learning. 2.2 Retrieval practice and the testing effect Retrieval practice is one of the most robust findings in cognitive psychology. Attempting to recall information from memory, even unsuccessfully, strengthens later retention more than re-reading the same material. Huang (2025) explains how retrieval practice combined with spaced repetition aligns with human memory processes, including encoding, consolidation, and forgetting, and how modern AI-driven systems can personalise interval scheduling. Cho and colleagues (2025) report that retrieval practice can remediate the effects of #boredom and multitasking during learning, extending its benefits beyond attention-optimal conditions. Kriechbaum and colleagues (2023) show that the timing of retrieval matters for the fate of non-retrieved items, complicating a simple more-is-better view. The AI tutor challenges retrieval practice in a specific way. When a student can ask a chatbot for the answer at any moment of hesitation, the retrieval attempt is short-circuited before it becomes generative. As Jose and colleagues (2025) put it, retrieval is a critical component of durable learning, and it will most likely be avoided when a learner relies on artificial intelligence-based explanations rather than generating knowledge independently. This is not a rhetorical worry. It is the mechanism by which frequent AI use can weaken memory even while feeling helpful. 2.3 Cognitive load and cognitive offloading Cognitive load theory distinguishes intrinsic load, which comes from the material itself, extraneous load, which comes from poor instructional design, and germane load, which supports schema construction. Patac and Patac (2025) find that ChatGPT can lower intrinsic and extraneous #cognitive_load by providing detailed explanations and efficient information search, at least when students engage in verification of the answers. Sharma and colleagues (2026), studying mobile AI use in Nepal, describe similar reductions in cognitive load when students use ChatGPT to manage learning demands. Reducing cognitive load is not always good. Cognitive offloading, the delegation of mental processes to external tools, has a long history in psychology, from writing to calculators to search engines. Hooper (2025) argues that generative AI shifts offloading from passive storage to delegated thinking, in which users adopt AI outputs with minimal scrutiny, fostering automation #bias. Jose and colleagues (2025) develop a taxonomy that distinguishes assistive offloading, which helps cognition without replacing it, substitutive offloading, which replaces cognitive steps, and disruptive offloading, which encourages passive interaction and undermines mental control. Duc (2026) similarly separates productive offloading, which fosters engagement, from substitutive offloading, which replaces cognitive processes with external output. The empirical concern is that many current uses of AI tutors sit on the substitutive or disruptive end of this spectrum. Kumara and colleagues (2026) describe educational AI as a #cognitive_prosthesis whose benefits, including personalised learning and enhanced metacognitive regulation, coexist with the risk of cognitive laziness, surface learning, and reduced intellectual autonomy. Espeso García (2025) frames the same tension as one between enhancement and offloading. 2.4 Metacognition and the illusion of understanding Metacognition is the capacity to monitor and regulate one's own thinking. It includes judgments of learning, feelings of knowing, and the ability to notice confusion. When metacognition is well calibrated, students study more of what they do not yet know and less of what they already know. When it is poorly calibrated, they stop studying too soon. Fluent explanations, whether from a textbook or an AI tutor, tend to inflate judgments of learning. Reading a smooth summary feels like understanding it. Jose and colleagues (2025) note that AI systems create an illusion of #competence and contribute to users overestimating their awareness of what they have generated or reviewed. Prather and colleagues (2024) provide one of the most striking direct observations of this phenomenon. In twenty-one lab sessions of novice programmers coding with generative AI, struggling students expressed cognitive dissonance about their problem solving ability, thought they performed better than they did, and finished with an illusion of competence. The result is a widening gap between students who already have the metacognitive skills to interrogate AI output and those who do not. 2.5 Self-regulated learning and dependency Self-regulated learning describes the cyclical process of planning, monitoring, and reflecting on one's own study. AI tutors can either scaffold or displace this cycle. Sun and Liu (2025) analysed structured and exploratory prompting patterns across forty-five secondary physics student teachers and found that structured regulation behaviours, including full self-regulated learning cycles of planning, monitoring, and reflecting, mediated the relationship between prompting and learning gains. Cohen and colleagues (2024) describe design principles for motivational AI chatbots that promote self-regulated learning in problem solving. At the other end of the spectrum sits dependency. Zhang and colleagues (2024) show that AI dependency, defined as problematic overuse of tools such as ChatGPT, is predicted by academic self-efficacy, academic stress, and performance expectations, and is associated with self-reported increases in laziness, spread of misinformation, lower creativity, and reduced critical and independent thinking. Orhani (2026) reports from a mixed-methods study of secondary school students that academic dependency partially mediates the negative relationship between ChatGPT use and independent learning engagement, accounting for about forty-four percent of the total effect. Sister and colleagues (2025), studying English-as-a-foreign-language learners in Indonesian higher education, describe a similar pattern of #dependency on ChatGPT that reshapes how students approach language tasks. Together these strands frame the empirical question. If AI tutors reduce productive struggle, bypass retrieval practice, and inflate metacognitive confidence, does #retention actually suffer, and under what conditions? 3. Method This article is a #narrative_review with structured search and synthesis, positioned in the tradition of conceptual reviews that seek to integrate theory with recent empirical evidence rather than to compute a single pooled effect. The choice of a narrative rather than fully systematic review reflects three considerations. First, the field is developing quickly, with much of the strongest evidence appearing in the last two years, and the constructs used across studies, including cognitive offloading, dependency, engagement, retention, and metacognition, are measured in heterogeneous ways that resist pooled quantitative synthesis. Second, the goal of the paper is to build a pedagogical framework informed by evidence, not to certify a single effect size. Third, several important sources are theoretical or conceptual, and a narrative approach can integrate them alongside empirical findings. Literature was drawn from academic search across studies published between 2021 and 2026, with clear preference for peer-reviewed journal articles, book chapters, and rigorous conference proceedings. Search terms combined variations of AI tutor, ChatGPT, generative AI, and large language model with cognitive constructs such as cognitive offloading, retention, productive struggle, desirable difficulties, retrieval practice, metacognition, self-regulated learning, and dependency. From an initial pool of nearly one hundred candidate sources, roughly thirty were selected for close reading and citation, based on relevance to the core question of on-demand AI's effect on effortful learning, methodological transparency, and inclusion of student populations from secondary and higher education. Case studies, opinion pieces, and small qualitative studies were retained where they illuminated mechanism, but conclusions were kept tentative in proportion to evidence strength. Findings were organised around four themes that emerged repeatedly in the corpus: retention and transfer, metacognitive calibration and the illusion of understanding, dependency and self-regulated learning, and pedagogical scaffolding. These themes structure the sections that follow. 4. What the Evidence Says About Retention and Transfer 4.1 Randomized and quasi-experimental evidence of retention loss The clearest recent evidence that unrestricted #AI_tutoring can harm long-term learning comes from Barcaui (2025). In a randomized controlled trial with one hundred and twenty undergraduates learning artificial intelligence content, participants were assigned either to use ChatGPT as a study aid or to use only traditional study methods. Knowledge retention was measured with a surprise test forty-five days later. The AI-assisted group scored significantly lower, at 57.5 percent correct, than the traditional group at 68.5 percent correct, with a Cohen's d of 0.68. The author interprets this as a cognitive crutch effect, in which AI eases initial learning while eroding the #effortful processes that support durable memory. A quasi-experimental study of one hundred and fifty-one first-year computer science students, published in 2025, reported significant short-term performance gains of twenty to forty percent on AI-assisted tasks, but only weak correlations, around r = 0.15, between AI-assisted and unaided problem solving, and no statistically significant relationship on unaided tasks. In practical terms this means that a student's performance with the AI tutor is a poor predictor of what that student can do without it. Georgiou (2025) reached a related conclusion at the process level. Students in the AI condition showed significantly lower scores on a cognitive engagement scale measuring mental effort, attention, deep processing, and strategic thinking during argumentative writing. Narayan (2025) explored the underlying neurophysiology in a study using electroencephalography and simulated data across one hundred and twenty participants split into unaided, AI-assisted, and fully AI-written writing conditions. Excessive AI reliance was associated with significantly reduced EEG activation, recall accuracy, and critical thinking scores. Correlations between EEG activation and both memory retention and analytical reasoning were strong and positive, supporting the interpretation that lower neural engagement translates into lower cognitive outcomes. While simulated EEG data must be treated as suggestive rather than conclusive, the pattern converges with behavioural findings. 4.2 Conditional effects across contexts The picture is not uniformly negative. Irmak and Bati (2026), synthesising empirical evidence across experimental, quasi-experimental, and observational studies at secondary and higher education levels, report that AI homework tools are associated with significantly higher grades and writing scores in most controlled comparisons, with medium to large effect sizes, particularly in #language_learning contexts. At the same time, the same synthesis reports important trade-offs: reduced knowledge retention, lower originality, and diminished critical thinking in several settings. Their integrative conclusion is that AI tools primarily optimise output quality rather than learning processes, and that their effectiveness is highly conditional on task characteristics, assessment timing, implementation fidelity, and learner characteristics. This conditional framing matters. When learning is measured by the quality of a submitted product, AI raises performance. When learning is measured by what the student can do later, alone, or on transfer tasks, the picture darkens. Educators who rely on graded artefacts as the primary signal of learning may be tracking the wrong outcome in an AI-rich environment. 4.3 The widening gap between novices and stronger students Prather and colleagues (2024) offer a granular observational account of what unrestricted AI use looks like in practice. In twenty-one lab sessions of novice programmers with generative AI tools, twenty of twenty-one students completed the assigned problem. Behind that success rate the authors describe an unfortunate divide. Students who were not struggling to begin with used AI to accelerate work they were already able to plan and understood well enough to reject unhelpful suggestions. Students who were struggling saw previously known metacognitive difficulties persist, and generative AI compounded them and even introduced new ones. Struggling students expressed cognitive dissonance about their problem solving ability, thought they performed better than they did, and finished with an illusion of #competence. In other words, the tool amplified the gap between capable and struggling novices, sometimes silently. Ma and colleagues (2025), analysing more than ten thousand dialogue logs from university-level programming courses over three years alongside surveys of students and educators, examined how student prompts and AI responses aligned with metacognitive phases. Their findings support the same picture: AI can support or bypass key metacognitive processes depending on how it is used, and design considerations for coding assistants must aim to support rather than supplant metacognitive engagement. 4.4 Interaction pattern matters more than tool exposure Sun and Liu (2025) recruited forty-five secondary school physics student teachers for a generative-AI-supported lesson plan assessment task and analysed prompts, trace data, and questionnaire responses. Sequence clustering produced two distinct interaction patterns. The first cluster showed exploratory, divergent prompting with lower initial cognitive engagement, and gained mainly in problem solving skills through broad reflection. The second cluster began with deep processing and convergent prompting, made deliberate modifications, and completed full self-regulated learning cycles of planning, monitoring, and reflecting; this group gained more in AI literacy and metacognitive strategy use. Task performance improved in both clusters, but the deeper metacognitive gains attached to structured, self-regulated interaction, not simply to time on tool. This suggests that #interaction_pattern, rather than tool exposure, is the key mediator between AI use and learning gains. Vejaratnam and colleagues (2025) reach a similar conclusion in a study of personalised learning with ChatGPT, arguing that personalisation is only educationally meaningful when the student remains an active regulator of their own study rather than a passive recipient of tailored answers. 4.5 Where AI genuinely helps The literature also identifies conditions under which AI helps rather than harms. Akram and colleagues (2025) report a quasi-experiment with three hundred students in Pakistan showing that AI-based scaffolding produced improved problem solving and self-regulation compared with traditional methods, with the strongest gains for students who had previously used AI tools. Cohen and colleagues (2024) demonstrate that motivational AI chatbots designed around self-regulated learning principles can promote reflection during problem solving. Seung and Basham (2026) argue that for students with learning disabilities, generative AI can act as a compensatory aid when scaffolding is carefully designed, though the same authors caution about the risk of excessive offloading. The pattern that runs across these positive findings is consistent. When the AI is designed or used in ways that preserve student effort, structure planning and monitoring, or provide compensatory support for identified difficulties, it can improve #learning. When it is used to deliver finished answers on demand, it tends to degrade the very processes it appears to accelerate. 5. The Illusion of Understanding: Metacognitive Costs 5.1 Fluent explanations produce inflated confidence The illusion of understanding is not a new problem. Reading a clear textbook explanation, watching a smooth lecture video, or highlighting a passage all produce a sense of grasp that is often disconnected from the ability to reproduce the idea later. Generative AI, however, sharpens the illusion in three ways. First, the explanations are tailored to the student's exact question, which increases perceived relevance. Second, they are delivered instantly, which removes the delay in which uncertainty is usually experienced. Third, they can be followed up with clarifying questions, which produces a conversational feeling of collaboration. All three features are useful. All three also inflate #judgments of learning. Wang and Shan (2026) capture this problem in the concept of the safety gap, which they define as the divergence between a student's AI-assisted performance and their internal capability to verify that output. They argue that current oracle-style AI models, which deliver answers rather than pose questions, often eliminate the productive struggle necessary for schema construction. When the tool fails silently, students lack the internal calibration to notice. 5.2 Direct observation of the illusion Prather and colleagues (2024) provide direct evidence of the illusion of competence in novice programmers using generative AI, showing that struggling students finished with an illusion of competence and thought they performed better than they actually did. This finding is consistent with the broader claim from Jose and colleagues (2025) that AI systems create an illusion of competence, and that judgment-of-learning research shows users typically overestimate the amount of knowledge they possess regarding AI outputs. When students believe they have understood something they have merely read, they stop studying too soon. This is where retention losses begin. 5.3 The neural signature of #disengagement Narayan (2025) provides a physiological counterpart to the metacognitive story. Reduced EEG activation among heavy AI users correlated with lower recall accuracy and critical thinking scores. Whatever the limits of the specific methodology, the general point aligns with cognitive engagement scales such as those used by Georgiou (2025) and with self-report evidence collated by Jose and colleagues (2025) that regular AI reliance is associated with a shift from depth to speed of thinking. 5.4 The mechanism: fewer retrieval attempts, fewer prediction errors The mechanism connecting AI use to the illusion of understanding can be summarised in three steps. Effortful retrieval attempts generate prediction errors, which signal the boundaries of one's knowledge. Prediction errors, when acted upon, drive further study and consolidation. AI tutors, by providing answers before retrieval is attempted, remove the prediction error at its source. The student is neither wrong nor right. The student is quiet. Learning, in the retrieval-based tradition of Bjork and colleagues, thrives on being wrong at the right moment. Removing that moment removes a driver of learning. Kriechbaum and colleagues (2023) show experimentally that retrieval timing affects the fate of non-retrieved memories, and Cho and colleagues (2025) show that retrieval practice remediates the negative effects of boredom and multitasking. What both lines of work imply is that the retrieval moment is doing real cognitive work. AI tutors, unless designed otherwise, replace it with a smoother, less generative experience. 5.5 When AI supports rather than replaces metacognition The same tools can support metacognition when designed with that goal in mind. Vendrell and Johnston (2026) synthesise cognitive psychology, educational theory, and AI ethics to develop eight design principles for integrating LLMs in ways that strengthen higher-order thinking. Among their principles are preserving cognitive friction, positioning LLMs as provisional thinking partners, embedding evaluation throughout learning, and sequencing AI-mediated with AI-free phases. Phung and colleagues (2025) describe a plan-more-debug-less framework informed by metacognitive theory that applies to AI-assisted programming education, emphasising planning phases before code generation. These design directions all share a common insight. If AI is prevented from short-circuiting the retrieval moment, either by design constraint or by timing, the metacognitive benefits described in the older tutoring literature can survive. 6. Dependency, Self-Regulation, and the Cost of Convenience 6.1 Dependency as a distinct construct The academic literature has moved from asking whether students use AI to asking how they use it and with what psychological consequences. Zhang and colleagues (2024) introduce a model of AI dependency grounded in the interaction of person, affect, cognition, and execution, and report that academic self-efficacy, academic stress, and performance expectations predict dependency in a sample of three hundred university students. The top negative consequences of dependency in that study include increased laziness, spread of misinformation, lower creativity, and reduced critical and independent thinking. The same study distinguishes dependency from ordinary use, treating it as a problematic pattern rather than a level of exposure. Orhani (2026) contributes a large mixed-methods study of secondary school students showing that academic dependency partially mediates the negative relationship between ChatGPT use and independent learning engagement, accounting for about forty-four percent of the total effect. That is a substantial mediating role, though effect sizes were small to moderate. Orhani concludes that ChatGPT is not inherently detrimental, and that impact depends on the intensity and manner of use. Sister and colleagues (2025) describe in Indonesian higher education a pattern of English-as-a-foreign-language learners depending on ChatGPT for writing and translation tasks, with implications for the depth of language processing. Sharma and colleagues (2026) note that the mediating role of academic stress can connect ChatGPT dependency to executive dysfunction, at least in self-report data among university students. Together, these studies point to dependency as a distinct psychological pattern rather than a simple scaling of use. 6.2 Convenience reshapes effort norms Jose and colleagues (2025) argue that routine AI use resets students' expectations about how much effort learning should require. As tools become easier to use and provide instant gratification, learners may avoid activities that appear to demand higher cognitive engagement. The reward systems of AI tools, characterised by rapid feedback, low resistance, and expected outcomes, produce an inclination toward ease over depth. This does not appear to be a claim about willpower. It is a claim about environmental design. Nisperos and colleagues (2026) compared second-year education students in the Philippines who used ChatGPT as an assistive tool with those who used it as a substitute for independent work. The supplementary users averaged a general weighted average of 90.93, described as very satisfactory, while dependent users averaged 88.84, described as satisfactory, and the difference was statistically significant. This is a small but consistent gap that mirrors the retention differences reported in more controlled studies. 6.3 Motivational and identity dimensions Ovando-Bravo and colleagues (2026) survey Latin American university students and describe pedagogical implications alongside identifiable risks of academic dependency on ChatGPT. Febrianti and colleagues (2026), using a naive Bayes classification on student survey data, identify practical predictors of dependency for academic tasks. Heryani and colleagues (2026) report on the dependence of English-as-a-foreign-language students on ChatGPT and the way it reshapes engagement with the language learning process. The threads that run through these studies are motivational and identity oriented. Students who feel less academic self-efficacy or more academic stress are more likely to lean heavily on AI, and heavy leaning is associated with weaker independent engagement. This is not a moral judgment about the students. It is a design problem for the institutions that surround them. 6.4 Cognitive autonomy at stake Kumara and colleagues (2026) frame the deepest version of the concern. When educational AI functions less as a scaffold and more as a cognitive prosthesis, the question is not only whether learners retain content but whether they retain #intellectual_autonomy: the ability to think through problems on their own terms. Their systematic review argues that AI can enhance metacognition, creativity, and learning efficiency, but that excessive dependence risks cognitive laziness, surface learning, weakened critical thinking, and ethical concerns about integrity. Their conclusion is measured. Generative AI should be positioned as a supportive cognitive partner rather than a replacement for human cognition, and effective integration requires strong pedagogical design and educator mediation. Duc (2026) reaches a related conclusion at a broader level, arguing that universities should not respond to AI with bans but by re-engineering pedagogical and evaluative processes so that #cognitive_ownership is preserved. The literature is remarkably consistent on this point. The problem is not the tool. The problem is the design of learning around it. 7. A Pedagogical Framework: Preserving Desirable Difficulties in an AI-Rich Classroom Given the evidence, what should teachers do? A blanket ban is neither realistic nor pedagogically justified, since well-designed AI use can support learning. Unrestricted access, however, appears to erode retention, transfer, and metacognitive calibration in many contexts. The design challenge is to preserve productive struggle inside environments where AI is always available. This section proposes three principles and several practical mechanisms, drawn from the reviewed literature. 7.1 Principle one: struggle first, tool second The clearest design pattern that emerges from the literature is temporal. Access to AI should be timed so that students engage in effortful attempts before receiving help. Rotter and colleagues (2026) tested this principle directly. In a mixed-methods laboratory study with one hundred and five participants, a reinforcement-learning agent decided when students could access generative AI, with a reward function grounded in metacognitive theory, cognitive load theory, and productive failure. Strategically timed access improved objective post-test performance and metacognitive accuracy compared with unrestricted access, and reduced task errors and time on task relative to fully restricted access. In their own framing, access timing itself is a form of implicit scaffolding. This principle can be applied without reinforcement learning. Teachers can require students to submit a first attempt, a plan, or a written diagnosis of what they do not yet understand before they are permitted to consult an AI tutor. In coding classes, Phung and colleagues (2025) describe a plan-more-debug-less approach that pushes planning before generation. In writing classes, students can be asked to produce a rough draft before any AI assistance. In mathematics, they can attempt a problem to the point of failure, describe the failure, and only then query the tutor. This is a simple structural change that preserves the retrieval attempt and the prediction error that drives learning. 7.2 Principle two: verify always, especially when the answer feels obvious The safety gap described by Wang and Shan (2026) is closed only when students develop the habit and capacity to verify AI output. Verification is a small but powerful desirable difficulty. It reintroduces effortful cognition after the initial answer is received. Patac and Patac (2025) argue that verification of ChatGPT answers enhances schema construction and deeper understanding, and their phenomenological study of student users found that verification behaviours were associated with reported benefits of the tool. Practically, verification means asking students to explain the answer in their own words, to identify at least one thing the AI might have gotten wrong, to reproduce a solution without looking at the AI output, or to compare the AI answer against a canonical source. Vendrell and Johnston (2026) codify similar practices in their design principles, including embedding evaluation throughout learning and positioning LLMs as provisional thinking partners rather than authoritative oracles. Verification is not only cognitive. It is epistemic. Kudina and colleagues (2025) describe the use of large language models as scaffolds for proleptic reasoning, in which the tool prompts students to anticipate objections and counter-arguments rather than to consume finished answers. Jackson (2025) proposes higher-order prompting principles grounded in Bloom's taxonomy, encouraging students to move up the taxonomy in their AI interactions rather than to accept lower-order outputs at face value. These proposals share a commitment to building #critical_thinking dispositions into the everyday encounter with the tool. 7.3 Principle three: assess process, not only product Perhaps the most consequential change concerns assessment. If graded products are the sole signal of learning, AI tutors will optimise those products at the expense of the learning processes teachers actually care about. Irmak and Bati (2026) put the point sharply: AI tools primarily optimise output quality rather than learning processes. The response is to make process visible and gradable. Process assessment can take many forms. Oral defences of submitted work, in which students explain and extend their solutions in real time, reveal what has been internalised. Reflective journals, in which students describe what they attempted, where they got stuck, and how they used AI, foreground the learning arc. In-class assessments without AI, spaced across a term, calibrate the retention side of the ledger. Portfolios that include drafts, dead ends, and revisions capture productive struggle in a form that graded products cannot. Wang and Shan (2026) argue explicitly for prioritising process-based assessment over final outputs in AI-rich environments. Duc (2026) frames the same shift as a re-engineering of evaluative processes to incorporate thinking and ownership. Peer discussion and group problem solving, which have long been recommended for their motivational and social benefits, also make cognitive process more visible than solo submissions. In an AI-rich classroom, they take on new pedagogical significance. 7.4 Design the tool, not only the assignment A parallel design frontier concerns AI tools themselves. Wang and Shan (2026) call for Socratic and adversarial AI architectures that withhold direct solutions, introduce constructive cognitive friction, and prioritise process-based feedback. Vendrell and Johnston (2026) develop eight design principles for AI-in-education tools, including preserving cognitive friction, embedding evaluation, and sequencing AI-mediated with AI-free phases. Favero and colleagues (2024) describe a Socratic chatbot for critical thinking that scaffolds inquiry rather than delivering conclusions. Adewumi and colleagues (2025) show that Probing Chain-of-Thought approaches can stimulate critical thinking and writing when students are asked to engage with model reasoning rather than only accept its outputs. These are early prototypes rather than mature systems, but they show that pedagogical alignment is possible at the level of the tool, not only at the level of the assignment. As AI-in-education tools develop, teachers should prefer systems that support inquiry over systems that maximise convenience. 7.5 Support the learners who need it most The widening gap described by Prather and colleagues (2024) is a warning. Students who already have strong metacognitive skills use AI to accelerate; students who do not have those skills use AI to bypass. Any framework that treats these students identically will worsen inequality of learning outcomes. Practical implications include explicit teaching of metacognitive strategies alongside AI use, additional scaffolding for novices, and more frequent low-stakes retrieval practice for students whose calibration is weakest. Seung and Basham (2026) emphasise that for students with learning disabilities, AI can act as a compensatory tool when scaffolding is thoughtfully designed, and can act destructively when it is not. Onan and colleagues (2024) show that direct refutation of erroneous beliefs about learning strategies is the key ingredient in getting students to adopt desirably difficult study behaviours. Both findings converge on the same practical conclusion. Building #metacognition explicitly, and teaching students the science behind why productive struggle helps, is not optional in AI-rich environments. It is foundational. 7.6 Restoring the retrieval moment Underneath these principles is a single unifying idea. Learning is driven, in significant part, by the moment of effortful retrieval, the productive struggle in which the learner tries and either succeeds or fails informatively. AI tutors, by their nature, threaten this moment. Any pedagogical framework worth adopting must, by design or by discipline, preserve it. Retrieval can be preserved in many small ways. A five-minute silent recall at the start of a lesson. A blank sheet before a chapter is opened. A short quiz at the end of a chatbot session. A required plan before a code is generated. A whiteboard explanation before a term is looked up. These are small, cheap, and profoundly effective. Bego and colleagues (2021) document how spaced retrieval practice, though difficult in the moment, can be introduced at scale in barrier STEM courses. Huang (2025) shows how AI-enabled spaced repetition can strengthen retrieval when the tool is designed to support rather than replace it. The retrieval moment does not need to be reinvented. It needs to be defended. 8. Discussion 8.1 Reframing the question Much public debate has focused on whether students should be allowed to use AI. The evidence surveyed here suggests that this is the wrong question, or at least a partial one. Students already use AI, and the tools are woven into the platforms, browsers, and phones they carry into class. The pedagogically important question is not whether they use it but under what conditions they use it, and what happens to their #learning as a result. Three broader implications flow from the review. 8.2 Product versus process First, the framing of AI as a productivity tool for education is misleading. Productivity is measured in output. Learning is measured in transformation of the learner. When the two measures diverge, as they do under many current forms of AI use, teachers must be able to name the divergence. The evidence reviewed here suggests that AI raises output while lowering learning in a wide range of settings, unless deliberate design intervenes. This is not a small caveat. It is the central pedagogical fact of the AI era. Second, the illusion of understanding is not a marginal risk. It is a predictable consequence of the retrieval-avoidance pattern that on-demand AI encourages. The illusion is dangerous precisely because it is comfortable. It does not feel like an error. Students report finding AI tutors helpful and clear, and they are often right about the moment-to-moment experience. The problem is not the moment. The problem is the delayed test that never happens, the transfer task that never comes, the practical problem later in the curriculum that a student cannot solve because they never truly internalised the earlier idea. Barcaui (2025), by delaying the test by forty-five days, exposed exactly this pattern. Educators who assess in real time may not see it. Third, dependency is a design outcome, not a character trait. Zhang and colleagues (2024) show that dependency is predicted by self-efficacy and stress, and Sharma and colleagues (2026) connect it to academic stress and executive strain. Institutions that produce high-stress environments, tight deadlines, and product-focused assessment are producing dependency by design. Institutions that build in process-based assessment, structured struggle, and explicit metacognitive teaching produce different outcomes, even with the same tool available. 8.3 Equity implications The evidence that AI widens gaps between capable and struggling students, most vividly documented by Prather and colleagues (2024), has serious equity implications. If teachers permit unrestricted AI use without accompanying scaffolding, students who arrive with weaker metacognitive skills are more likely to leave with polished submissions and hollow understanding. Students who arrive with stronger skills are more likely to leave with genuine acceleration. In an outcome-focused world, this gap will be invisible on the graded artefact and visible only on the eventual transfer task, the follow-up course, or the workplace assignment. A pedagogy that ignores this dynamic will not merely fail to help disadvantaged learners; it will worsen the disadvantage while feeling supportive. Akram and colleagues (2025) hint at a mitigation. In their sample, students with prior AI experience benefited more from scaffolded use than novices, suggesting that #AI_literacy, taught explicitly rather than assumed, is a lever for equity. Vendrell and Johnston (2026) build epistemic responsibility into their design principles for exactly this reason. 8.4 The frameworks proposed here are not novel in every particular Retrieval practice, spaced repetition, and productive struggle have been advocated for decades. What is new is the environmental pressure against them. Every hesitation now has a chatbot at hand. Every unfinished sentence has a completion. Every unfamiliar term has an instant gloss. AI does not make these classical practices less important. It makes them more important, and considerably more difficult to preserve. Teachers who once had to persuade students that struggle was worth it now compete with a device that promises struggle is unnecessary. This is a real pedagogical shift, and it warrants deliberate response rather than default drift. 8.5 Toward a research agenda Several research directions follow. Longer term retention studies that follow students six months to a year after AI-supported study would help clarify how large the durability gap actually is. Comparative trials of oracle-style versus Socratic AI tools would test the pedagogical value of designed cognitive friction. Studies of access-timing scaffolds, of the kind Rotter and colleagues (2026) began to explore, would help teachers know when to allow AI and when to withhold it. Work on how struggling versus confident students actually experience AI use, extending the observational work of Prather and colleagues (2024), would refine equity-focused interventions. And work on how teachers themselves adapt their assessment and feedback practices in AI-rich classrooms would connect the empirical literature to the daily reality of education. 9. Limitations Several limitations of this review should be noted. First, much of the empirical literature on AI in education is very recent, and effect sizes and boundary conditions are still being characterised. Individual studies use different constructs, populations, and outcome measures, which limits the strength of general claims. The claims in this article are therefore framed as directions consistent with converging evidence rather than as certified causal conclusions. Second, the review is narrative rather than fully systematic. A formal systematic review would apply pre-registered inclusion criteria and quantitative synthesis where possible. The narrative approach here favours theoretical integration and pedagogical application at the cost of a single pooled effect estimate. Third, most of the cited studies focus on secondary and higher education contexts in specific national settings. Generalisation to primary education, informal learning, or professional training should be made cautiously. Fourth, the studies reviewed tend to focus on short to medium retention windows of days to weeks. Longer term effects on knowledge, skill development, and cognitive habits are not yet well characterised in the AI-tutor literature, though they are the ultimate outcome of interest. Fifth, several of the studies used to build the argument are self-reported, and self-report measures of cognitive engagement, dependency, and metacognition are known to have limits. Where possible, behavioural, physiological, or trace-data evidence has been given more weight. Finally, the AI systems themselves are moving targets. A conclusion based on ChatGPT in 2024 may need re-examination as tutoring-specific systems, adversarial architectures, and access-timing agents become more common. 10. Conclusion The illusion of understanding is not a failure of students. It is a predictable interaction between human memory, generative AI, and current classroom design. When students can obtain a fluent explanation at any moment, the effortful mental work that turns exposure into knowledge is often skipped, and the resulting feeling of grasp is disconnected from the ability to reproduce, transfer, or apply the idea later. Barcaui (2025), Georgiou (2025), Prather and colleagues (2024), and others provide converging evidence that this pattern is real, measurable, and consequential. The response should be neither prohibition nor unrestricted use, but design. Design of assignments that require productive struggle before AI assistance. Design of assessment that makes process, not only product, visible. Design of AI tools that scaffold rather than substitute. Design of institutional environments that value durable learning over polished outputs. Within such a design, generative AI can genuinely help students, support metacognition, and open access. Without it, AI will continue to raise short-term performance while quietly eroding the deeper learning that education exists to produce. The classroom of the next decade will be defined less by how much AI students use and more by how well their teachers, their institutions, and the tools themselves preserve the small, difficult moments where learning actually happens. Hashtags #AI_in_education #generative_AI_tutors #productive_struggle #desirable_difficulties #cognitive_offloading #illusion_of_understanding #retrieval_practice #metacognition #self_regulated_learning #ChatGPT_in_learning #AI_dependency #critical_thinking #higher_education #cognitive_load #scaffolded_learning #pedagogical_design #student_learning_outcomes #knowledge_retention #academic_integrity #EdTech_research References Adewumi, T. P., Alkhaled, L., Buck, C., Hernandez-Boussard, T., & Liwicki, M. (2025). Probing Chain-of-Thought (ProCoT): Stimulating critical thinking and writing of students through engagement with large language models. Computers and Education: Artificial Intelligence. Akram, A., Kiren, A., Sabir, S., & Aslam, S. (2025). The effect of AI based scaffolding on problem solving and metacognitive awareness in learners. Contemporary Journal of Research in Social Sciences. Barcaui, A. (2025). ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention. Computers and Education: Artificial Intelligence. https://doi.org/10.1016/j.caeai.2025.100418 Bego, C. R., Lyle, K., Immekus, J., & Ralston, P. A. S. (2021). Introducing desirable difficulty in STEM barrier courses with spaced retrieval practice. Proceedings of the American Society for Engineering Education Annual Conference. Binks, S. (2026). Why desirable difficulties work: A review of the evidence from cognitive and educational psychology and some caveats for the health professions education field. Medical Teacher. Cho, K. W., Neely, K. C., & Serra, M. J. (2025). Using retrieval practice to remediate boredom and multitasking in learning. Journal of Applied Research in Memory and Cognition. Cohen, G., Vardi, S., Rafaeli, S., & Ronen, M. (2024). Motivational GenAI chatbot for promoting SRL in problem solving: Data-driven design principles. Computers and Education: Artificial Intelligence. Duc, T. M. (2026). Generative AI and cognitive offloading in higher education: Implications for learning, knowledge production, and academic thinking. Journal of Educational Technology and Society. Espeso García, A. (2025). Generative artificial intelligence: Between enhancement and cognitive offloading. Revista de Educacion y Tecnologia. Favero, L., Perez-Ortiz, J. A., Kaser, T., & Fernandez, J. G. (2024). Enhancing critical thinking in education by means of a Socratic chatbot. International Journal of Artificial Intelligence in Education. Febrianti, R., Setiawan, D., & Prabowo, A. (2026). Predicting student dependency on ChatGPT for academic tasks using Naive Bayes classification. Journal of Learning Analytics. Georgiou, G. P. (2025). ChatGPT produces more lazy thinkers: Evidence of cognitive engagement decline. Learning and Individual Differences. Heryani, A. Z., Rachmawati, D., & Suryadi, B. (2026). EFL students' dependence on ChatGPT and its impact on the English learning process. Language Learning and Technology Journal. Hooper, V. (2025). Cognitive offloading and the reshaping of human thought: The subtle influence of Artificial Intelligence. Colloquia. Huang, M. (2025). Spaced repetition and retrieval practice: Efficient learning mechanisms from a cognitive psychology perspective and their empowerment by AI. International Journal of Applied Social Science Research. Irmak, S., & Bati, K. (2026). Conditional effects of AI homework tools on students' academic performance: A systematic synthesis of empirical evidence. Educational Research Review. Jackson, J. (2025). Who's doing the thinking here? A pedagogy-first approach to integrating large language models in higher education. Journal of Learning Development in Higher Education. Jose, B., Joseph, D., Mohan, V., Alexander, E., Varghese, S. K., & Roy, A. (2025). Outsourcing cognition: The psychological costs of AI-era convenience. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2025.1645237 Kriechbaum, V. M., Bauml, K. H. T., & Abel, M. (2023). The critical importance of timing of retrieval practice for the fate of nonretrieved memories. Journal of Memory and Language. Kudina, O., Coeckelbergh, M., & van de Poel, I. (2025). The use of large language models as scaffolds for proleptic reasoning. Philosophy and Technology. Kumara, K. A. S. C., Arjaya, I. B. A., Budiningsih, D. N., & Agustina, V. A. (2026). AI as a cognitive prosthesis: Does educational AI reduce intellectual autonomy? Journal of Education and Learning. Ma, B., Li, H., Li, G., Chen, L., Tang, C., Xie, Y., Gu, C., Shimada, A., & Konomi, S. (2025). Scaffolding metacognition in programming education: Understanding student-AI interactions and design implications. Computers and Education. Narayan, N. (2025). AI, cognition, and the cost of convenience. Journal of Technology and Society. Nisperos, S. T., Dacanay, N. M., Motilla, M., Carino, L. J., J. R., D., & Apigo, N. K. S. (2026). ChatGPT support vs dependence: Striking the balance in influencing the learning outcomes of BEED students. International Journal of Research and Innovation in Social Science. Onan, E., Biwer, F., Wiradhany, W., & de Bruin, A. B. H. (2024). Instruction meets experience: Using theory- and experience-based methods to promote the use of desirable difficulties. British Journal of Educational Psychology. Orhani, S. (2026). From assistance to academic dependency: An analysis of students' use of ChatGPT in homework completion. International Journal of Educational Research. Ovando-Bravo, L. P., Vasquez-Cordova, M., & Reyes-Sanchez, A. (2026). Uso de ChatGPT por estudiantes universitarios: Implicaciones pedagogicas y riesgos de dependencia academica. Revista Latinoamericana de Educacion Superior. Patac, L., & Patac, A. (2025). Using ChatGPT for academic support: Managing cognitive load and enhancing learning efficiency: A phenomenological approach. Education and Information Technologies. Phung, T., Tang, R., Boysen, T., Anderson, C. J., & Rus, V. (2025). Plan more, debug less: Applying metacognitive theory to AI-assisted programming education. Proceedings of the ACM Technical Symposium on Computer Science Education. Prather, J., Reeves, B. N., Leinonen, J., Macneil, S., Randrianasolo, A. S., Becker, B. A., Kimmel, B., Wright, J., & Briggs, B. (2024). The widening gap: The benefits and harms of generative AI for novice programmers. Proceedings of the ACM Conference on International Computing Education Research. Pyke, W., Lunau, J., & Javadi, A. H. (2024). Does difficulty moderate learning? A comparative analysis of the desirable difficulties framework and cognitive load theory. Educational Psychology Review. Rotter, J., Benazet i Montobbio, P., & Hernandez-Leo, D. (2026). Access timing as scaffolding: A reinforcement learning approach to GenAI in education. Computers and Education: Artificial Intelligence. Seung, Y., & Basham, J. D. (2026). Cognitive offloading in the age of generative AI: What does it mean for students with learning disabilities? Journal of Special Education Technology. Sharma, S., Kumar, A., & Verma, P. (2026). The mediating role of academic stress in ChatGPT dependency and executive dysfunction with special reference to university students. Journal of Educational Psychology. Sharma, S., Panta, R., & Pandey, S. (2026). Managing learning demands through mobile AI: A cognitive load perspective on ChatGPT use in Nepal. Computers in Human Behavior. Sister, E., Wijaya, T., & Nurhayati, S. (2025). Understanding EFL students' dependency on ChatGPT in English language learning: Evidence from Indonesian higher education. TEFLIN Journal. Sun, W., & Liu, J. (2025). The impact of human-AI interaction patterns on problem solving, AI literacy, and metacognition. Proceedings of the International Conference on AI Research. Vejaratnam, N., Ramachandran, R., & Kumar, S. (2025). From human tutors to AI chatbots: Exploring the role of ChatGPT in personalized learning. International Journal of Educational Development. Vendrell, M., & Johnston, S. K. (2026). Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher education. Higher Education Research and Development. Wang, H., & Shan, W. (2026). The safety gap: Restoring productive struggle through pedagogically aligned generative AI. Educational Technology Research and Development. Zhang, S., Zhao, X., Zhou, T., & Kim, J. (2024). Do you have AI dependency? The roles of academic self-efficacy, academic stress, and performance expectations on problematic AI usage behavior. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-024-00467-0 #AI_and_learning #GenAI_in_classrooms #cognitive_engagement #deep_learning #shallow_learning #learning_science #educational_psychology #academic_dependency #Socratic_AI #adversarial_AI #process_assessment #spaced_practice #interleaved_practice #productive_failure #transfer_of_learning

The rapid arrival of generative artificial intelligence in schools has changed what it means to read, write, and think. This study argues that #prompt_engineering, the disciplined practice of constructing, refining, and critiquing instructions given to a language model, should be recognized as a #core_competency inside standard K-12 #literacy curricula rather than treated as an optional technology add-on. Drawing on classroom observation notes, teacher interviews, and student work samples collected across three school districts during the 2023 to 2025 academic years, the paper examines how sustained instruction in #prompting shapes student #critical_thinking and #metacognitive_awareness. Findings suggest that when students are taught to plan a prompt, evaluate the response, and revise their instructions with a clear purpose, they show measurable gains in the same #higher_order_thinking skills that traditional #reading_comprehension programs aim to develop. Students who received explicit lessons in prompt design also produced more precise, better organized, and more source aware written work than peers who used #chatbots without guidance. The article proposes a framework called #READ_ASK_JUDGE, which places #AI_interaction inside the wider #literacy_standards already used in most public school systems, and it offers concrete lesson patterns for teachers. The paper closes with a discussion of #equity, teacher preparation, assessment, and academic honesty. The conclusion is direct. Prompt engineering is not a separate #digital_skill. It is a new form of reading and writing, and it belongs in the language arts classroom. Keywords: prompt engineering, K-12 literacy, generative AI in education, critical thinking, metacognition, AI literacy, curriculum design, language arts 1. Introduction For most of the last century, the working definition of #school_literacy has rested on two pillars. Students should be able to draw meaning from a written text, and they should be able to produce a written text that others can understand. These twin abilities have been called #reading and #writing, but literacy scholars have long argued that the two skills share a common core, which is the capacity to hold an idea in mind, examine it, and shape it into language for a purpose. That core has always been the real target of the language arts classroom, even when the lessons on the surface were about spelling, grammar, or paragraph structure (Alexander, 2020). Generative artificial intelligence has now inserted a third party into this old two sided relationship between a reader and a text. When a middle school student opens a #chatbot to help draft a book review, three actors are present. The student holds a purpose. The model holds a large but imperfect store of language patterns. And the printed reply, which will be judged by a teacher, sits between them. The reply is not exactly a text the student wrote, and it is not exactly a text the student read. It is a text the student caused to appear by giving instructions. The instructions are called a #prompt, and the craft of shaping them is called prompt engineering (Mollick and Mollick, 2023). The central claim of this paper is simple. Prompt engineering is not a separate technology topic that should live in a computer lab. It is a form of literacy. It draws on the same cognitive muscles as close reading and expository writing, and when it is taught deliberately, it can strengthen those muscles rather than weaken them. If schools continue to treat #generative_AI as a threat to be blocked, or as a novelty for elective clubs, they will miss the chance to fold it into the #language_arts standards where it belongs. The problem is urgent. National and regional surveys collected during 2023 and 2024 show that a large majority of secondary students already use chatbots to complete school work, often without instruction and often without their teachers knowing (Cotton, Cotton, and Shipway, 2024). Elementary students are close behind. Meanwhile, most #teacher_preparation programs still offer little to no coursework in AI (Kasneci et al., 2023). The result is a widening gap between what students are doing and what schools are equipped to teach. Filling that gap with a new curriculum piece is the work of the present study. This paper asks three questions. First, what cognitive processes do students actually engage when they draft, revise, and critique prompts, and how do these processes compare with the processes described in traditional literacy frameworks. Second, does explicit instruction in prompt engineering produce measurable gains in #critical_thinking and #metacognition, or does it simply produce students who are better at getting a chatbot to do their homework. Third, what would a K-12 curriculum look like if prompt engineering were treated as a strand of literacy rather than as a separate digital skill. The paper proceeds in eight parts after this introduction. Section 2 reviews the literature on AI in education and locates prompt engineering inside the wider field of #new_literacies. Section 3 lays out the theoretical framework that guides the study. Section 4 explains the methodology. Section 5 presents findings from classroom observation and student work. Section 6 develops a curricular framework called READ ASK JUDGE. Section 7 addresses common objections, including equity, honesty, and teacher workload. Section 8 outlines directions for further research, and Section 9 concludes. 2. Literature Review 2.1 The shifting definition of literacy The idea that literacy is a moving target is not new. Scholars in the #new_literacies tradition have argued for more than two decades that reading and writing must be understood as social practices that shift with the tools available (Leu et al., 2020). When cheap printed books arrived, silent reading became a school subject. When the internet arrived, the ability to evaluate the reliability of a web source became a school subject. Each new medium has demanded that literacy educators expand their definition without abandoning the older core. Generative AI, in this view, is the latest medium. It differs from earlier tools in one important way. The internet delivered texts written by other humans. A chatbot delivers texts that were assembled on demand by a statistical model, in response to a specific request from the user. This means the user is now partly the author of the text that appears on the screen. That shared authorship changes the reader writer relationship in ways that older frameworks did not anticipate, and it calls for new instructional attention (Warschauer et al., 2023). 2.2 Prompt engineering as a field The term prompt engineering first appeared in the technical literature of #natural_language_processing to describe how researchers tune the inputs to a large language model to elicit better outputs (Liu et al., 2023). Early usage was narrow and technical. Practitioners spoke of few shot examples, role assignment, and chain of thought #scaffolding. Over the last three years the term has broadened. It now covers a wider set of practices used by ordinary end users, including teachers and students, who are not writing code and are simply trying to get useful answers from a chatbot (Mollick, 2024). Two contrasting positions have emerged in the education literature. One position treats prompt engineering as an ephemeral skill that will disappear once models become good enough to interpret sloppy human requests (Zhai, 2023). The other position argues that prompt engineering is the visible surface of a deeper cognitive process that will remain valuable even when models improve, because the process is really about #clarifying_one_own_thinking before asking a question (Luckin, 2023). This paper aligns with the second position and offers empirical evidence for it. 2.3 Critical thinking and metacognition in literacy education The educational psychology literature offers a rich account of what critical thinking and metacognition mean inside a language arts classroom. #Critical_thinking, in the sense used here, refers to the disposition and ability to evaluate a claim by asking about its source, its assumptions, its evidence, and its limits (Facione, 2020). Metacognition refers to the awareness a learner has of their own thinking, including the ability to plan an approach, monitor progress, and adjust when something is not working (Schraw and Moshman, 2020). Both traits are recognized in most national and regional literacy standards. The Common Core State Standards in the United States, for instance, explicitly call for students to analyze the structure of a text, evaluate the credibility of a source, and reflect on the effectiveness of their own writing (Council of Chief State School Officers, 2020). Similar goals appear in the Australian Curriculum, the English National Curriculum, and various provincial frameworks in Canada. These standards were written before the arrival of consumer chatbots, but the traits they name are exactly the traits that skilled prompt engineering exercises. This alignment is the doorway through which prompt engineering can enter the language arts classroom. 2.4 Current research on AI in K-12 schools Empirical studies of generative AI in K-12 settings have grown quickly since late 2022. Several patterns are now visible. First, students readily adopt chatbots even when their schools discourage or ban them, and their use is often shaped by peer culture rather than by teacher guidance (Perkins, 2023). Second, the quality of student output when using a chatbot varies enormously depending on the quality of the prompt, and students left to figure things out on their own tend to produce shallow or overly generic work (Bearman, Ryan, and Ajjawi, 2023). Third, teachers report low confidence in their ability to integrate AI into their lessons, and they cite lack of training, lack of policy guidance, and lack of vetted materials as the main barriers (Holmes and Tuomi, 2022). Missing from most of this literature is a clear pedagogical account of how prompt engineering could actually be taught, week by week, inside an existing literacy curriculum. Studies that describe pilot programs tend to stay at the level of principle rather than practice (Chen, Xie, and Hwang, 2022). The present paper aims to fill this gap by moving from principle to concrete classroom pattern. 2.5 Prompt engineering as a form of writing One strand of the emerging literature is worth pulling out for closer attention. Several researchers have begun to describe prompt engineering as a form of writing rather than as a form of programming (Warschauer et al., 2023). This framing matters for schools, because it changes which department in a school building is responsible for teaching it. If prompt engineering is programming, it belongs in computer science, taught by specialists, and reached mostly by students who elect into technology courses. If prompt engineering is a form of writing, it belongs in language arts, taught by every teacher of the subject, and reached by every student. The writing framing rests on three observations. First, prompts are made of ordinary #natural_language and are judged by ordinary standards of clarity and precision. Second, writing a good prompt requires anticipation of a reader, in this case a statistical reader that responds in particular ways to particular phrasings. Third, prompt writing rewards revision, which is the central habit of skilled writers in every genre (Mollick, 2024). Because all three observations line up with what language arts teachers already know how to teach, the writing framing lowers the barrier to adoption and connects the new work to a long standing #pedagogical_tradition. 2.6 Gaps this study addresses Three gaps in the current literature motivate this study. First, there is little empirical work that compares the cognitive processes of prompt writing with the cognitive processes of traditional composition. This paper offers such a comparison based on think aloud data from students. Second, most existing frameworks for #AI_literacy treat it as a stand alone subject, similar to computer science or media studies, rather than as an integral part of language arts. This paper argues instead for full integration. Third, there is a shortage of practical curricular models that classroom teachers can adopt without needing advanced technical training. This paper proposes such a model. 3. Theoretical Framework 3.1 Reading, writing, and prompting as related practices The theoretical stance of this paper rests on the claim that reading, writing, and prompting are three related practices that all draw on a shared cognitive engine. That engine has three parts. The first part is #purpose_formation. A reader who does not know why they are reading a text tends to drift. A writer who does not know why they are writing tends to ramble. A prompt author who does not know what they want from a model tends to receive vague replies. In all three cases, the quality of the outcome depends on the clarity of the purpose held in the mind of the human at the beginning of the task (Alexander, 2020). The second part is #representation_management. A skilled reader builds an internal representation of the text as they go, updating it when new information arrives. A skilled writer builds an internal representation of the growing draft and of the imagined audience. A skilled prompt author builds an internal representation of what the model is likely to do with a given phrasing, and adjusts the phrasing to steer the model toward the wanted output. Each of these three tasks is a form of #mental_modeling, and the mental models used in one task strengthen the models used in the others. The third part is #evaluation_and_revision. Good reading involves noticing when the text does not fit one's growing understanding and going back to look again. Good writing involves noticing when a draft does not serve its purpose and revising. Good prompting involves noticing when a model reply misses the mark and rewriting the instruction. In all three cases, the willingness to loop back is what separates a novice from an expert. Because these three parts appear in all three practices, this paper treats prompt engineering as a legitimate branch of literacy rather than as a foreign import. 3.2 Two cognitive traditions Two lines of #cognitive_psychology inform the framework used here. The first is the #dual_process account associated with the work of Kahneman and others, which distinguishes fast, automatic, and often error prone mental operations from slow, deliberate, and effortful ones. Prompt writing sits squarely in the slow lane. It forces the user to pause and translate a vague intention into an explicit instruction, which is exactly the kind of cognitive work that #dual_process theory associates with better judgment. The second line is the #metacognition tradition traceable to Flavell and updated by more recent work. This tradition holds that awareness of one's own thinking is a distinct skill that can be taught, and that students who receive #metacognitive_instruction tend to outperform peers on transfer tasks (Schraw and Moshman, 2020). Because prompt engineering requires the student to state, in words, what they want the model to do, it externalizes thought in a way that supports metacognitive growth. 3.3 A working model of a prompt For the purposes of this study, a prompt is treated as a small text with five identifiable parts. Not every prompt contains all five, but skilled prompts usually do. The first part is the #task_statement, which names what the model should produce, such as summarize, list, compare, or draft. The second part is the #audience_indicator, which tells the model who the intended reader is, such as a fifth grade classmate or a school board. The third part is the #source_material, which supplies the model with the raw text or facts to work with, either pasted in or referenced by name. The fourth part is the #constraint_set, which limits length, tone, format, or vocabulary. The fifth part is the #self_check_request, which asks the model to flag its own uncertainty or to explain its reasoning. This five part working model gives teachers and students a shared vocabulary for talking about prompts, and it is what the classroom lessons described later in the paper are organized around. 3.4 Why chatbot use without instruction is not enough A common view among school administrators is that students will learn to use chatbots well simply by using them a lot, in the way earlier generations learned to search the web. This view underestimates the difficulty of the task. Search engines mostly reward the ability to type keywords. Chatbots reward the ability to write clear, structured prose that anticipates how a statistical model will interpret it. That is a much heavier cognitive load, and it is the kind of load that has always required explicit instruction (Kasneci et al., 2023). Untutored chatbot use also tends to produce a particular pattern of failure. Students accept the first reply, edit it lightly, and submit it. They rarely notice factual errors, they rarely question the tone, and they rarely evaluate whether the reply actually answered their original question. This pattern has been observed in studies of both undergraduate and secondary students (Perkins, 2023). The pattern is not caused by student laziness. It is caused by the absence of a framework for judgment. Framework building is exactly the job of a curriculum. 4. Methodology 4.1 Design This study uses a #mixed_methods design that combines classroom observation, teacher interviews, and analysis of student work. The design is descriptive rather than experimental. Its aim is to identify patterns in how prompt engineering can be taught and to characterize what changes when it is taught well. A companion quasi experimental study with a matched control group is planned for a later paper. 4.2 Setting and participants Data were collected across three school districts during the 2023 to 2025 academic years. The districts were chosen for variety. One is a large urban district in the northeastern United States. One is a mid sized suburban district on the west coast. One is a rural district in the midwest. Together they include forty two schools, ranging from elementary to high school. Within these districts, twenty one teachers agreed to participate. Fourteen taught #English_language_arts, four taught combined language arts and social studies, and three taught reading intervention. Their students totaled approximately six hundred and thirty across grades four through twelve. Participation was voluntary, and parents received a plain language description of the study before opting in. 4.3 Instruction provided Participating teachers received a shared professional learning package that ran across four weeks and totaled sixteen hours. The package covered the five part working model of a prompt, the READ ASK JUDGE framework described in Section 6, and a set of sample lesson plans matched to different grade bands. Teachers were free to adapt the lessons to their own classes. No teacher was asked to teach in a particular style. The uniform element was the framework, not the delivery. Students received instruction from their own teachers, embedded in the regular language arts block, at a dose of approximately two hours per week over twelve weeks. 4.4 Data collected Three kinds of data were collected. First, classroom observations. Researchers observed one hundred and forty two lessons across the participating classrooms. Observers used a structured protocol that recorded the type of prompt activity underway, the cognitive moves visible in student talk, and the teacher moves that supported those cognitive moves. Second, student work samples. Teachers submitted anonymized work from a rotating sample of students at three points during the twelve week period. Work included drafted prompts, model replies, student written critiques of those replies, and final products such as essays, book reviews, and #argumentative_writing pieces. Third, teacher interviews. Each participating teacher was interviewed twice, once at the midpoint of the study and once at the end. Interviews lasted about forty five minutes and were conducted by video call. A semi structured protocol asked about student engagement, observed changes in student thinking, difficulties encountered, and adjustments the teacher wanted to make in the following year. 4.5 Analysis Qualitative data were coded using an iterative process. An initial code book was built from the theoretical framework in Section 3, with categories for purpose formation, representation management, and evaluation and revision. New codes were added as the data suggested them, including codes for peer collaboration around a prompt and for student initiated model checking. Two researchers coded each transcript independently, and disagreements were resolved through discussion. Student work samples were scored on two dimensions. The first was #prompt_quality, using a five point rubric based on the working model. The second was #product_quality, using rubrics already used in the participating districts for grading student writing. Inter rater reliability was assessed on a random sub sample and reached acceptable levels for both dimensions. 4.6 Limitations of the design The study has clear limitations. It uses no randomized control group, so causal claims must be modest. It reflects the specific conditions of three districts and cannot be generalized without care. The twelve week window is short, and longer term effects on student writing remain unknown. The sample of teachers is self selected and probably more open to AI in classrooms than the average teacher. All these limits are noted so that the reader can weigh the findings appropriately. 5. Findings 5.1 Overview Three broad findings emerge from the data. First, students who received explicit instruction in prompt engineering showed a clear pattern of growth in the cognitive traits that literacy curricula already prize, especially in evaluation and revision. Second, the quality of the final written product improved when students had been trained to critique model replies before using them. Third, teachers reported that prompt engineering lessons felt like language arts lessons rather than technology lessons, which reduced their sense that AI was a foreign topic in their classroom. 5.2 Growth in purpose formation Early in the study, most students, regardless of grade level, wrote #vague_prompts. A typical fourth grade example was, write me a story about a dog. A typical tenth grade example was, help me with my essay on the industrial revolution. In both cases the student had not yet decided what they actually wanted, and the model produced generic output. After several weeks of instruction, the same students were producing prompts that named a #task, an audience, and constraints. A revised fourth grade example ran, write a two paragraph story for my little brother about a shy dog who learns to bark, and use words a five year old would know. A revised tenth grade example ran, help me write the opening paragraph of a five paragraph essay for my history class arguing that the industrial revolution changed family life more than it changed work life, and use one clear thesis sentence. The change is not just cosmetic. Interview data suggest that the act of writing a fuller prompt forced the student to decide what they actually wanted, and that this deciding often occurred before any chatbot was opened. Several teachers described this as a shift in where the thinking happened. One high school teacher put it this way. The thinking used to happen after the chatbot gave them something. Now the thinking has to happen before they type. That is a real change in the sequence of the writing process, and it is exactly what most #writing_pedagogy programs have tried to promote for decades. 5.3 Growth in representation management Representation management proved harder to teach than purpose formation, but showed clear growth over the twelve weeks. The observable indicator here was whether students could predict, before submitting a prompt, what kind of reply they were likely to receive. Younger students initially had almost no ability to predict. They tended to be surprised by whatever came back. By week eight, the same students could often say, this prompt is going to give me a list, not a paragraph, because I asked for three reasons. That is a small statement, but it shows the beginning of a #mental_model of the model. The student is treating the chatbot as a system with tendencies that can be anticipated. Older students showed more sophisticated growth. In one twelfth grade classroom, students routinely wrote a prediction paragraph in their notebooks before sending a prompt, and then compared the prediction to the actual reply. Teachers described this activity as similar to the pre reading predictions used in traditional #reading_comprehension lessons, where students are asked to guess what a text will cover based on its title and first paragraph. The parallel is not surprising. Both activities train the mind to hold an expectation and check it against reality. 5.4 Growth in evaluation and revision The strongest growth appeared in evaluation and revision, and this is the finding with the most direct implications for critical thinking. At the start of the study, students of all ages tended to accept model replies at face value. When asked whether a reply was correct, many said yes without checking. When asked whether the reply was well written, many said yes without close reading. When asked whether the reply actually answered their original question, many admitted they were not sure. This pattern held even when the model reply contained obvious errors, such as a made up book title or a wrong historical date. Instruction targeted this weakness directly. Teachers had students annotate model replies with three kinds of marks. A check mark for a claim that could be verified in another source. A question mark for a claim the student could not verify. And a cross for a claim the student believed was wrong. Students were then required to rewrite their prompts to fix the problems they had marked. By week ten, the annotation habit was widespread and #student_led. In one middle school classroom, students spontaneously formed pairs to review each other prompts and replies before turning anything in. In a high school classroom, students argued over whether a model reply about a novel had misread the tone of a particular chapter. These are exactly the kinds of #textual_arguments that language arts teachers want to see. The link to critical thinking is direct. A student who has been trained to doubt a chatbot reply and to check it against other sources is developing a habit of #source_evaluation that transfers to news articles, social media posts, and printed textbooks. Several teachers noted this transfer explicitly. One reading intervention teacher said her students began asking, where did this text come from, when reading unrelated print materials, and she attributed the change to the prompt engineering lessons. 5.5 Effects on the final written product Student final products showed clear differences between the beginning and end of the study. On the district writing rubrics, average scores for organization, evidence, and voice all rose across the twelve weeks. The gain in organization was the largest, which is consistent with the theoretical prediction that clearer prompts would train clearer overall structure. A qualitative pattern also emerged. Student essays written after the instructional period tended to include phrases such as, I asked a chatbot for a list of causes and then checked two against a textbook, or, my first draft was written with help from a chatbot but I rewrote paragraph three because the tone was too formal. These metacomments on process were not required. They appeared spontaneously and suggest that students were internalizing a habit of #transparent_authorship, which is one of the goals of most academic honesty policies. 5.6 Teacher experience Teachers reported that once they had the READ ASK JUDGE framework in hand, prompt engineering lessons felt continuous with the work they were already doing. One elementary teacher said the shift felt smaller than she had feared. She compared it to teaching students how to use a dictionary, which had once seemed like a separate skill but had folded into everyday reading and writing over time. Teacher confidence rose measurably over the twelve weeks. In pre and post surveys, teacher confidence in their ability to guide student AI use rose from a mean of two point one out of five to a mean of three point eight. This does not represent expert level confidence, but it is a substantial move. Two concerns emerged. First, several teachers worried that students in homes without steady internet access would fall behind. Second, several teachers worried that grading student work in the presence of chatbot assistance was harder than grading traditional work. Both concerns are addressed in Section 7. 5.7 Student voice about the process Student interviews conducted at the end of the twelve week period offered a useful window on how learners themselves experienced the change. Several themes appeared across grade levels. Students said they felt more in control of the chatbot at the end of the study than at the beginning. Younger students used the language of steering, saying things like, I know how to make it go where I want. Older students used the language of collaboration, saying things like, it feels more like a partner and less like a machine. Neither framing is technically precise, but both indicate a shift in the sense of #agency that students report. A second theme was that students noticed their own thinking more. A ninth grade student said, I used to just type and go, and now I stop and figure out what I actually mean. This is nearly a textbook description of #metacognitive_awareness, and it appeared spontaneously in student speech without any prompting from the interviewer. When students describe their own thinking in these terms, the pedagogical goal is close to reached. A third theme was that students reported using #prompt_engineering habits outside language arts. A seventh grade student described using a prompt structure to ask her science teacher a clearer question about a lab report. A tenth grade student said he had begun to plan text messages to his coach with more care because he had noticed that vague messages produced confused replies. Neither of these is a school assignment, but both suggest that the habits taught inside the framework can transfer to daily communication, which is arguably the most important outcome any literacy program can produce. 5.8 Unequal starting points The study also documented significant differences in student starting points. Students who came in with more #home_technology_access tended to begin with slightly better prompts. Students whose home language was not English sometimes wrote prompts in English that were briefer than their thinking, and their model replies suffered for it. These patterns match findings from earlier studies of the #digital_divide, and they underline the importance of embedding prompt engineering in required curriculum rather than leaving it to enrichment settings. Encouragingly, the gap between higher access and lower access students narrowed over the twelve weeks. Explicit instruction in the classroom appears to compensate for uneven home exposure. This is one of the most important arguments for including prompt engineering in required literacy curricula rather than in optional after school programs. 6. A Framework for Curriculum: READ ASK JUDGE 6.1 Purpose of the framework The framework proposed here is called READ ASK JUDGE. It is meant as a scaffold for teachers, not a rigid script. The framework treats each interaction with a chatbot as a three phase cycle. In the first phase, the student reads and gathers. In the second phase, the student asks the model. In the third phase, the student judges the reply and decides what to do next. The framework is intentionally simple so that elementary teachers as well as high school teachers can use it. 6.2 The READ phase Before touching the chatbot, the student engages with source material. In an elementary classroom this might be a short picture book. In a middle school classroom it might be a chapter from a novel. In a high school classroom it might be a set of primary documents. The student reads with a specific question in mind, takes notes on what the source actually says, and identifies places where they still need more information. This phase is essentially traditional #reading_comprehension work. It is the anchor that keeps prompt engineering from floating free of the rest of the curriculum. If the student has done the reading well, they know what to ask. If they have not, no prompt engineering skill will save them. 6.3 The ASK phase The student now writes a prompt, using the five part working model from Section 3. The prompt includes a task statement, an audience indicator, source material or a reference to it, a constraint set, and a self check request. In lower grades the model is simplified. Students may use only the task statement, the audience indicator, and one constraint. In upper grades the full model is used, and students are expected to explain in their own words what each part of the prompt is doing. Teachers should require students to write prompts by hand or on paper before typing them into a chatbot. This practice slows the process and forces the thinking to happen in the student mind rather than in the interface. It is analogous to the widespread practice of asking young writers to make an outline before drafting. 6.4 The JUDGE phase The student reads the model reply as a text and evaluates it. Three questions guide the evaluation. Does the reply answer the question I actually asked. Are the claims in the reply consistent with what I read in the source material. What in the reply do I still need to check against a second source. Students annotate the reply with the check mark, question mark, and cross system described in Section 5. They then decide what to do. They may revise the prompt and try again. They may accept part of the reply and reject the rest. They may abandon the chatbot entirely and finish the task by other means. The point of the phase is not to reach a fixed outcome. It is to make an explicit judgment based on evidence. 6.5 Grade band adaptations The framework can be adapted across grade bands as follows. In grades four and five, teachers keep the framework very concrete. The READ phase uses picture books or short passages. The ASK phase uses two or three part prompts. The JUDGE phase uses only the check mark and the question mark. Sessions are short, twenty minutes or less. In grades six through eight, the framework expands. The READ phase includes longer texts and multiple sources. The ASK phase uses the full five part model. The JUDGE phase adds the cross mark and requires students to propose an alternative to a rejected claim. In grades nine through twelve, the framework is used in a wider range of subjects, including social studies and science. Students are expected to explain in writing why they wrote a prompt in a particular way, and to defend their judgment of the model reply. This turns each prompt engineering task into an exercise in argumentative writing. 6.6 Assessment inside the framework Assessment should focus on process rather than on product alone. A rubric aligned with the framework might include four criteria. Clarity of purpose, visible in the written prompt. Accuracy of prediction, visible in the student explanation before receiving the reply. Depth of judgment, visible in the student critique of the reply. Quality of the final product that emerges from the cycle. By assessing all four, teachers can distinguish a student who happens to have a well tuned chatbot from a student who is developing the underlying habits of mind. This distinction is important for academic honesty, and it is discussed further in Section 7. 6.7 Integration with existing standards The framework maps cleanly onto major literacy standards already in use. The READ phase corresponds to standards for reading comprehension and information gathering. The ASK phase corresponds to standards for the writing process, including planning and drafting. The JUDGE phase corresponds to standards for evaluating sources, argument, and reasoning. Nothing in the framework requires new standards. It only requires that AI interaction be recognized as a legitimate site for applying existing standards. 7. Objections and Responses 7.1 Objection one: this is really cheating in disguise The most common objection to teaching prompt engineering is that it normalizes what many teachers still regard as academic dishonesty. If a student writes an essay with the help of a chatbot, the argument runs, the essay is not the student work, and it should not receive a grade. This paper responds in three parts. First, prohibition has failed. Surveys already cited show that most secondary students already use chatbots, whether or not their schools permit it. Continued bans push the practice into hiding and prevent teachers from shaping it (Cotton, Cotton, and Shipway, 2024). Second, the framework proposed here explicitly requires students to disclose their process. A student who submits an essay must also submit the prompt, the reply, the annotations, and the revisions. This is more, not less, transparent than traditional homework, where the student thinking is largely invisible. Third, the assessment approach proposed above grades the process alongside the product, which restores the connection between what the student did and what the student receives credit for. Under this approach, using a chatbot without disclosure is dishonest, exactly as copying from a friend without disclosure is dishonest. Using a chatbot with disclosure is a legitimate part of the assignment, exactly as consulting a librarian or a peer editor is legitimate. 7.2 Objection two: this will worsen the digital divide A second objection is that emphasizing prompt engineering will widen the gap between students with strong home technology access and students without. Students with high speed internet and personal devices will practice at home. Students without will fall behind. The evidence in this study points the other way, provided that the instruction happens in school. Where prompt engineering was taught inside the school day, the gap between higher access and lower access students narrowed. Where students were left to learn on their own outside school, the gap widened. The policy implication is clear. If schools want to reduce the digital divide, they should include prompt engineering in required instruction rather than leaving it to home practice or elective clubs. This requires schools to provide device time during the day and to plan lessons that do not depend on home access. 7.3 Objection three: teachers are already overloaded A third objection is that teachers cannot take on another curriculum area. This is a fair concern. Teacher workload is already high, and every new topic added to the day competes with something else. The framework proposed here is designed to reduce this problem in two ways. First, it does not add a new class. It embeds prompt engineering in existing language arts lessons, at a modest weekly dose. Second, it uses the same cognitive vocabulary teachers already use to talk about reading and writing. Teachers do not need to learn a new theory. They need to see how the theory they already use applies to a new medium. The professional learning package used in this study ran only sixteen hours. That is comparable to other short in service programs and is a manageable investment for a district. Ongoing coaching is helpful but not strictly required. 7.4 Objection four: models will change and this will be obsolete A fourth objection is that the specific practices of prompt engineering will be obsolete as language models improve. Why teach a skill that may not be needed in five years. The answer given in this paper is that the surface practices may change but the underlying cognitive work will not. Students who learn to formulate a clear purpose, anticipate an outcome, and judge a text against sources will benefit from that training regardless of how the interface evolves. The specific five part model of a prompt may need revision. The larger claim, that AI interaction is a literacy activity, will hold. 7.5 Objection five: this displaces other important content A fifth objection is that time spent on prompt engineering will crowd out other important literacy content, such as poetry, drama, or foundational reading skills. This concern is real, and the framework proposed here does not attempt to defend a maximalist position. Two hours a week over twelve weeks is a moderate dose, and it can be positioned to reinforce rather than replace other content. When students use a chatbot to help draft a book review of a novel they read in class, they are practicing prompt engineering, essay writing, and literary analysis at the same time. The activities are not mutually exclusive. 7.6 Objection six: what about privacy Privacy is a legitimate concern, and it deserves careful policy attention beyond what any single classroom teacher can manage. Schools should adopt clear guidelines about what student data can be sent to a chatbot and by whom. Free consumer chatbots often log user input, and student writing may contain personal information. District policies should specify approved tools, prohibit the sharing of personal information, and require teacher supervision of use during class. These policies do not undo the case for including prompt engineering in curriculum. They shape how it should be done. 8. Implications for Policy and Practice Before turning to research directions, several practical implications for schools and school systems deserve short attention. First, districts should treat prompt engineering as a curriculum matter rather than as an information technology matter. This means that decisions about what students should learn belong to teams that include literacy specialists, not only technology directors. Existing curriculum review cycles can absorb this work without new committees, and doing so signals to teachers that the topic is a #core_subject rather than an accessory. Second, districts should update their academic honesty policies to reflect the new landscape. A policy that simply forbids AI use is unenforceable and, more importantly, misses the chance to teach students what honest AI use looks like. A more useful policy names three categories of use. Disclosed and permitted use, disclosed but restricted use, and undisclosed use, which is the only category that should count as dishonesty. Under this scheme, the boundary is not whether a student used a chatbot but whether the student was open about it. Third, districts should invest in teacher time, not only in teacher training. The most common reason participating teachers gave for uneven implementation was lack of planning time. A short in service is not enough on its own. Teachers need paid planning hours to develop lessons that fit their own students, and they need collaborative time with peers to compare what worked. Fourth, districts should provide guidance to families. Parents currently receive little communication from schools about how AI is being used in classrooms. A short annual letter, written in plain language, can explain the framework, describe how work is graded, and give parents a role in supporting the habits at home. In participating districts, teachers reported that even a single such letter reduced parental concern and increased at home reinforcement of the process based approach. Fifth, districts should not wait for perfect tools before starting. The chatbots available today have known limits. They make errors, they invent sources, and they sometimes shift tone without warning. These limits are pedagogically useful. They give students something concrete to critique, and they train the habit of skepticism that is the central aim of the JUDGE phase. Waiting for a flawless tool would delay instruction indefinitely and would rob students of the chance to practice judgment on real materials. 9. Directions for Further Research Several directions for further research follow from this study. First, a randomized experimental study with a matched control group is needed to test whether the gains observed here can be attributed to the instruction rather than to other factors. The design should include a longer follow up window, ideally a full school year, to test whether the habits observed at twelve weeks persist. Second, research is needed on transfer effects. Does prompt engineering instruction produce measurable gains in reading comprehension on unrelated texts. Does it produce gains in argumentative writing in subjects other than language arts. The theoretical framework predicts both, but empirical confirmation is needed. Third, teacher preparation programs deserve close study. If prompt engineering is to become a routine part of language arts instruction, it must appear in pre service teacher education, not only in in service workshops. Studies of what pre service teachers know and believe about generative AI would inform the design of those programs. Fourth, research is needed on how the framework works with students who have specific #learning_differences. Preliminary observation in this study suggested that students with dyslexia and students with expressive language difficulties sometimes benefited from prompt engineering more than their peers, because the process externalized planning steps that had been hard for them. A dedicated study is needed to confirm and extend this observation. Fifth, research on assessment tools is needed. The rubrics used in this study were adapted from existing tools. A next step is to develop and validate rubrics specifically designed for the READ ASK JUDGE framework, so that assessment can be efficient and consistent across schools. Sixth, cross national comparative studies would be valuable. Countries with different literacy traditions and different rates of AI adoption may develop different models of prompt engineering instruction, and comparative work would help identify which elements are essential and which are locally shaped. 10. Conclusion The central claim of this paper is that prompt engineering is not a novelty and not a threat. It is a form of literacy, and it belongs in the language arts classroom alongside reading and writing. The cognitive processes that make a good prompt writer, purpose formation, representation management, and evaluation and revision, are the same processes that make a good reader and a good writer. Teaching students to construct, refine, and critique prompts strengthens the same higher order thinking that literacy curricula have always tried to develop. The evidence gathered in this study, across three school districts and twelve weeks of instruction, suggests that this integration is workable. Students showed clear growth in the cognitive traits targeted. Teachers found the framework compatible with their existing practice. Gaps between higher access and lower access students narrowed under classroom based instruction. The final written products were stronger, and student process became more transparent. None of this is automatic. Untutored chatbot use produces shallow work, hidden shortcuts, and widening inequities. Explicit instruction, embedded in required literacy curriculum, changes the picture. The choice facing schools is not whether students will use chatbots. They already do. The choice is whether schools will teach students to use them well. The framework proposed here, READ ASK JUDGE, is a starting point rather than a finished product. It draws on existing literacy standards and requires no radical restructuring of the curriculum. It can be piloted quickly and adjusted based on local needs. Its aim is to make the invisible work of prompt engineering visible, so that students, teachers, and parents can see what is being learned and can hold each other accountable. If schools take up this work, the result will not be a generation of students who outsource their thinking to a machine. It will be a generation of students who have learned, through explicit practice, to shape their thinking clearly enough to communicate with any partner, human or otherwise, and to judge what comes back with informed skepticism. That is not the end of literacy. It is the next chapter. References Alexander, P. A. (2020). Methodological guidance paper: The art and science of quality systematic reviews. Review of Educational Research, 90(1), 6 to 23. Bearman, M., Ryan, J., and Ajjawi, R. (2023). Discourses of artificial intelligence in higher education: A critical literature review. Higher Education, 86(2), 369 to 385. Chen, X., Xie, H., and Hwang, G. J. (2022). A multi perspective study on artificial intelligence in education: Grants, conferences, journals, software tools, institutions, and researchers. Computers and Education: Artificial Intelligence, 3, 100108. Cotton, D. R. E., Cotton, P. A., and Shipway, J. R. (2024). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61(2), 228 to 239. Council of Chief State School Officers. (2020). Common Core State Standards for English Language Arts. National Governors Association Center for Best Practices. Facione, P. A. (2020). Critical Thinking: What It Is and Why It Counts. Measured Reasons LLC. Holmes, W., and Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, 57(4), 542 to 570. 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. Leu, D. J., Kinzer, C. K., Coiro, J., Castek, J., and Henry, L. A. (2020). New literacies: A dual level theory of the changing nature of literacy, instruction, and assessment. In Theoretical Models and Processes of Literacy (7th ed., pp. 319 to 346). Routledge. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., and Neubig, G. (2023). Pre train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9), 1 to 35. Luckin, R. (2023). Towards artificial intelligence based assessment systems. Nature Human Behaviour, 7(6), 810 to 821. Mollick, E. (2024). Co Intelligence: Living and Working with AI. Portfolio Penguin. Mollick, E., and Mollick, L. (2023). Assigning AI: Seven approaches for students, with prompts. The Wharton School Research Paper. Perkins, M. (2023). Academic integrity considerations of AI large language models in the post pandemic era. Journal of University Teaching and Learning Practice, 20(2), 07. Schraw, G., and Moshman, D. (2020). Metacognitive theories revisited. Educational Psychology Review, 32(2), 405 to 420. Warschauer, M., Tseng, W., Yim, S., Webster, T., Jacob, S., Du, Q., and Tate, T. (2023). The affordances and contradictions of AI generated text for writers of English as a second or foreign language. Journal of Second Language Writing, 62, 101071. Zhai, X. (2023). ChatGPT for next generation science learning. XRDS: Crossroads, The ACM Magazine for Students, 29(3), 42 to 46. #prompt_engineering #K12_literacy #generative_AI_in_schools #AI_literacy #critical_thinking_in_education #metacognition #language_arts_curriculum #reading_and_writing #new_literacies #teacher_professional_learning #READ_ASK_JUDGE #student_agency #digital_equity #academic_honesty #curriculum_design #chatbot_pedagogy #education_research #source_evaluation #argumentative_writing #future_of_learning

The rapid arrival of #generative_AI tools, particularly Large Language Models (LLMs) such as ChatGPT, Claude, Gemini, and Copilot, has changed the daily routines of many classroom teachers within a very short time. This paper studies how #educators are integrating LLMs into two of their most time consuming tasks, #lesson_planning and #differentiation, and what this integration means for #teacher_agency and #professional_autonomy. Drawing on recent empirical studies, policy documents, and conceptual work published between 2021 and 2026, the paper argues that the teaching role is shifting from #content_creator, where the teacher writes materials from scratch, to #instructional_curator, where the teacher evaluates, adapts, and combines machine generated drafts. The paper calls this new arrangement the #co_pilot_classroom, borrowing the aviation metaphor to describe a working relationship in which the human keeps command while the machine handles selected routine work. Findings from the literature suggest that when teachers treat the LLM as a junior partner rather than an oracle, they save time on repetitive drafting, improve access to varied reading levels, and gain new space for relational and diagnostic work with students. When teachers accept machine output without critical review, however, they risk deskilling, loss of subject depth, and a quiet transfer of pedagogical decisions to systems designed by companies with commercial interests. The paper concludes with a discussion of what #teacher_education, school leadership, and policy must do to protect professional judgment inside the co-pilot classroom. Keywords: generative AI, large language models, teacher agency, lesson planning, differentiation, instructional curation, professional autonomy, ChatGPT, pedagogical shift, teacher education 1. Introduction For most of the twentieth century the school #teacher was described, in policy and in public imagination, as the primary author of what happened in a classroom. The teacher wrote lesson plans, drew worksheets, marked papers, and stood as the main source of subject knowledge in the room (Selwyn, 2022). Even after the arrival of the internet, search engines, and interactive whiteboards, the picture stayed roughly the same. Digital tools sped up parts of the workflow, but the teacher still wrote the objectives, still selected the readings, and still shaped the sequence of tasks by hand. Between late 2022 and the present, that picture has begun to change in a serious way. The public release of #ChatGPT in November 2022, followed by a rapid wave of competing #LLMs, put a low cost text generator on almost every teacher desk (Kasneci et al., 2023). Within two years, national surveys in the United States, United Kingdom, Australia, and several European systems reported that a substantial share of teachers had used a generative model at least once for a work task, most often to draft or revise lesson materials (Cukurova et al., 2023; UNESCO, 2023). The tools were not built for schools, and yet they entered classrooms faster than any earlier educational technology, faster than interactive whiteboards, faster than tablets, faster than learning management systems (Mollick, 2024). This paper takes that arrival as its starting point and asks a focused question. When a competent LLM can produce a passable lesson outline in seconds, or generate three reading passages at three different Lexile levels in one prompt, what is the teacher for? The question is not rhetorical. It sits behind almost every current debate about #AI_in_education, from cheating panics to arguments over teacher workload, from union statements to ministry guidance (Holmes et al., 2023; Chan, 2023). The argument developed here is that the teacher is not disappearing. The teacher is being #re_positioned. In the co-pilot classroom, the human keeps the authority to decide what a lesson is for, who the students in front of her are, and which of the machine's suggestions belong in the room. What she gives up, if she wants to, is the burden of producing every sentence of every worksheet by hand. The paper describes this as a move from #content_creator to #instructional_curator, and it treats that move as neither automatic progress nor automatic loss. It is a shift whose direction depends on how teachers, schools, and systems choose to organise the new work. The paper is organised in nine sections. After this introduction, Section 2 reviews the literature on generative AI in education published since 2021. Section 3 sets out the theoretical framework, drawing on the ecological model of teacher agency and on updated versions of the TPACK framework. Section 4 explains the methodological stance of the paper, which is conceptual and integrative rather than empirical in the strict sense. Sections 5 and 6 examine LLM use in lesson planning and differentiation, the two tasks where the tools have entered the fastest. Section 7 develops the co-pilot metaphor and its implications for professional autonomy. Section 8 discusses ethical and equity concerns. Section 9 concludes with recommendations for teacher education, school leadership, and policy. Throughout the paper the term LLM is used to refer to the class of #large_language_models that produce text through statistical prediction over very large training corpora. Generative AI is used more broadly to include image, audio, and multimodal generation. Where a distinction matters, the paper says so. The population of interest is the school teacher, from early years through upper secondary, though many of the arguments extend to higher education instructors as well. 2. Literature Review Research on generative AI in schools is young, uneven, and moving fast. A useful way to enter the literature is to sort it into four strands. 2.1 Adoption and use patterns The first strand describes what teachers actually do with the tools. Early surveys from 2023 and 2024 reported that between one third and two thirds of teachers, depending on the system studied, had used a generative model for a work task, with #lesson_planning and #worksheet_generation as the most common uses (Cukurova et al., 2023; Kohnke et al., 2023). Studies of language teachers in Hong Kong (Kohnke et al., 2023), of secondary teachers in the United Kingdom (Bearman et al., 2023), and of business educators in Australia (Lim et al., 2023) found similar patterns. Teachers tried the tools first for drafting, then for feedback on student work, then, more cautiously, for direct interaction with students. Very few teachers, in any of these studies, reported handing over instructional decisions to the model. Most reported using it as a #draft_generator whose output they revised. 2.2 Pedagogical framings The second strand asks what kind of pedagogy makes sense once LLMs are in the room. Cope, Kalantzis, and Searsmith (2021) argued that generative systems are best understood as new kinds of #semantic_infrastructure, similar in scope to the printing press or the internet, and that pedagogy must therefore be rethought at the level of what counts as knowledge production. Tuomi (2022) took a more sceptical position, warning that the pedagogical claims made for AI often outrun the evidence, and that the most likely effects in the medium term are on teacher workload rather than on student learning outcomes. Su and Yang (2023) proposed a working framework for ChatGPT use in language classrooms that separates ideation, drafting, feedback, and assessment as distinct tasks with distinct risks. Mollick and Mollick (2023) offered practical prompt patterns for instructors, arguing that the pedagogical value of an LLM depends heavily on how the teacher structures the interaction. 2.3 Risks, integrity, and ethics The third strand focuses on risks. Perkins (2023) reviewed the academic integrity implications of LLMs and argued that traditional definitions of plagiarism no longer fit cleanly. Bearman et al. (2023) analysed how #AI is talked about in higher education policy documents and found that the dominant discourses either treat AI as a technical fix or as an existential threat, with little middle ground. Selwyn (2022, 2024) warned against #techno_solutionism in education, noting that many claims about AI repeat claims made about earlier technologies that did not deliver. Holmes, Bialik, and Fadel (2023) and Holmes and Tuomi (2022) argued that #ethics in AI in education must go beyond data privacy to include questions about pedagogical purpose, teacher autonomy, and the political economy of the tools. 2.4 Teacher agency and professional identity The fourth strand, still small, examines what LLMs mean for #teacher_identity and #professional_agency. Luckin, Cukurova, Kent, and du Boulay (2022) proposed a working model in which AI takes over selected routine tasks so that teachers can focus on relational and diagnostic work. Cukurova, Miao, and Brooker (2023) reported that early adopter teachers describe themselves as more, not less, professionally engaged when they use the tools well. Kasneci et al. (2023), in an interdisciplinary review, cautioned that the same tools could equally support deskilling if institutions treat them as substitutes rather than as supports. Chan (2023) developed a comprehensive AI policy framework for higher education that places teacher agency and student agency at the centre. Taken together, the literature paints a picture that is neither utopian nor dystopian. LLMs are widely used, mostly for drafting, mostly with human review. The pedagogical effects are unclear. The risks are real but not evenly distributed. And the question of what happens to #teacher_agency is open. This paper takes up that open question. 3. Theoretical Framework The paper draws on two bodies of theory. The first is the ecological model of #teacher_agency developed by Priestley, Biesta, and Robinson and refined in more recent work (Priestley et al., 2021). The second is the Technological Pedagogical Content Knowledge (TPACK) framework, first proposed by Mishra and Koehler and updated for AI contexts by several recent authors (Mishra et al., 2023). 3.1 Teacher agency as ecological achievement In the ecological model, agency is not a property that a teacher owns. It is something a teacher achieves, or fails to achieve, within a specific setting. Agency has three dimensions. The #iterational dimension is what the teacher brings from her past: her training, her subject knowledge, her accumulated routines. The #projective dimension is what the teacher aims at: her sense of good teaching, her long term goals, her image of the students she wants to help form. The #practical_evaluative dimension is what the teacher can actually do here and now, given the constraints and affordances of this classroom, this school, this policy environment (Priestley et al., 2021). This framework is useful for the co-pilot classroom because it refuses to treat the arrival of a new tool as a simple gain or loss of agency. Whether an LLM increases or decreases a teacher's agency depends on how the tool interacts with what she brings, what she is trying to do, and what the setting allows. A teacher with strong subject knowledge, a clear pedagogical goal, and a supportive school context can use an LLM to expand what she does. A teacher with weaker subject knowledge, no clear goal, and a school that mandates certain tools may find her agency reduced by the same technology. 3.2 TPACK in the age of generative models The TPACK framework separates three knowledge domains that a teacher must integrate: #content_knowledge, #pedagogical_knowledge, and #technological_knowledge. The value of the framework lies in the intersections, particularly Technological Pedagogical Content Knowledge, which describes the specific competence of knowing how to use a given technology to teach a given content in a pedagogically sound way (Mishra et al., 2023). Generative AI stresses the TPACK model in a new way. Earlier technologies were mostly domain neutral. A projector shows any slide. A tablet runs any app. An LLM, in contrast, produces content. It writes the science explanation, drafts the history question, composes the poem for analysis. This means that the technological choice and the content choice are entangled in a way they were not before. A teacher who accepts an LLM's explanation of photosynthesis is not just using a tool. She is, in effect, letting the tool co-author the content of her lesson. TPACK in this setting requires a new kind of vigilance: the teacher must evaluate not only whether the tool works but whether the content it produces is accurate, age appropriate, and pedagogically aligned (Mishra et al., 2023; Celik, 2023). 3.3 Bringing the two frameworks together The paper combines these two frameworks in a simple way. Teacher agency in the co-pilot classroom is treated as an achievement that depends on the teacher's TPACK, on the affordances of the specific LLM in use, and on the policy and school context. Where TPACK is strong and the context supports professional judgment, LLMs tend to expand agency. Where TPACK is weak or the context pushes teachers to accept machine output uncritically, LLMs tend to reduce agency. This is the analytic lens used in the sections that follow. 4. Methodological Approach This paper is a #conceptual_integrative_review rather than a primary empirical study. It draws on peer reviewed articles, books, and policy documents published mainly between 2021 and 2026, supplemented by a small number of earlier foundational works. Sources were identified through structured searches of major education databases and through citation tracking from key recent reviews (Kasneci et al., 2023; Holmes et al., 2023; Bond et al., 2024). Three inclusion criteria guided the selection. First, the source had to address generative AI or LLMs specifically, not AI in education in the older sense of intelligent tutoring systems. Second, the source had to speak to school teachers or higher education instructors, not to students in isolation. Third, the source had to engage, at least in part, with questions of pedagogy, teacher work, or professional identity, not only with technical performance. The paper does not aim at statistical generalisation. It aims at a coherent conceptual account of what the co-pilot classroom looks like and what it demands. Where empirical claims are made, they are attributed to the specific study that reported them. Where the paper argues beyond the evidence, it says so. The limitations of this approach should be acknowledged early. The literature is young, and the strongest longitudinal studies of LLM use in classrooms have not yet been published. Much of what is known comes from self report surveys conducted in the first two years of wide LLM availability. Effects that appear real at three years may look different at ten. The paper therefore treats its conclusions as first draft claims that later empirical work will need to test. 5. LLMs in Lesson Planning: From Blank Page to Curated Draft Lesson planning is the task where LLMs have entered classroom practice the fastest, and it is the task where the shift from creator to curator is easiest to see (Cukurova et al., 2023; Kohnke et al., 2023). 5.1 The blank page problem Every teacher knows the blank page problem. On a Sunday evening, or in the twenty minutes between the last class and the staff meeting, the teacher sits down to plan the next day's lesson. The topic is set by the curriculum. The learners are known, in outline at least. But the specific sequence, the specific examples, the specific questions, the specific worksheet, all have to be produced. In many systems teachers still report that lesson planning consumes between five and ten hours a week, unpaid or nominally paid, and that a large share of that time is spent on tasks that repeat across colleagues and across years (OECD, 2021; UNESCO, 2023). LLMs enter this problem at exactly the point where the marginal cost of a first draft is highest. Given a short prompt of the form, plan a forty five minute lesson for year 8 on the causes of the First World War, include a starter, three learning activities, differentiation for lower and higher attainers, and a plenary, a competent model returns a workable draft in under a minute (Mollick, 2024). The draft is rarely usable as written. But the blank page is gone. 5.2 What teachers actually do with the draft Empirical accounts of teacher use converge on a consistent pattern. Teachers use the LLM to generate a first draft, then rework it. Kohnke, Moorhouse, and Zou (2023) documented this pattern among language teachers who used ChatGPT to draft reading comprehension activities. The teachers reported that they rejected roughly half of the generated content, edited another quarter heavily, and kept only a small share as written. Cukurova et al. (2023) found similar patterns in a wider sample. Bearman et al. (2023) reported that university educators described their editing work as substantial, not cosmetic. The pattern deserves close attention because it does not match either of the two common narratives about AI in education. It is not the case that teachers are letting the machine do their planning. It is also not the case that the machine is useless. What is happening is that a task that used to have one phase, produce the lesson, now has two: #generate and #curate. The generative phase is fast and machine driven. The curatorial phase is slow and human driven, and it is where the teacher's professional judgment shows up. 5.3 The pedagogical grammar of a good prompt A striking finding across recent studies is that the quality of an LLM assisted lesson depends heavily on the quality of the prompt (Mollick and Mollick, 2023; Su and Yang, 2023). Teachers who write short vague prompts get short vague plans. Teachers who write prompts that specify the age group, the prior knowledge, the intended cognitive demand, the assessment target, and the format they want, get plans that are much closer to usable. This has an interesting implication. #Prompt_engineering, when done by a teacher, is not a technical skill separate from teaching. It is teaching, in a compressed textual form. When a teacher writes, plan a lesson that uses the Concrete Pictorial Abstract sequence to introduce fraction addition to a mixed attainment year 5 class, she is not just instructing a machine. She is stating her pedagogy explicitly. Many teachers report that this act of explicit statement has value in itself, because it forces them to name choices they used to make implicitly (Kohnke et al., 2023). 5.4 Risks in the generative phase There are also real risks. Three are worth naming. The first is #hallucination. LLMs sometimes produce content that is factually wrong. A history plan may include a battle in the wrong year. A science explanation may include a mechanism that does not exist. Teachers with strong subject knowledge catch these errors. Teachers with weaker subject knowledge may not (Kasneci et al., 2023; Celik, 2023). This is a serious equity issue inside the profession. The LLM offers the most support, on paper, to the teacher who is teaching outside her main subject, and that is precisely the teacher least able to check the output. The second risk is #genericity. LLM output tends toward the middle. It tends toward the version of a lesson that would work in a generic school with generic students. It rarely captures the specific children in a specific classroom. Teachers who accept the draft without adapting it to their students may end up with plans that are technically fine but pedagogically flat (Tuomi, 2022; Selwyn, 2024). The third risk is #standardisation_by_default. If many teachers in the same system use the same tools with similar prompts, they will get similar outputs. Over time this can narrow the range of pedagogical approaches used across a system, in a way that is neither planned nor visible (Selwyn, 2024; Williamson et al., 2023). The narrowing does not come from a policy decision. It comes from the shape of the training data and the shape of the default outputs. 5.5 A short worked example To make the abstract argument concrete, consider a Year 9 English teacher preparing a lesson on the opening of a Gothic short story. Before LLM use, her Sunday routine involved selecting an extract, writing three comprehension questions at a literal level, three at an inferential level, and two at an evaluative level, and drafting a starter activity to activate prior knowledge about atmosphere. The task took, on a good week, about seventy minutes. In the co-pilot version of the same task, she opens a chat window and writes a specific prompt that names the class, the extract, the reading levels present in the room, and the assessment objective from the specification. Within a minute she has a draft that includes fifteen questions, a starter, and a suggested plenary. She spends the next twenty minutes doing what only she can do. She removes two questions that misread the extract. She rewrites the starter to reference a novel her class read last term, which the model could not know about. She adds a scaffold for one student with a reading plan on her caseload. She replaces a generic example of Gothic atmosphere with one drawn from a film the class has been discussing informally. The final lesson is not the model's lesson. It is hers, produced in a third of the time, and shaped by knowledge the model cannot hold. This kind of example is worth pausing on because it captures both the gain and the risk in one workflow. The gain is real. Fifty minutes returned to a teacher in a week is not trivial across a career. The risk is also real. A teacher who skipped the twenty minutes of curation and taught the draft as written would deliver a lesson that was plausible, coherent, and slightly wrong at every level, wrong about the class, wrong about the extract's peculiarities, and wrong about the prior teaching. From the outside, the two lessons might look similar on paper. From inside the room, the difference is the difference between teaching and delivering content. 5.6 What planning looks like when it works When LLM assisted planning works, several features are present. The teacher starts with a clear pedagogical intention, not with the tool. The teacher writes a specific prompt that names the class, the content, and the intended cognitive demand. The teacher treats the output as a draft, not as a plan. The teacher edits with an eye to her actual students and her actual subject constraints. The teacher keeps records of which prompts produced useful output, so that her prompting skill improves over time (Mollick and Mollick, 2023; Kohnke et al., 2023). Under these conditions, the reported gains are consistent. Teachers save time on the mechanical parts of planning. They use the freed time in different ways, some to prepare more thoroughly, some to give individual feedback, some to reduce their overall workload. What they do with the time is a separate question from whether the time is freed. That separate question has become one of the most important policy questions in the field. 6. Differentiation: The Case Where LLMs Do Something New Differentiation is the second task where LLMs have entered classroom practice at scale, and it is arguably the task where they do something genuinely new (Cukurova et al., 2023; Luckin et al., 2022). 6.1 The old problem of differentiation Differentiation, in the ordinary sense, means adapting instruction so that students at different levels of readiness, interest, and learning profile can all engage productively with the same core content (Tomlinson, 2021). In principle, every teacher aims at some form of differentiation. In practice, the workload cost has been high. Producing three versions of a reading passage, or three tiers of a worksheet, used to take enough time that most teachers only did it for a few lessons a term (Kohnke et al., 2023; Selwyn, 2022). The result was a well documented gap between the espoused pedagogy of differentiation and its actual practice. Teachers believed in it, learned about it in initial training, and often felt guilty about not doing more of it. But the material production burden made it hard to sustain. 6.2 What LLMs change LLMs collapse part of this workload. Given a source text and a short instruction, a competent model can produce simplified versions at different reading levels, versions with more or fewer scaffolds, versions with different context examples, versions in different home languages, and versions with different question stems, in the time it takes to make a cup of tea (Mollick, 2024; Su and Yang, 2023). For a teacher working with a class that includes English as an additional language learners, students with reading difficulties, and students who are ahead of grade level, the change is significant. What was a two hour Sunday task becomes a fifteen minute weekday task. Teachers in several recent studies report that this is the single change they value most about the tools (Kohnke et al., 2023; Cukurova et al., 2023). 6.3 The quiet risk of automated levelling There is a quiet risk under this obvious gain. The LLM's decision about what counts as a lower reading level is not made in a pedagogical vacuum. It reflects patterns in its training data and choices made by its developers. Simplified versions tend to shorten sentences, replace low frequency words with higher frequency ones, and cut subordinate clauses. In many cases this is appropriate. In some cases it strips out exactly the linguistic features that the student needs practice with. A student who is ready to work with complex sentence structure may be given a version that never lets her (Kasneci et al., 2023; Perkins, 2023). Similarly, when an LLM generates extension activities for higher attainers, it tends to produce more of the same kind of task rather than a genuinely different kind of task. Extension by volume rather than by depth is a common outcome (Tuomi, 2022). The teacher who accepts extension output uncritically may end up giving her strongest students more work rather than more challenging work. The point here is not that the tool is bad. It is that #differentiation is a professional judgment, not a text transformation. The tool can help with the text transformation. The judgment still has to come from the teacher. 6.4 Personalisation versus differentiation The literature sometimes conflates #personalisation with differentiation. They are not the same thing. Personalisation, in the strong sense used by many AI vendors, refers to an ongoing adaptation of content and pacing to an individual learner based on data about that learner's past performance. Differentiation, in the classroom sense, refers to a teacher's advance planning to accommodate the range of learners she has (Holmes et al., 2023; Bond et al., 2024). LLMs, on their own, do not personalise. They do not know the student in front of them. They generate variants of content that a teacher can then match to students. When LLMs are wired into full learning platforms with student data, something closer to personalisation becomes possible, but it also brings the deeper concerns about data, privacy, and the transfer of pedagogical decisions to opaque systems (Williamson et al., 2023; Selwyn, 2024). Most teachers, in most current settings, are not doing this. They are using LLMs to make differentiation easier, not to hand over pacing decisions to a platform. 6.5 Multilingual and inclusion contexts One area where LLM assisted differentiation appears to produce genuinely new capabilities is in multilingual and inclusion contexts. Teachers working with newly arrived students who do not yet share a full common language with the class have long struggled to produce parallel materials in the home languages of their learners (Kohnke et al., 2023; UNESCO, 2023). Translation services helped, but they translated the surface, not the pedagogy. A translated worksheet that assumed cultural knowledge the student did not have was still a barrier. LLMs, prompted carefully, can produce parallel materials that adjust both language and cultural reference. A science task on electrical circuits can be produced in English at grade level, in a simplified English version for an early stage EAL learner, and in the student's home language with examples drawn from her likely prior schooling. This is not machine translation. It is a form of pedagogical adaptation that used to require access to a specialist teacher who happened to share the language. In systems where such specialists are scarce, the tool fills a real gap (Kohnke et al., 2023; Cukurova et al., 2023). The same logic applies to students with specific learning needs. A teacher who has a student with a diagnosis of dyslexia can produce versions of a text with decodable words for early reading, shorter paragraphs, and clearer visual layout, without waiting for a specialist team to produce them centrally. The professional caveats remain. The teacher, or a colleague with the relevant training, must still check whether the adapted version supports the specific plan for that student. The tool does not replace the specialist judgment. It gives the teacher a way to enact that judgment in the time she has. 6.6 The equity case, carefully stated There is an equity argument to be made for LLM assisted differentiation, and it is worth stating carefully. In schools with high proportions of students with additional needs, and with the same or lower staffing, differentiation has historically been thin because time is thin. If LLMs let teachers in these schools produce genuinely varied materials in the time they actually have, students in those schools may gain access to a range of tasks they did not previously have (Cukurova et al., 2023; Luckin et al., 2022). The argument only holds if teachers in those schools also have the training, the review time, and the professional autonomy to check and adapt the output. Without those conditions, the same tool can widen gaps rather than close them, by handing under supported teachers a stream of plausible looking but generic material (Kasneci et al., 2023; Selwyn, 2024). This is one of many places in the co-pilot classroom where the tool is not the driver. The system around the tool is. 7. The Co-Pilot Metaphor and Professional Autonomy The paper's central claim can now be stated more sharply. The teacher in the co-pilot classroom is not being replaced. She is being asked to work in a new posture, in which some of what she used to do is done, in draft form, by a machine, and her job is to command, correct, and complete that draft. 7.1 Why the aviation metaphor The #co_pilot metaphor is borrowed from commercial aviation, where the captain and the first officer share the flight deck under a clear command structure. The captain has final authority. The first officer executes many routine tasks and flags anomalies. Automation supports both, but neither is replaced by it. Crew resource management, developed after several fatal accidents in the 1970s and 1980s, insists that the human pilots stay in command and that automation is a support, not a substitute (Mollick, 2024). The metaphor is useful in classrooms for three reasons. First, it names the shared work without pretending it is equal. The teacher has final authority. Second, it insists on the possibility of anomaly detection. The captain notices when the co-pilot is wrong. Third, it treats automation as a support that requires training, discipline, and clear procedures, not as magic (Luckin et al., 2022; Mollick, 2024). 7.2 Where the metaphor breaks Every metaphor has limits. The aviation metaphor breaks in at least two places when applied to teaching. First, aircraft automation is designed and certified within tight regulatory frameworks. Classroom LLMs are not. They are built and updated by commercial actors under weak or absent educational regulation (Williamson et al., 2023; UNESCO, 2023). Second, aviation has one primary goal, safe transport. Teaching has many goals, some of them in tension with one another. What counts as good teaching is contested in a way that what counts as safe flight is not (Biesta, 2022). These limits do not destroy the metaphor. They tell us where care is needed. The teacher must keep command precisely because the tools she is working with are less certified and less bounded than aviation automation. And she must keep her pedagogical goals in view precisely because the tools have no view of them at all. 7.3 Professional autonomy at stake #Professional_autonomy is often defined as the capacity of a teacher to make principled judgments about her practice, based on subject knowledge, pedagogical training, and knowledge of her students, within a framework of accountability (Biesta, 2022; Priestley et al., 2021). Autonomy in this sense is not the freedom to do anything. It is the freedom to exercise professional judgment. LLMs affect this autonomy in complex ways. On the positive side, when a teacher uses an LLM to reduce mechanical work, she can spend more time on the judgment heavy parts of her job: knowing her students, responding to their work, choosing what to emphasise. This is the argument that Luckin et al. (2022) and Cukurova et al. (2023) develop most strongly. Autonomy, in this reading, is expanded by the tool. On the negative side, several mechanisms can reduce autonomy. If schools mandate specific tools with specific default prompts, teachers lose the freedom to choose their own materials (Williamson et al., 2023). If pacing platforms use LLM output to drive minute by minute instructional decisions, teachers lose the freedom to shape the flow of a lesson (Holmes et al., 2023; Selwyn, 2024). If teacher training moves away from subject depth on the assumption that the tools will fill the gap, teachers lose the knowledge base on which their judgment rests (Kasneci et al., 2023). If accountability systems begin to measure teachers by their use of specified AI tools rather than by student learning, teachers lose the ability to say no to a tool that does not fit their class (Selwyn, 2024). None of these mechanisms is hypothetical. All are visible in current policy debates. Which of them dominate over the next decade is not a technical question. It is a political and professional question. 7.4 The curator posture The paper suggests that the healthiest posture for a teacher in the co-pilot classroom is that of an #instructional_curator. The metaphor comes from museum work. A curator does not create every object in a collection. She selects, arranges, contextualises, and interprets. She has authority over the exhibition even though she did not paint the pictures. She is expected to know the field deeply enough to make principled choices and to explain them (Cope et al., 2021). Applied to teaching, the curator posture has several features. The teacher generates or accepts a range of drafts from various sources, including LLMs, colleagues, textbooks, and open educational resources. She evaluates each against her knowledge of her subject and her students. She combines, adapts, and rewrites. She takes responsibility for what ends up in her classroom. And she keeps her own writing muscles active, because a curator who cannot write cannot judge writing. The curator posture is compatible with heavy LLM use. It is not compatible with passive LLM use. The difference is not visible in the outputs on a given day. It is visible over years, in whether the teacher's judgment sharpens or dulls. 8. Ethical and Equity Considerations The co-pilot classroom raises ethical and equity questions that cut across the whole paper. Five are worth naming. 8.1 Data and privacy Almost every commercial LLM sends prompts to a remote server owned by a private company. If a teacher pastes student work into the tool, or writes a prompt that names a specific child and describes her needs, she may be transferring personal data outside the school's control (Holmes et al., 2023; UNESCO, 2023). Many teachers do this without realising it. Guidance from ministries and regulators has been slow to catch up. Schools that want to use LLMs responsibly need clear rules about what can and cannot be entered, and those rules need to be teachable in a two hour training session, not buried in a fifty page policy (Chan, 2023). 8.2 Bias in outputs LLMs are trained on very large text collections that carry the biases of the societies that produced them. When a teacher asks for example sentences, biographies, or historical narratives, the outputs can reflect these biases in ways that are hard to detect at a glance (Bender et al., 2021; Kasneci et al., 2023). This is a curriculum issue as well as a technical one. A teacher who accepts LLM generated content for a unit on historical figures may end up with a set of examples that quietly narrows the field. The curator posture is the main defence, but it presumes that the teacher has enough #cultural_knowledge to notice the narrowing. 8.3 The digital divide, again Access to strong LLMs is not equally distributed. Better performing models are typically behind paywalls. Schools in wealthier systems and wealthier districts can afford enterprise licences with better privacy terms and stronger models. Schools in poorer systems get free tier tools with weaker output, more advertisements, and fewer protections (Williamson et al., 2023; UNESCO, 2023). The #digital_divide, once a matter of hardware and connectivity, has been renewed as a matter of AI access. 8.4 Effect on students' own writing Although this paper focuses on teacher work, the effects on students cannot be entirely bracketed. If teachers use LLMs to produce polished, uniform materials, and students use LLMs to produce polished, uniform assignments, the classroom may lose one of its oldest functions: giving students the experience of struggling with a blank page and finding their own voice (Perkins, 2023; Bearman et al., 2023). This is not an argument against LLM use. It is an argument for designing at least some tasks in which the struggle is the point. 8.5 The political economy of the tools Finally, the tools are made by a small number of large companies, most of them based in a small number of countries, whose business models depend on wide adoption (Williamson et al., 2023). Decisions about what these tools can and cannot do, what they default to, and how they are updated, are made in corporate settings with limited public accountability. Teacher agency, at the systemic level, is not just a question of what individual teachers do. It is a question of whether the profession, and the systems that employ it, have any real voice in how the tools are built (Selwyn, 2024; UNESCO, 2023). 9. Implications for Teacher Education, Leadership, and Policy If the argument of this paper is right, then #teacher_education, school leadership, and policy all need to adjust. 9.1 Teacher education Initial teacher education has to prepare new teachers for the curator posture, not just for the creator posture. This means several concrete changes. Trainee teachers should learn to write and evaluate prompts as part of their pedagogy modules, not as an add on. They should be given exercises in which they compare LLM generated lessons with human generated ones on the same topic, and identify the differences. They should be taught to spot common LLM errors in their subject, including hallucinations, over simplifications, and cultural narrowing (Mishra et al., 2023; Celik, 2023). At the same time, teacher education must not weaken subject depth on the assumption that the tools will fill the gap. The opposite is true. Because LLMs produce plausible looking content in every subject, the teacher's ability to judge that content depends on her subject knowledge being stronger, not weaker, than before (Kasneci et al., 2023). Programmes that reduce subject specific study in favour of generic technology modules will produce teachers who cannot see when the co-pilot is wrong. Continuing professional development matters at least as much as initial training. Most current teachers were trained before generative AI was available. Their route into the co-pilot classroom is through in service learning, and that learning has to be practical, ongoing, and tied to their actual subjects. One off webinars are not enough (Cukurova et al., 2023; Luckin et al., 2022). 9.2 School leadership School leaders sit at the point where policy meets practice, and their choices shape the co-pilot classroom in ways that individual teachers cannot. Several leadership tasks stand out. Leaders need to make explicit choices about which tools the school supports, on what terms, and with what data protections. Ad hoc adoption, in which individual teachers use whatever they like without guidance, is a real risk to student data and to consistency (Chan, 2023; UNESCO, 2023). Leaders need to protect professional judgment. If a tool produces a suggested lesson that a teacher decides not to use, the school should back the teacher, not the tool. This sounds obvious. It is not obvious in systems where accountability metrics begin to reward tool use as such. Leaders need to build time for curation. The gains from LLM use appear only when teachers have time to evaluate and adapt the output. A school that installs the tool and then cuts planning time on the assumption that the tool will make up the difference has misunderstood the arrangement (Luckin et al., 2022; Cukurova et al., 2023). 9.3 Policy At the policy level, three lines of work are pressing. The first is #regulation. Educational LLMs need clearer rules about what data they can collect, how they can be updated, and what claims their vendors can make about learning outcomes. The current situation, in which schools sign generic terms of service written for consumer use, is not sustainable (Holmes et al., 2023; UNESCO, 2023). The second is #procurement. Public education systems have significant purchasing power. When they use that power thoughtfully, they can push vendors toward better privacy, better transparency, and better alignment with local curricula. When they use it thoughtlessly, they lock in tools whose defaults may not suit their students (Williamson et al., 2023). The third is #evaluation. The next decade needs serious empirical work on what LLM use actually does to teacher workload, teacher judgment, and student learning, over years and not months. Much current evidence is short term and self reported. Longitudinal studies with strong designs are the only way to move the debate past its current mix of hype and panic (Kasneci et al., 2023; Bond et al., 2024). 9.4 The role of the professional community Beneath all of this sits the professional community of teachers themselves. Historically, professions have kept their autonomy by holding, collectively, the knowledge base on which their judgments rest. Doctors have medical journals and colleges. Lawyers have bars and case law. Teachers have their unions, associations, and, less formally, their staffrooms and their subject networks (Biesta, 2022). In the co-pilot classroom, these professional structures matter more, not less. It is through them that teachers can share which prompts work, which outputs are unreliable, which tools are trustworthy, and which policies should be resisted. No individual teacher can evaluate a general purpose LLM in isolation. A profession can. Whether teachers use their collective structures for this purpose in the coming years will shape the balance between agency and automation more than any single technical development (Selwyn, 2024; Cukurova et al., 2023). 10. Conclusion Generative AI has arrived in classrooms faster than any earlier educational technology, and it has arrived in a form that touches the core of teacher work. Lesson planning and differentiation, the two tasks where LLMs have moved fastest, are exactly the tasks where teachers have historically spent the most invisible hours. The tools do not remove those tasks. They change their shape. Where the teacher used to write from a blank page, she now edits from a machine draft. Where she used to produce one worksheet because she had no time for three, she can now produce three and choose the best two. This paper has argued that the pedagogical shift underneath this change is best described as a move from #content_creator to #instructional_curator, and that the working relationship between teacher and machine is best described as a co-pilot arrangement in which the human keeps command. The shift is neither automatic gain nor automatic loss. Under the right conditions, it expands teacher agency by freeing time for judgment heavy work. Under the wrong conditions, it reduces teacher agency by transferring pedagogical decisions to systems that do not know the students, the subject, or the setting. The right conditions are not mysterious. They involve strong subject knowledge, explicit training in prompting and evaluation, school policies that protect professional judgment, procurement that resists lock in, regulation that catches up with the tools, and a professional community that shares what it learns. None of these is easy. All are within reach. The wrong conditions are also not mysterious. They involve thin training, tool mandates without teacher voice, accountability that measures tool use rather than learning, and a slow erosion of the subject knowledge on which teacher judgment rests. These are the ways the co-pilot classroom can become a #ghost_flight, technically operating but with no human really in charge. The direction is not settled. It will be settled, over the next several years, by choices made by teachers, school leaders, teacher educators, and policymakers, in every system where these tools are used. This paper's central claim is that the choices should be made in the open, with the pedagogical purpose of schooling in view, and with the understanding that the teacher's role has changed but has not shrunk. What has shrunk is the burden of mechanical production. What has grown is the burden, and the privilege, of curation. In a room full of plausible drafts, the person who knows which draft belongs in front of these children, on this morning, matters more than ever. References Bearman, M., Ryan, J., and Ajjawi, R. (2023). Discourses of artificial intelligence in higher education: A critical literature review. Higher Education, 86(2), 369 to 385. https://doi.org/10.1007/s10734-022-00937-2 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 to 623. https://doi.org/10.1145/3442188.3445922 Biesta, G. (2022). World Centred Education: A View for the Present. Routledge. Bond, M., Khosravi, H., De Laat, M., Bergdahl, N., Negrea, V., Oxley, E., Pham, P., Chong, S. W., and Siemens, G. (2024). A meta systematic review of artificial intelligence in higher education: A call for increased ethics, collaboration, and rigour. International Journal of Educational Technology in Higher Education, 21(1), 4. https://doi.org/10.1186/s41239-023-00436-z Celik, I. (2023). Towards Intelligent TPACK: An empirical study on teachers professional knowledge to ethically integrate artificial intelligence based tools into education. Computers in Human Behavior, 138, 107468. https://doi.org/10.1016/j.chb.2022.107468 Chan, C. K. Y. (2023). A comprehensive AI policy education framework for university teaching and learning. International Journal of Educational Technology in Higher Education, 20(1), 38. https://doi.org/10.1186/s41239-023-00408-3 Cope, B., Kalantzis, M., and Searsmith, D. (2021). Artificial intelligence for education: Knowledge and its assessment in AI enabled learning ecologies. Educational Philosophy and Theory, 53(12), 1229 to 1245. https://doi.org/10.1080/00131857.2020.1728732 Cukurova, M., Miao, X., and Brooker, R. (2023). Adoption of artificial intelligence in schools: Unveiling factors influencing teachers engagement. In Artificial Intelligence in Education, Lecture Notes in Computer Science, vol 13916. Springer. https://doi.org/10.1007/978-3-031-36272-9_14 Holmes, W., Bialik, M., and Fadel, C. (2023). Artificial Intelligence in Education: Promises and Implications for Teaching and Learning. Globethics Publications. Holmes, W., and Tuomi, I. (2022). State of the art and practice in AI in education. European Journal of Education, 57(4), 542 to 570. https://doi.org/10.1111/ejed.12533 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 Kohnke, L., Moorhouse, B. L., and Zou, D. (2023). ChatGPT for language teaching and learning. RELC Journal, 54(2), 537 to 550. https://doi.org/10.1177/00336882231162868 Lim, W. M., Gunasekara, A., Pallant, J. L., Pallant, J. I., and Pechenkina, E. (2023). Generative AI and the future of education: Ragnarok or reformation? A paradoxical perspective from management educators. The International Journal of Management Education, 21(2), 100790. https://doi.org/10.1016/j.ijme.2023.100790 Luckin, R., Cukurova, M., Kent, C., and du Boulay, B. (2022). Empowering educators to be AI ready. Computers and Education: Artificial Intelligence, 3, 100076. https://doi.org/10.1016/j.caeai.2022.100076 Mishra, P., Warr, M., and Islam, R. (2023). TPACK in the age of ChatGPT and generative AI. Journal of Digital Learning in Teacher Education, 39(4), 235 to 251. https://doi.org/10.1080/21532974.2023.2247480 Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Portfolio. Mollick, E. R., and Mollick, L. (2023). Assigning AI: Seven approaches for students, with prompts. The Wharton School Research Paper. https://doi.org/10.2139/ssrn.4475995 OECD (2021). Teachers Getting the Best out of Their Students: From Primary to Upper Secondary Education. TALIS, OECD Publishing. Perkins, M. (2023). Academic integrity considerations of AI large language models in the post pandemic era: ChatGPT and beyond. Journal of University Teaching and Learning Practice, 20(2). https://doi.org/10.53761/1.20.02.07 Priestley, M., Nieveen, N., and Sinnema, C. (2021). Understanding Teacher Agency in Curriculum Making. In Priestley, M., Alvunger, D., Philippou, S., and Soini, T. (Eds.), Curriculum Making in Europe: Policy and Practice within and across Diverse Contexts, 187 to 213. Emerald. Selwyn, N. (2022). Education and Technology: Key Issues and Debates (3rd ed.). Bloomsbury Academic. Selwyn, N. (2024). On the limits of artificial intelligence (AI) in education. Nordisk tidsskrift for pedagogikk og kritikk, 10(1). https://doi.org/10.23865/ntpk.v10.6062 Su, J., and Yang, W. (2023). Unlocking the power of ChatGPT: A framework for applying generative AI in education. ECNU Review of Education, 6(3), 355 to 366. https://doi.org/10.1177/20965311231168423 Tomlinson, C. A. (2021). So Each May Soar: The Principles and Practices of Learner Centered Classrooms. ASCD. Tuomi, I. (2022). Artificial intelligence, 21st century competences, and socio emotional aspects of learning. European Journal of Education, 57(4), 552 to 567. https://doi.org/10.1111/ejed.12534 UNESCO (2023). Guidance for Generative AI in Education and Research. United Nations Educational, Scientific and Cultural Organization. Williamson, B., Macgilchrist, F., and Potter, J. (2023). Re examining AI, automation and datafication in education. Learning, Media and Technology, 48(1), 1 to 5. https://doi.org/10.1080/17439884.2023.2167830 #generative_AI #teacher_agency #LLMs_in_education #co_pilot_classroom #instructional_curation #lesson_planning #differentiation #professional_autonomy #TPACK #ChatGPT_in_schools #AI_pedagogy #teacher_education #prompt_engineering #critical_edtech #future_of_teaching

Download the book (PDF): Before choosing a journal or writing a word, an author needs an accurate model of what Scopus is, what it contains, and how a journal becomes and stays part of it. Misconceptions on these points are the source of most wasted effort and most avoidable mistakes. 1. What Scopus Is Scopus is an abstract and citation database of peer-reviewed literature, owned and operated by the publisher Elsevier and launched in 2004. It indexes the metadata, abstracts, and cited references of scholarly outputs across the sciences, social sciences, health sciences, and arts and humanities. It is not a full-text library: in most cases Scopus stores the descriptive record and the reference list of a document and links out to the full text held elsewhere, rather than hosting the article itself. Two functions make Scopus important to authors. First, it is a discovery system: researchers worldwide search it to find prior work, which means an indexed article is more likely to be read and cited. Second, it is a measurement system: because it records who cites whom, it can compute citation counts, author h-indexes, and journal-level metrics. Many universities and funding agencies, particularly in Asia, the Middle East, Eastern Europe, and parts of Latin America, tie hiring, promotion, degree completion, and financial incentives to publication in Scopus-indexed journals. That institutional dependence, more than any intrinsic property of the database, is why the question “Is it in Scopus?” carries so much weight. How Scopus has developed Scopus launched in 2004 as Elsevier’s entry into a field long dominated by the citation indexes that became Web of Science. From the outset it aimed for broader coverage than its established rival, and over two decades it has expanded the range of disciplines, languages, document types, and back-file references it includes, while building the curation machinery that decides what is admitted and what is removed. Its reference coverage now reaches back to 1970, and more recent additions have brought in preprints from selected servers and a growing body of open access content. The direction of travel matters to authors in one respect above all: coverage is actively curated and continually revised, so the database an author consults today is not the database of five years ago, and the status of any particular journal must be checked at the point of use rather than assumed. Scale and coverage Scopus is among the two largest curated citation databases, alongside Web of Science. Its scale is large and growing, and Elsevier reports rounded figures that are revised regularly. The values below are drawn from Elsevier’s own content materials and should be read as approximate and dated rather than exact. What Scopus indexes Scopus covers several distinct source types, and the route to inclusion differs by type. Authors should know which type their intended output belongs to, because the evaluation pathway and the relevant metrics differ. Journals. The core of Scopus. Peer-reviewed serial publications with an ISSN, selected by the Content Selection and Advisory Board against published criteria. Conference proceedings and conference series. Especially important in engineering, computer science, and the physical sciences, where conference papers can carry weight comparable to journal articles. Book series and standalone books. Selected through a publisher-level evaluation rather than title-by-title review. Trade journals. Industry-oriented periodicals included where users find their articles scientifically relevant. Preprints. Non-peer-reviewed early versions from selected servers, displayed in author profiles but treated as a distinct, non-core content type. Routes other than journals: conference papers, books, and chapters Authors often assume that a Scopus publication must be a journal article. It need not be. Several other output types are indexed, and in some fields they are the primary currency. Knowing the route that fits your work, and how it is evaluated, prevents both missed opportunities and misplaced effort. Conference papers. In computer science, engineering, and parts of the physical sciences, peer-reviewed conference papers carry weight comparable to journal articles, and major conferences are more selective than many journals. Scopus indexes conference papers when the proceedings meet its criteria, often through established proceedings series. The practical points are to target conferences whose proceedings are indexed, to confirm that indexing on the Scopus Sources list rather than trusting the conference’s own claim, and to be alert that “predatory conferences” exist in the same way predatory journals do, soliciting fees for proceedings that are never properly indexed. Books and monographs. Scholarly books, including monographs, edited volumes, and reference works, are indexed where the publisher and title meet Scopus’s book criteria. Unlike journals, books are evaluated at the level of the publisher and the book programme rather than title by title, so the reputation and editorial standards of the press matter a great deal. For authors in the humanities and many social sciences, where the book remains a central form of scholarship, this is an important and sometimes overlooked route into the database. Book chapters. Individual chapters in an indexed book or book series are themselves indexed and citable, with their own metadata and reference lists. A chapter in a well-regarded edited volume from a reputable academic publisher can therefore be a legitimate Scopus output. The same caution applies as for journals and conferences: confirm that the book or series is covered, and weigh the standing of the publisher, because chapters in obscure or pay-to-publish volumes carry little of the value authors expect. The metrics differ by type. Journal-level indicators such as CiteScore apply to serials; books and chapters accrue citations individually and are assessed more on the standing of the publisher and the citations the specific work attracts. Whichever route you take, the verification discipline in Part III applies unchanged. Subject classification and why it shapes quartiles Scopus assigns every source to one or more subject categories under a classification scheme commonly known by its abbreviation, ASJC (the All Science Journal Classification). A journal may sit in a single narrow category or in several broad ones, and this placement is not a cosmetic label: it is the basis on which percentiles and quartiles are calculated, because a journal is ranked against the other journals in the same category. The same journal can therefore be Q1 in a niche category where it is a leader and Q2 or Q3 in a broad category where it competes with stronger titles. Two practical consequences follow for authors. First, when a rule or a colleague refers to a journal’s quartile, ask which category it refers to, because the answer can differ. Second, when judging fit, look at the categories a journal is classified under: they tell you how Scopus understands the journal’s scope, which is often more precise than the journal’s own marketing description of its aims. Scopus, Web of Science, and Google Scholar Authors frequently conflate these three systems. They are not interchangeable, and a journal’s presence in one says little about the others. Practical consequence. If your institution requires “Scopus-indexed,” a journal’s Impact Factor or its presence in Web of Science does not satisfy that requirement. Likewise, appearing in Google Scholar means only that Google found the article; it is not a quality marker and does not imply Scopus coverage. Always check the specific database your obligation names. What “Scopus-indexed” does and does not mean Being indexed in Scopus means a journal passed the database’s selection criteria and that its content is discoverable and citable within the system. It does not mean the journal is prestigious, high-impact, or even reliably stable. Scopus spans the full quality range from leading journals to titles that are weak and at risk of removal. Two misreadings are common and costly: assuming that any Scopus journal is a safe, high-quality venue, and assuming that indexing is permanent. Neither is true. The chapters that follow give the tools to tell strong Scopus journals from weak ones and to confirm that a title is still active before you submit. Author profiles, affiliation profiles, and search tools Scopus is not only a record of journals; it also builds a profile for every author whose work it indexes and for every institution. These profiles are what make the database a measurement system, and an author benefits from understanding them early. · Author profile. Scopus automatically groups an author’s indexed documents into a profile with a unique Scopus Author Identifier, and computes that author’s document count, citation count, and h-index from it. Because the grouping is automatic, profiles are sometimes split or merged incorrectly, especially for common names or after a name change; authors can request corrections, and linking an ORCID iD keeps the record coherent across databases. · Affiliation profile. Institutions have their own profiles, which aggregate the output of their affiliated authors and underpin university analytics and rankings. Consistent use of an institution’s official name on submissions keeps an author’s work attributed to the right place. · Search and discovery tools. Beyond basic search, Scopus offers analytical views — citation overviews, document and source comparisons, and, for subscribers, an AI-assisted search layer that summarises and surfaces relevant literature. These are aids to finding and assessing work, not substitutes for reading it. For an author, the practical lesson is to claim and check your Scopus author profile, correct any misattributed or duplicated records, and register an ORCID iD so that your output is counted accurately. A fragmented profile understates a researcher’s record precisely when committees are looking at it. Hashtags: #AcademicPublishing #Scopus #ResearchHandbook #PeerReview #ScholarlyPublishing #JournalSelection #PublishOrPerish #AcademicWriting #Elsevier #CitationIndex #HIndex #CiteScore #JournalQuartiles #ASJC #AuthorProfile #ORCID #WebOfScience #GoogleScholar #HigherEducation #AcademicCareer #ResearchImpact #TenureTrack #PhDLife #PostdocLife #UniversityRankings #AcademicTwitter #BookExcerpt #MustRead #PredatoryJournals #ConferenceProceedings #ResearchMethodology #AcademicBestPractices

The global tourism industry has entered a period of deep transformation in which travelers increasingly search for experiences that feel authentic, ethical, and rooted in place. Among the most visible expressions of this shift is #farm_to_fork tourism, a hybrid form of #agritourism and #culinary_tourism that connects visitors directly with the farmers, fishers, herders, artisans, and cooks who shape a region's food identity. This article examines how #gastronomic_leisure functions as a mechanism for strengthening local #supply_chains, financially sustaining small and medium sized agricultural producers, protecting regional food heritage, and reducing the carbon footprint associated with long distance food sourcing at destinations. Drawing on a narrative review of peer reviewed literature published between 2020 and 2025, along with documented practices from rural regions across Europe, Asia, Latin America, and Africa, the study builds a conceptual framework that links visitor demand, producer participation, and territorial governance. Findings suggest that farm-to-fork tourism improves farm household income diversification, keeps a larger share of tourist spending within rural communities, encourages the retention of traditional cultivars and #heritage_livestock, and reduces food kilometers when destination restaurants prioritize short chains. Barriers include seasonality, weak logistics, uneven digital literacy, and fragile cooperation between hospitality and agriculture. The paper proposes an integrative model of #short_food_supply_chains embedded in destination management, and offers practical recommendations for #policy_makers, hospitality operators, and rural cooperatives. The findings contribute to debates on #sustainable_tourism, #food_sovereignty, and #rural_regeneration, and offer a reference framework that students and early career researchers can use to design their own studies of place based food tourism. Keywords: farm-to-fork tourism, agritourism, culinary tourism, short food supply chains, sustainable gastronomy, rural development, food miles, regional food identity 1. Introduction For much of the late twentieth century, food at tourist destinations was treated as a background service. Hotels imported standardized ingredients from global wholesalers, resorts served menus that could have been printed in any country, and the farms surrounding popular destinations often had no commercial relationship with the visitors passing through. This decoupling between #travel and #terroir produced a paradox in which regions rich in #agricultural_heritage exported their labor and land to feed distant markets while importing bland foodservice supplies for their own guests. In the last decade this pattern has begun to reverse. Travelers today expect that the food on their plate tells a story of the place they are visiting, and that story now carries measurable value in the form of longer stays, higher spend, and stronger destination loyalty (Sidali et al., 2021). Farm-to-fork tourism sits at the intersection of two established trends. The first is #agritourism, defined broadly as tourism activities that take place on working farms or in rural landscapes and that involve visitors in agricultural production, processing, or #rural_lifestyle. The second is #culinary_tourism, understood as travel motivated wholly or partly by the desire to eat and drink distinctive foods that express a #cultural_identity. Where these two streams meet, a new category emerges that emphasizes the traceable path of food from the field to the visitor's plate, with the shortest possible distance and the highest possible cultural fidelity in between (Testa et al., 2022). This article uses the term farm-to-fork tourism to describe that meeting point, and treats #gastronomic_leisure as its behavioral core: the pleasure of eating becomes inseparable from the pleasure of understanding where the meal was born. The rise of this category is not accidental. It reflects a convergence of consumer values around #sustainability, health, #authenticity, and skepticism toward industrial food systems that has been amplified by pandemic era attention to fragile global logistics (Hall et al., 2021). It also reflects a strategic response by rural regions that have watched their populations age, their farmland fragment, and their small producers struggle against price pressure from supermarket chains. For these regions, tourism offers a demand channel that pays a premium for exactly the products that industrial systems undervalue: heirloom vegetables, native breed meats, unpasteurized cheeses, wild fermented drinks, and slow ripening fruits. When that channel is organized carefully, it produces what scholars call #territorial_value, meaning the retention of economic surplus, cultural meaning, and ecological function within the place where the food is grown (Everett, 2022). Yet the promise is not automatic. A farm-to-fork label placed on a menu does not guarantee that the beef came from the neighboring valley or that the farmer received a fair share of the price. Weak logistics, seasonal mismatch between supply and visitor peaks, thin cooperation between hospitality and agriculture, and the temptation for restaurants to greenwash their sourcing all threaten to hollow out the concept. If tourism scholars and #policy_makers want farm-to-fork tourism to deliver on its promise, they need to look past the marketing surface and analyze the actual structure of the supply chains that connect farms to forks, along with the governance arrangements that hold those chains together (Sims, 2023). This article contributes to that analysis in three ways. First, it synthesizes recent peer reviewed evidence on how #short_food_supply_chains behave when they intersect with tourism demand, drawing on studies from #Mediterranean, #Andean, Southeast Asian, and Sub Saharan contexts. Second, it develops a conceptual framework that positions farm-to-fork tourism as a triadic relationship between the #visitor, the #producer, and the #destination, mediated by #hospitality operators and #public_governance. Third, it translates the framework into practical guidance for students, entrepreneurs, and rural planners who want to design farm-to-fork initiatives that survive beyond their first grant cycle. The remainder of the article proceeds through a literature review, a methodological note, the presentation of the framework, a discussion of empirical evidence and challenges, and a set of policy recommendations followed by a conclusion. 2. Literature Review 2.1 The evolving definition of gastronomic tourism Early treatments of food related travel described eating as an incidental part of the tourist experience, subordinate to sightseeing or leisure. The scholarship of the last five years has moved that framing decisively. #Culinary_tourism is now understood as a primary motivation for a substantial share of travelers, and food is theorized as a #cultural_heritage asset comparable to architecture or music (Ellis et al., 2020). Researchers distinguish between food tourism, which is oriented around consumption of dishes in restaurants or markets, and gastronomic tourism, which extends into the entire production system, including farms, orchards, vineyards, fisheries, and processing workshops. Farm-to-fork tourism belongs to the gastronomic tourism family because its defining feature is engagement with the productive origin of the meal, not only the meal itself. Recent literature also emphasizes the affective and identity dimensions of food travel. Visitors who cook a family recipe with a local host, harvest olives with a farming couple, or watch a fisher return with the morning catch report levels of place attachment that traditional sightseeing rarely generates (Björk and Kauppinen Räisänen, 2022). These emotional bonds translate into concrete economic outcomes such as repeat visits, word of mouth recommendation, and willingness to pay premiums for #origin_labeled products long after the trip ends. Farm-to-fork tourism thus operates simultaneously as a leisure product, a marketing channel for #regional_food, and a soft infrastructure for building durable consumer relationships with rural producers. 2.2 Agritourism and the diversification of farm income Parallel to the food tourism literature, agritourism scholarship has documented how farm household income has grown increasingly precarious under global commodity pressure, climate volatility, and generational turnover. Studies from Italy, Portugal, and Slovenia report that farms which added tourism activities such as farm stays, tasting rooms, cooking workshops, and open days achieved income diversification that reduced their exposure to single crop failure and improved their resilience during price shocks (Ammirato et al., 2020). The magnitudes vary by region and by farm size, but a consistent pattern emerges: the tourism euro spent on the farm is not simply added to the agricultural euro, it changes the composition of what the farm produces. Farms with tourism activities are more likely to retain #native_varieties, maintain landscape features such as hedgerows and stone walls, and adopt lower input practices because these features attract visitors who pay for #authenticity. The literature has also documented the limits of agritourism. Not all farms can convert into tourism operators, and those that try face steep learning curves in hospitality, safety, and marketing. Small farms in remote locations may lack the visitor volume necessary to sustain a full tourism enterprise on their own. This is precisely where the farm-to-fork extension of agritourism becomes strategic. Instead of asking every farm to become a hotel or a restaurant, farm-to-fork tourism allows farms to remain farms while participating in tourism value chains through supply relationships with restaurants, hotels, cooperatives, and destination management organizations (Cavicchi and Ciampi Stancova, 2021). 2.3 Short food supply chains and territorial economies The economics of farm-to-fork tourism depends heavily on how short the underlying supply chains actually are. #Short_food_supply_chains are commonly defined by three criteria: physical proximity between producer and consumer, social proximity in the sense of direct communication and trust, and organizational proximity in that few or no intermediaries stand between the field and the plate. Meta analyses of European and North American cases suggest that when tourism restaurants source through short chains, they retain between 60 and 80 percent of the food expenditure within the region, compared to as little as 20 percent when they source through global wholesalers (Kneafsey et al., 2021). The difference translates directly into rural employment, farm investment, and #tax_revenue for local governments. Short chains also produce ecological benefits. Life cycle analyses of restaurant meals show that transport related emissions can be reduced substantially when ingredients travel tens rather than thousands of kilometers, although the net carbon effect depends on production practices and the mode of transport used within the short chain (Poore and Nemecek, 2022 update). Farm-to-fork tourism therefore carries a plausible claim to lower #food_miles, but this claim needs to be verified case by case rather than assumed. A restaurant that flies in truffles from another continent to serve them with locally grown pasta has not honored the concept, even if the pasta is regional. 2.4 Regional food identity and heritage A third body of literature relevant to this study addresses the role of food in #regional_identity. Traditional dishes, cultivars, and preservation techniques function as living #cultural_heritage that is at risk of erosion when younger generations move to cities and when supermarkets standardize taste (Bessiere, 2022). Tourism has an ambiguous relationship with this heritage. On one hand, tourist demand can reinvent markets for endangered products and encourage the transmission of skills from older to younger practitioners. On the other, tourist demand can flatten heritage into caricature, replacing complex local knowledge with photogenic simplifications that satisfy visitors without sustaining the underlying practice. Recent work on #geographical_indications such as protected designations of origin and protected geographical indications shows that formal recognition schemes can help balance these forces by tying tourism marketing to verified production standards (Belletti et al., 2023). Farm-to-fork tourism benefits when regions have such recognized products because visitors can trust that what they are eating corresponds to a documented tradition. Even in regions without formal designations, informal community codes about what counts as authentic can play a similar role if they are respected by hospitality operators. 2.5 Sustainability, carbon, and the ethics of tourist food The last stream of literature considers the environmental ethics of tourist food. Global tourism is a substantial source of greenhouse gas emissions, and food and beverage services form a nontrivial share of that footprint (Lenzen et al., 2020 update; Gossling and Peeters, 2021). Studies estimate that food related emissions in a typical destination represent between 10 and 25 percent of the tourism sector's total emissions, depending on cuisine, transport, and waste patterns. Farm-to-fork strategies offer a partial remedy by shortening transport distances, reducing packaging, and often favoring seasonal plant based ingredients that carry lower per calorie footprints. They are not a silver bullet, because emissions from tourist travel to the destination usually dwarf food emissions at the destination, but they contribute to a plausible narrative of #responsible_tourism that increasingly influences traveler choice. 3. Methodology This article uses a narrative review approach combined with conceptual modeling. A narrative review is appropriate when a field is developing rapidly and includes contributions from tourism studies, rural sociology, agricultural economics, food systems analysis, and hospitality management, because it allows integration across disciplinary vocabularies. The search strategy covered peer reviewed articles, book chapters, and grey literature published between January 2020 and early 2026 in English, Spanish, Italian, and Portuguese, indexed in Scopus, Web of Science, and Google Scholar. Search terms combined variants of farm-to-fork, agritourism, culinary tourism, short food supply chains, food miles, and rural regeneration. Sources were screened for relevance to the intersection of tourism and local supply chains, and additional sources were identified through backward and forward citation tracking. The synthesis privileges evidence that is grounded in empirical case study or survey work over purely theoretical contributions, without excluding the latter. The conceptual framework presented in section four was developed inductively from the reviewed literature. Recurrent themes and mechanisms were coded and clustered, and the resulting clusters were arranged into a triadic model of visitor, producer, and destination, with mediating roles for hospitality operators and public governance. The framework is offered as a heuristic device that students and practitioners can adapt to their own contexts rather than as a definitive theoretical claim. 4. Conceptual Framework: The Triadic Model of Farm-to-Fork Tourism Farm-to-fork tourism is best understood not as a single product but as a #relational_system in which three principal actors negotiate value. The first actor is the visitor who arrives at the destination with expectations about food, culture, and experience. The second is the producer who grows, raises, catches, or processes the food. The third is the destination itself, understood as the territorial community that includes local governments, resident consumers, landscape features, and #cultural_institutions. Between these actors, two mediating layers operate. Hospitality operators, meaning restaurants, hotels, guesthouses, cooperatives, and food tour organizers, translate producer supply into visitor experience. Public governance, meaning regional and municipal authorities, extension services, and agencies that manage geographic indications and #food_safety, sets the rules and provides the infrastructure that allows the system to function. 4.1 Visitor demand as pull Visitor demand for farm-to-fork experiences has grown along a spectrum of intensity. At one end, visitors want a restaurant menu that lists the farms of origin next to the dishes. At the other, visitors want to spend a day on the farm, cook with the family, and take home preserved products they helped make. Between these extremes lie many intermediate formats such as farm tours, market visits, cheese and wine pairings, foraging walks, fishing trips, and harvest festivals. Each format implies a different level of producer engagement and a different price point, and the demand for each is shaped by traveler demographics, trip duration, and #digital_channels through which experiences are booked. Younger travelers report stronger interest in experiential and educational components, while older travelers often prefer comfort formats that still deliver #authenticity, such as tasting menus paired with visits from the producer to the table. Solo travelers, families, and small groups all show growing willingness to pay for experiences that promise learning, connection, and #storytelling. This diversity implies that destinations should not offer a single farm-to-fork product but a portfolio calibrated to different segments. 4.2 Producer capacity as push On the supply side, producer capacity determines what tourism demand can absorb. Small farms typically face constraints in volume, consistency, packaging, invoicing, and #food_safety_compliance. A farm that raises heritage breed pigs may only slaughter a small number of animals per season, and a fisher may not know from one week to the next whether the catch will be sufficient to supply a hotel. These constraints are not signs of failure but structural features of small scale production that also underpin the quality tourism demands. Farm-to-fork systems succeed when they are designed to accommodate these features rather than force producers into industrial reliability standards they cannot meet. Cooperative structures, producer groups, and shared aggregation points play an important role here. When several farms pool their supply, they can meet minimum order sizes for hospitality clients while preserving individual identities on the menu. Shared #cold_storage, joint transport, and coordinated planning reduce logistics costs that no single small farm could absorb. 4.3 Destination as context The destination is more than a location. It is the fabric of landscapes, memory, and community that gives farm-to-fork tourism its meaning. A destination with strong food heritage, active markets, and a lively calendar of #food_festivals offers a rich context in which farm-to-fork initiatives can flourish. A destination without these features can still develop farm-to-fork tourism but must invest more heavily in building the missing infrastructure. Destination management organizations play a crucial role by curating the offer, ensuring quality, and communicating with visitors in ways that individual farms and restaurants cannot afford on their own. The destination also acts as the level at which environmental impact is measured. Reducing #food_miles is meaningful only if the reduction is calculated across the destination's food supply as a whole, including hotels, restaurants, cafes, and events. A single restaurant sourcing locally is admirable but has limited effect. A destination in which the majority of hospitality operators source through short chains creates a systemic change that shows up in the region's environmental accounts. 4.4 Hospitality operators as translators Restaurants, hotels, and other hospitality actors sit at the crucial junction where producer supply meets visitor experience. Their menu design decisions, procurement contracts, and staff training determine whether the farm-to-fork promise reaches the plate. A chef who takes the time to visit farms, learns the seasonal rhythm, and rewrites the menu around what is available creates a genuinely place based cuisine. A chef who insists on a fixed menu regardless of season forces the kitchen back into industrial supply. Hospitality operators also act as narrators. The way they present the farm-to-fork story to guests shapes visitor understanding and willingness to pay. Training front of house staff to speak knowledgeably about the producers behind the dishes is a small investment with high returns in guest satisfaction and repeat business. 4.5 Public governance as scaffolding Public governance provides the scaffolding without which the whole system becomes fragile. #Food_safety regulations must accommodate small scale processing without lowering health standards. Zoning and land use policies must protect agricultural land near tourism zones from conversion. #Geographical_indications, quality labels, and origin schemes must be enforced. Public procurement, for example in schools and hospitals, can complement tourism demand and stabilize producer revenue across seasons. Extension services and vocational training must build farmer capacity in areas such as hospitality, digital marketing, and financial planning. Where these public functions are strong, farm-to-fork tourism scales more easily. Where they are weak, individual initiatives struggle to persist beyond charismatic founders or single grants. 5. Findings and Discussion 5.1 Economic sustainability of local agriculture The strongest and most consistent finding across the reviewed literature is that farm-to-fork tourism improves the economic sustainability of participating farms, provided the chain design is favorable. Studies from #Tuscany, #Andalusia, the #Douro_Valley, #Kerala, and the #Andes report that farms which supply tourism markets through short chains capture prices that are between 20 and 60 percent higher than those they would receive through conventional wholesale channels (Torres and Momsen, 2021; Contini et al., 2023). The premium reflects the willingness of visitors to pay for perceived authenticity, quality, and story, and it is retained by producers because intermediary layers are minimized. Higher prices are only part of the picture. Farms participating in farm-to-fork tourism also report more stable demand once relationships with hospitality operators become routine, which reduces the classic vulnerability of small producers to spot market fluctuations. Contract farming arrangements adapted to the tourism sector, in which restaurants commit to purchasing a share of a farm's output at agreed prices for a season, have been documented in several European regions and provide farmers with the predictability they need to plan investments in equipment, land improvements, and labor (Cavicchi and Ciampi Stancova, 2021). The presence of tourism demand also encourages diversification of farm output, which further stabilizes income. Equally important is the effect on #generational_renewal. Rural areas across the world face a demographic crisis in which farm succession has broken down and young people leave for cities. Farm-to-fork tourism reintroduces a set of activities on the farm, including hosting, teaching, and #storytelling, that appeal to younger family members who might not have wanted to farm in the traditional sense. Case studies from #Puglia, #Basque_country, and rural #Japan describe returning children who took over family farms because tourism gave them a viable and meaningful role that pure agricultural production did not (Sidali et al., 2021). The economic effect of retaining even a few young families in a village extends far beyond the farm itself, touching schools, health services, and the sustainability of the community as a whole. 5.2 Retention of tourist spending in local economies A second important finding concerns the leakage of tourist expenditure. Conventional resort tourism suffers from high leakage rates when hotels are foreign owned, food is imported, and staff wages are repatriated. Regions dominated by such tourism find that visitor arrivals do not translate proportionally into local prosperity. Farm-to-fork tourism significantly changes this arithmetic. When restaurants source ingredients locally, when hotels feature regional products in breakfast and minibars, and when tour operators route visitors to farm experiences, the share of tourist spending that remains in the local economy rises sharply. Empirical estimates from Italian and Spanish rural regions suggest that a full farm-to-fork strategy can increase the local retention of tourist food expenditure by 30 to 50 percentage points compared with conventional sourcing patterns (Kneafsey et al., 2021). The multiplier effect is substantial because farmers who receive higher revenues spend a significant part locally on inputs, wages, and services, and the money circulates through village economies before eventually leaking outward. This retention effect is one of the strongest arguments for public support of farm-to-fork initiatives, because it aligns tourism policy with #regional_development in ways that other tourism strategies rarely do. 5.3 Preservation of regional food identity Farm-to-fork tourism has been credited with helping to preserve endangered aspects of #regional_food identity. The mechanism is straightforward. Tourist demand creates markets for products that industrial food systems would abandon, and the visibility of tourism gives cultural weight to practices that might otherwise be dismissed as backward. Native maize varieties in the Andes, ancient wheat landraces in southern Italy, rare cheese making techniques in the Balkans, and small scale coastal fisheries in Southeast Asia have all been documented as beneficiaries of tourism driven demand (Belletti et al., 2023; Bessiere, 2022). The preservation function is not automatic and depends on how tourism markets are structured. When restaurants insist on a rotating cast of ancient cultivars, farmers respond by planting them. When restaurants prefer versatile modern varieties that are easier to cook, ancient cultivars retreat. The choices made by chefs and menu designers therefore have surprisingly deep consequences for #biodiversity_conservation in agricultural landscapes. There is also a knowledge transmission dimension. When tourists come to a farm to learn how to make a traditional cheese, the farmer must articulate techniques that were previously tacit. This articulation forces a codification of knowledge that can then be transferred to younger family members, students, and other producers. Several regional #culinary_academies and vocational schools have partnered with farm-to-fork networks to build curricula rooted in living tradition rather than in generic hospitality textbooks. At the same time, scholars warn against the risk of #commodification, in which heritage is reduced to spectacle. A farm that stages a folkloric performance twice a day for tour buses is not preserving heritage, it is manufacturing a product. Genuine preservation requires that the practices be maintained in their productive function first, with tourism as a complement rather than a replacement. The distinction is subtle but crucial, and it explains why some initiatives that appear successful in visitor numbers actually erode the heritage they claim to protect. 5.4 Environmental impact and food miles The environmental case for farm-to-fork tourism rests primarily on the reduction of #food_miles and the associated transport emissions. Evidence in this area is more nuanced than popular narratives suggest. Life cycle assessments show that transport emissions per kilogram of food are usually a smaller share of total food emissions than production emissions, which means that shortening the chain does not always produce dramatic reductions in the overall #carbon_footprint (Poore and Nemecek update, 2022). Nonetheless, farm-to-fork sourcing does deliver measurable environmental benefits when combined with other sustainable practices such as seasonal menus, plant forward composition, reduced packaging, and low input production. A more powerful environmental case emerges when farm-to-fork sourcing is combined with #agroecological_practices on the supplying farms. Farms that use cover crops, integrate livestock and plants, and rely on natural pest management tend to produce lower per unit emissions than industrial farms of the same crop. When tourism supply chains reward such practices with premium prices, they align farmer incentives with ecological outcomes in a way that regulatory approaches alone often fail to achieve (Everett, 2022). This alignment is arguably the deepest environmental contribution of farm-to-fork tourism, because it changes what is being produced, not only how it is transported. Waste reduction is another important dimension. Short chains reduce the losses that occur in long distance transport and cold storage, and menus built around what is available limit the pressure to discard imperfect or slower moving ingredients. Restaurants that participate seriously in farm-to-fork networks report significant reductions in kitchen waste as chefs learn to use the whole animal, the whole vegetable, and the whole season (Testa et al., 2022). 5.5 Social capital and community cohesion Beyond economic and environmental effects, farm-to-fork tourism strengthens the social fabric of rural communities. The regular interaction between farmers, chefs, tourism operators, and local officials builds trust and #social_capital that can be mobilized for other purposes such as landscape restoration, disaster response, and youth programming. Case studies describe the emergence of village level food councils, joint marketing associations, and shared brands that would not have formed in the absence of tourism cooperation (Ammirato et al., 2020). The social effect is not universally positive. Where tourism grows rapidly and unevenly, tensions can emerge between farms that benefit and farms that do not, or between long resident families and newcomers who arrive to capitalize on the trend. Effective governance mechanisms are needed to distribute benefits, mediate conflicts, and protect the collective goods on which the whole system depends, such as the reputation of the region and the quality of its landscapes. The role of #public_governance in maintaining these commons cannot be overstated. 5.6 Regional examples Several regional experiences illustrate the potential and challenges of farm-to-fork tourism. In the Slow Food movement's home region of Italy, decades of network building have produced a mature ecosystem in which farms, restaurants, cooperatives, and municipalities interact routinely around food heritage. Visitors to regions such as Piedmont, Emilia Romagna, and Sardinia now encounter menus that name individual farms, cheeses, and vineyards, and a large share of what they eat travels less than 50 kilometers to reach their plate (Sidali et al., 2021). The economic and cultural benefits are visible in the persistence of small farms, the vitality of village markets, and the ongoing transmission of artisan skills. In the Basque region of Spain and France, culinary institutions such as cooking schools and gastronomic societies have partnered with farms, fishers, and cider houses to develop networks in which chefs source from producers they know personally. The resulting cuisine has attracted international attention, generated employment for young people, and elevated the region's identity while sustaining producers in mountain and coastal villages (Contini et al., 2023). In Latin America, high altitude regions of Peru and Bolivia have developed farm-to-fork itineraries around native potato varieties, quinoa cultivation, and Andean cuisine. These itineraries have created markets for smallholder farmers whose products previously had no commercial value beyond subsistence and have contributed to the international profile of #Andean_gastronomy (Torres and Momsen, 2021). Similar dynamics can be observed in Oaxaca and other regions of Mexico that have organized farm-to-fork experiences around #maize, #chocolate, and #mezcal. In Asia, Japanese rural regions have integrated farm stays into recovery strategies after disasters and population decline, using #agritourism to reintroduce younger visitors to villages that were losing residents. Rice paddies, tea plantations, and mountain vegetable farms have become the scenes of educational tourism that supports both preservation and economic renewal. In parts of Southeast Asia, coastal communities have used farm-to-fork frameworks to elevate small scale fisheries as tourism assets, generating revenue that supports #marine_conservation. Kerala and Sri Lanka have developed spice route tourism that links farms, kitchens, and heritage sites into coherent visitor experiences. In Africa, examples from South Africa, Rwanda, and Ethiopia demonstrate how coffee, cocoa, and dairy origin tourism can channel visitor interest into producer income, though these programs often face steeper infrastructure and marketing challenges than their European counterparts (Sims, 2023). 6. Challenges and Constraints 6.1 Seasonality and supply variability The most persistent operational challenge in farm-to-fork tourism is the misalignment between agricultural and tourism seasons. In many regions, peak tourist arrivals occur during a limited window that does not coincide with the harvest of most local products. A summer beach destination may see millions of visitors between June and August but have limited local production ready during that period. A mountain destination may host winter sports enthusiasts when agricultural production is at its lowest. Bridging these mismatches requires investment in #preservation techniques such as curing, fermentation, drying, and cold storage that extend the availability of local products beyond their fresh window. Supply variability compounds the challenge. A restaurant that promises a local product cannot always guarantee that the product will be available on a given day, and this uncertainty conflicts with the operational patterns of the hospitality industry, which typically requires predictable menus and reliable inventory. Some restaurants have responded by adopting flexible menu strategies in which items rotate with availability and the daily specials board becomes a central marketing tool (Testa et al., 2022). Others have invested in relationships with multiple small suppliers so that when one falls short another can compensate. These strategies work but require chefs and managers who are willing to abandon the industrial predictability model. 6.2 Logistics and cold chain infrastructure Small producers often lack the logistics infrastructure required to serve hospitality clients efficiently. Delivering ten kilograms of vegetables to five restaurants across a region can consume a significant share of a farmer's day and eat into the price premium the short chain was supposed to deliver. Shared logistics platforms, cooperative delivery routes, and #food_hubs where multiple producers aggregate their output for pickup by hospitality clients are practical responses that have been documented in several European and North American cases (Kneafsey et al., 2021). Investment in such infrastructure often requires public support because no individual actor can justify the capital cost alone. Cold chain gaps are particularly problematic for perishable products such as fish, dairy, and leaf vegetables. In regions with limited electrical reliability or road quality, farm-to-fork projects need to invest in ice production, insulated transport, and efficient scheduling to keep products safe and fresh. Failure at this level can rapidly damage the reputation of a farm-to-fork initiative, even if all other elements are in place. 6.3 Regulatory frictions Regulations designed for industrial food systems often place disproportionate burdens on small producers. Meat processing rules that require large scale slaughterhouse throughput make it difficult for heritage breed farms to sell whole carcasses to restaurants. Dairy rules that require pasteurization may exclude traditional raw milk cheeses that are precisely what tourists come to eat. Fisheries rules designed for large vessels can criminalize the small scale artisanal fishers whose catches would be central to a coastal farm-to-fork offer. Reforming these rules is politically complex because #food_safety concerns are legitimate and food borne illness at tourism destinations has serious consequences. Approaches that maintain safety while adapting to small scale realities include mobile slaughter units, on farm processing under supervised protocols, and heritage exemptions that allow traditional production under documented practices. Regions that have implemented such reforms have seen faster growth of farm-to-fork markets, but the political work required is often underestimated (Belletti et al., 2023). 6.4 Digital literacy and marketing Farm-to-fork initiatives depend increasingly on digital channels for booking, marketing, and customer engagement. Small producers and rural hospitality operators often lack the digital skills needed to operate effectively in this environment. Beautiful farms with excellent products may remain invisible to potential visitors because their #digital_footprint is limited. Training, shared digital platforms, and mentoring programs can help, but the digital divide between urban and rural regions remains a real obstacle. Recent work on rural entrepreneurship stresses the importance of intermediary organizations that can bring digital capacity to producers who would otherwise be excluded (Cavicchi and Ciampi Stancova, 2021). 6.5 Risk of superficial adoption A subtle but important challenge is the risk that farm-to-fork branding is adopted superficially by hospitality operators who continue to source most of their inputs conventionally while advertising a small share as local. This kind of #greenwashing degrades consumer trust and undermines producers who have genuinely committed to short chain participation. Verification schemes, transparent sourcing disclosures, and independent auditing can address this risk, but they add cost and complexity that not every destination is willing to bear (Ellis et al., 2020). Consumer education is also part of the solution, because visitors who ask specific questions about sourcing are less easily satisfied by generic claims. 6.6 Carrying capacity and overtourism Farm-to-fork initiatives that become highly successful can generate a new problem in the form of pressure on the very landscapes and communities that made them attractive. Village restaurants that once served a modest evening trade can be overwhelmed by tour buses. Small farms that welcomed a handful of visitors per week can be swamped by daily crowds that disrupt farm work and stress family members. Overtourism erodes the authenticity that farm-to-fork tourism promises and can push producers back out of the tourism economy if their farm operations suffer. Managing #carrying_capacity requires deliberate planning. Booking systems that cap visitor numbers, distribution of tourism across a wider set of farms and villages, and pricing strategies that discourage bulk arrivals are among the tools available. Destinations that ignore these questions risk building fragile success that collapses when the strain becomes too great. 7. Policy Implications and Recommendations The evidence reviewed above supports a set of policy recommendations directed at rural planners, tourism authorities, and educational institutions. First, destinations should treat farm-to-fork tourism as a #cross_sectoral strategy rather than a hospitality niche. This means aligning agricultural policy, tourism policy, and rural development policy around shared objectives. Governance structures such as inter-ministerial working groups, regional food councils, and territorial development contracts can institutionalize this alignment (Cavicchi and Ciampi Stancova, 2021). Second, public investment should prioritize enabling infrastructure such as food hubs, shared cold storage, small scale processing facilities, and mobile slaughter units. These investments are the physical scaffolding of short chains and are unlikely to be built by the private sector alone due to their public goods character. Investment should also include digital infrastructure such as broadband, shared marketing platforms, and #training in digital literacy for producers. Third, regulatory frameworks should be adapted to accommodate small scale production while maintaining #food_safety. Approaches such as risk based inspection, tiered regulation by scale, and heritage exemptions offer models that have been tested in several countries. Consultation with producers, chefs, and public health experts is essential to design regulations that work in practice. Fourth, quality assurance schemes should be developed and enforced to protect the credibility of farm-to-fork claims. #Geographical_indications, certified networks, and community codes of practice can all play a role, and independent verification adds trust. Consumer facing communication about these schemes should be simple enough for visitors to understand without overwhelming them with technical detail. Fifth, education and training should be prioritized at multiple levels. Vocational schools, hospitality academies, and #culinary_institutes should integrate farm-to-fork content into their curricula. Farmer training programs should include hospitality, marketing, and financial literacy. Higher education programs in tourism and rural development should offer specialized tracks in gastronomic tourism informed by empirical research, and student capstone projects can contribute to knowledge building at the local level (Bessiere, 2022). Sixth, environmental accounting should be integrated into destination management. Regions should track food miles, kitchen waste, and packaging use as part of their tourism metrics, and set targets for improvement. Transparency about these metrics builds credibility with environmentally conscious travelers and provides a basis for public policy adjustment. Seventh, planning tools should address #carrying_capacity from the outset. Destinations should map the fragility of their farm-to-fork ecosystems and set limits on visitor volumes, seasonal distribution, and #site_intensity that protect the integrity of the underlying agriculture and community life. Eighth, governance should include #community_voice mechanisms so that residents and producers can shape the direction of tourism development rather than have it imposed on them by external investors or platforms. Genuine participation is not only ethically required but is also practically necessary to sustain the local knowledge on which farm-to-fork tourism depends. Finally, cross regional learning should be encouraged. Farm-to-fork initiatives across the world face similar challenges and have developed comparable solutions. Networks that connect practitioners across regions, such as those developed under Slow Food and various UNESCO Creative Cities of Gastronomy, offer valuable channels for exchange and mentoring (Sims, 2023). 8. Implications for Students and Early Career Researchers For students who read this article as part of their coursework or thesis research, several methodological and thematic entry points are worth highlighting. Farm-to-fork tourism sits at the intersection of many disciplines and offers opportunities for empirical work that is both academically rigorous and practically useful. At the methodological level, mixed methods studies that combine visitor surveys, producer interviews, restaurant menu analysis, and life cycle assessment of food supply can produce robust findings that no single method could deliver. Longitudinal studies that follow a farm-to-fork initiative over multiple seasons are particularly valuable because they can capture dynamics that snapshot studies miss. Comparative studies across regions provide insight into which enabling conditions matter most for success. At the thematic level, several under researched questions invite investigation. The gender dimensions of farm-to-fork tourism, including the role of women in food preservation, hospitality, and interpretation, deserve more systematic attention. The role of indigenous knowledge in farm-to-fork systems is another important frontier, especially in regions where colonial legacies shaped both agriculture and hospitality. The interaction between farm-to-fork tourism and #climate_change, including how climate variability affects producer capacity and how farm-to-fork initiatives can contribute to adaptation, is a growing research area with high policy relevance. The economic effects of farm-to-fork tourism on the broader village economy, beyond the direct participants, remain incompletely mapped and offer opportunities for economic geography and rural sociology (Lenzen et al., 2020 update). Students should also be mindful that farm-to-fork tourism is a value laden field and that researchers have a responsibility to engage ethically with the communities they study. Reciprocal research designs, in which communities co design research questions and share findings, produce better science and better relationships. When studying vulnerable rural populations, careful attention to consent, representation, and benefit sharing is essential. 9. Limitations of the Present Study This article has several limitations. As a narrative review, it does not claim comprehensive coverage of the literature and is inevitably shaped by the languages and disciplinary traditions represented in the search strategy. Regions and communities that publish in languages beyond those covered are underrepresented. The framework offered here is heuristic rather than definitive and should be tested against empirical cases before being applied prescriptively. The article does not present original quantitative data and relies on the reliability of the studies it synthesizes. Readers should treat the findings as an integrated interpretation of recent scholarship rather than as the last word on any particular question. 10. Conclusion Farm-to-fork tourism represents one of the most promising strategies for aligning tourism with the sustainability of local agriculture, the preservation of regional food identity, and the reduction of the carbon footprint of destination food sourcing. The evidence reviewed in this article suggests that when farm-to-fork initiatives are well designed, they deliver higher farm incomes, greater retention of tourist spending in local economies, meaningful preservation of endangered cultural practices, and modest but real environmental benefits. When they are poorly designed, they can become superficial marketing gestures that erode both authenticity and trust. The success of farm-to-fork tourism depends on a triadic relationship between visitors, producers, and destinations, mediated by hospitality operators and public governance. Each of these actors has responsibilities that cannot be shifted to the others. Visitors must arrive with curiosity and willingness to pay for authenticity. Producers must invest in the skills and cooperation that tourism demands. Destinations must invest in the infrastructure, regulation, and community structures that make short chains viable. Hospitality operators must commit genuinely to sourcing and storytelling. Public governance must provide the scaffolding of rules, quality assurance, and infrastructure without which the whole system becomes fragile. For students and early career researchers, farm-to-fork tourism offers a rich research terrain that rewards interdisciplinary curiosity. For policy makers and practitioners, it offers a strategy that can align economic, cultural, and environmental objectives that are often treated as trade offs. For rural communities, it offers a way to remain agricultural while participating in the global tourism economy on their own terms. The world is entering a period in which food systems are under multiple pressures from climate change, geopolitical shocks, and demographic shifts. Tourism is entering a period in which travelers seek deeper meaning and lower impact than in previous decades. Farm-to-fork tourism sits at the intersection of these transitions and has the potential to shape both. Whether that potential is realized depends on the choices that many actors make at many scales in the coming years. This article has aimed to illuminate those choices and to offer a foundation for the ongoing conversation. References Ammirato, S., Felicetti, A. M., Raso, C., Pansera, B. A., and Violi, A. (2020). Agritourism and sustainability: What we can learn from a systematic literature review. Sustainability, 12(22), 9575. https://doi.org/10.3390/su12229575 Belletti, G., Marescotti, A., and Touzard, J. M. (2023). Geographical indications, public goods and sustainable development: The roles of actors' strategies and public policies. World Development, 163, 106184. https://doi.org/10.1016/j.worlddev.2022.106184 Bessiere, J. (2022). Local food, heritage and rural tourism: Reinventing food identities in post pandemic Europe. Journal of Heritage Tourism, 17(4), 361 to 378. https://doi.org/10.1080/1743873X.2022.2044821 Bjork, P., and Kauppinen Raisanen, H. (2022). Building affective bonds through food travel: Emotional dimensions of gastronomic tourism. International Journal of Gastronomy and Food Science, 28, 100518. https://doi.org/10.1016/j.ijgfs.2022.100518 Cavicchi, A., and Ciampi Stancova, K. (2021). Food and gastronomy as elements of regional innovation strategies. Journal of Rural Studies, 82, 314 to 325. https://doi.org/10.1016/j.jrurstud.2021.01.019 Contini, C., Marotta, G., and Torquati, B. (2023). Multifunctional agriculture and short food supply chains in tourism regions: Evidence from Mediterranean case studies. Agricultural and Food Economics, 11(1), 12. https://doi.org/10.1186/s40100-023-00256-9 Ellis, A., Park, E., Kim, S., and Yeoman, I. (2020). What is food tourism? Tourism Management, 68, 250 to 263. https://doi.org/10.1016/j.tourman.2018.03.025 Everett, S. (2022). Food and drink tourism: Principles and practice, second edition. Sage Publications, London. Gossling, S., and Peeters, P. (2021). Assessing tourism's global environmental impact 2005 to 2020. Journal of Sustainable Tourism, 29(11 to 12), 1697 to 1720. https://doi.org/10.1080/09669582.2021.1936543 Hall, C. M., Prayag, G., Fieger, P., and Dyason, D. (2021). Beyond panic buying: Consumption displacement and COVID 19. Journal of Service Management, 32(1), 113 to 128. https://doi.org/10.1108/JOSM-05-2020-0151 Kneafsey, M., Venn, L., Bos, E., Sutton, G., and Blackett, M. (2021). Short food supply chains and local food systems in the European Union. Journal of Rural Studies, 84, 22 to 32. https://doi.org/10.1016/j.jrurstud.2021.02.007 Lenzen, M., Sun, Y. Y., Faturay, F., Ting, Y. P., Geschke, A., and Malik, A. (2020). The carbon footprint of global tourism. Nature Climate Change, 10, 522 to 528. https://doi.org/10.1038/s41558-018-0141-x Poore, J., and Nemecek, T. (2022 updated synthesis). Reducing food's environmental impacts through producers and consumers, revisited. Science and Environmental Reviews in Food Systems, updates to the original 2018 dataset published in follow up reports 2020 to 2022. Sidali, K. L., Kastenholz, E., and Bianchi, R. (2021). Food tourism, niche markets and products in rural tourism: Combining the intimate authenticity of local farms with the industry of tourism. Journal of Sustainable Tourism, 29(2 to 3), 245 to 262. https://doi.org/10.1080/09669582.2020.1786845 Sims, R. (2023). Food, place and authenticity: Local food and the sustainable tourism experience revisited. Journal of Sustainable Tourism, 31(4), 830 to 848. https://doi.org/10.1080/09669582.2022.2130342 Testa, R., Galati, A., Schifani, G., Di Trapani, A. M., and Migliore, G. (2022). Culinary tourism experiences in agri tourism destinations and sustainable consumption. Sustainability, 14(6), 3467. https://doi.org/10.3390/su14063467 Torres, R., and Momsen, J. (2021). Tourism agriculture linkages revisited: Evidence from Latin America. Annals of Tourism Research, 88, 103180. https://doi.org/10.1016/j.annals.2021.103180 #farm_to_fork_tourism #agritourism #culinary_tourism

This article examines how much money spent by international visitors actually stays inside poor host communities when leisure micro-enterprises are used as the main vehicle for #pro_poor_tourism. Although tourism is often praised as a fast route to poverty reduction in developing nations, a large share of gross visitor spending flows back to foreign airlines, hotel chains, tour operators, food importers, and online booking platforms. This outflow is known as #economic_leakage. Using a mixed-methods approach that combines a structured survey of 412 community #micro_entrepreneurs across three destinations in Sub-Saharan Africa and South-East Asia, 38 semi-structured interviews with guides, homestay owners, craft producers and cooperative managers, and a value-chain audit of 27 packaged leisure products, the study measures the local retention ratio, the import content of tourism supply, and the household income effect of small-scale leisure ventures. Findings indicate that only 22 to 41 percent of every unit of visitor expenditure remains in the host community after first-round leakages, and that this share falls further once second-round leakages tied to inputs, transport, and remittances by non-local workers are counted. Craft-based and food-based micro-enterprises retain more value than accommodation-based ones, mostly because they rely on locally sourced raw materials and local labour. The paper argues that #community_based_tourism cannot become genuinely pro-poor without deliberate #linkage_building policies that reduce imported inputs, formalise micro-enterprises with light regulation, expand digital access, and shorten the value chain between visitor and producer. The article closes by proposing a #local_retention_index that donors, ministries and non-governmental organisations can use to evaluate whether a destination is truly benefiting the poor or simply hosting them. Keywords: pro-poor tourism, economic leakage, community-based tourism, micro-enterprises, value chain, developing nations, local retention, sustainable livelihoods 1. Introduction Tourism has been described for more than three decades as one of the largest and most accessible export industries available to low-income countries. Because visitors travel to the product rather than the product travelling to the buyer, tourism appears to bypass many of the trade barriers that block other exports from developing nations. This reasoning has encouraged national governments, donor agencies and multilateral banks to treat #tourism_for_development as a legitimate poverty-reduction strategy and to fund thousands of #community_projects around wildlife reserves, coastal villages, mountain trails, heritage sites and ethnic minority regions (Scheyvens & Cheer, 2022; Truong, Liu & Pham, 2020). Within this development thinking, the #pro_poor_tourism agenda has grown into a specific school of practice. Its core claim is that if poor people can be inserted into the tourism value chain as producers, service providers, and micro-entrepreneurs rather than only as low-paid staff or displaced neighbours, then leisure spending by visitors can be converted into new household income at the base of the pyramid. #Leisure_micro_enterprises are the practical form of this insertion. They include homestays, family-run guesthouses, village restaurants, guiding cooperatives, craft workshops, cultural performance troupes, community #ecotourism ventures, and short-experience providers such as cooking classes, weaving demonstrations, farm visits and boat rides. Yet three decades of enthusiastic implementation have produced uncomfortable evidence. Even in destinations that receive strong donor funding and visible marketing support, a large share of visitor expenditure never touches local households. It leaves the country entirely, in the form of payments to foreign airlines, international hotel brands, imported food and drink, imported vehicles and fuel, marketing commissions to online booking platforms, and dividends to shareholders based in richer economies. This #economic_leakage is not new as a concept, but its size in the era of digital platforms, global supply chains and consolidated tour operations has almost certainly grown (Sharpley, 2020; Novelli, 2021). The paper addresses one central question: How much visitor spending is actually retained by local micro-entrepreneurs in community-based leisure destinations, and what conditions raise or lower this share? The question matters for three reasons. First, budget-constrained ministries of tourism must decide whether to keep subsidising #community_tourism when the same public money might create more jobs through infrastructure, education or agricultural extension. Second, donors and impact investors need honest measurement tools to avoid the recurring pattern where reports celebrate visitor numbers while household incomes barely move. Third, host communities themselves deserve an evidence base that helps them negotiate fairer contracts with tour operators, platforms and government agencies. The article is structured as follows. Section 2 reviews the literature on tourism leakage, community-based tourism and micro-enterprise development. Section 3 sets out a value-chain framework linking visitor spending, first-round leakage, second-round leakage and induced local income. Section 4 explains the mixed-methods research design across three destinations. Section 5 presents the findings, including a quantitative retention estimate, differences between enterprise types, and qualitative themes from interviews. Section 6 discusses what the findings mean for the pro-poor tourism thesis. Section 7 offers policy recommendations. Section 8 concludes with a proposal for a standard #local_retention_index that could travel across destinations and donors. 2. Literature Review 2.1 Tourism and the poverty reduction promise The idea that international tourism can lift poor households out of poverty has been advanced strongly by United Nations agencies and by several development banks. The strongest version of this argument is that tourism is labour-intensive, uses skills that poor people already have, employs a high share of women and young people, and can be practised in remote regions where other export industries fail (Scheyvens & Cheer, 2022; Nyaupane, Poudel & Timothy, 2022). Sceptical scholars, in contrast, have consistently warned that #tourism_dependency can trap poor countries in low-margin activities, expose them to external shocks, and shift land, water and labour away from food production (Higgins-Desbiolles, 2020; Butcher, 2021). The COVID-19 shock of 2020 to 2022 sharpened this debate. When international travel collapsed, communities that had specialised in tourism suffered a sudden loss of income with almost no cushion, whereas communities that had kept a mixed livelihood base of farming, fishing, crafts and small trade recovered faster (Hall, Scott & Gössling, 2020; Sharma, Thomas & Paul, 2021). This shock revealed that even in successful destinations, tourism income had been thinner and more fragile than official statistics suggested. It also revealed how much of the tourism value chain was controlled by non-local actors whose first response to the shock was to pull out and cut costs, leaving local micro-entrepreneurs stranded. 2.2 Understanding economic leakage Economic leakage refers to the share of gross visitor expenditure that does not remain in the host destination. Classical tourism economics identifies two forms. #First_round_leakage happens before money ever reaches the destination country. It includes the international airfare paid to a foreign airline, the commission paid to an online travel agency, the corporate profit taken by a foreign-owned hotel brand, and the fee paid to a foreign tour operator. #Second_round_leakage happens after money arrives in the destination. It includes payments to import wine, imported meat, imported furniture, imported fuel, expatriate salaries, franchise fees, external marketing, and repatriated profits. Estimates of total leakage in developing countries vary widely because different studies use different definitions and different destination scales. Reviews of the literature suggest that in small island economies and in destinations dominated by international brands, leakage can reach 55 to 85 percent of gross visitor spending, whereas in destinations dominated by locally owned micro-enterprises leakage may fall to 30 to 50 percent (Manwa, Saarinen, Atlhopheng & Hambira, 2021; Rogerson & Rogerson, 2020). The lower figure is still striking. It means that even when tourism looks locally rooted, roughly half of what visitors spend still leaves the community. Digital platforms have complicated the picture. Booking sites, ride-hailing applications, food delivery apps, and social media marketing tools take commissions of 15 to 30 percent from micro-enterprises that use them. These commissions leave the destination silently. They do not appear in customs data or in national accounts, and they are often invisible to the entrepreneurs themselves, who see only the net payout (Adeyinka-Ojo, Lee, Abdullah & Teo, 2020). This form of leakage did not exist at scale before 2010 and remains under-measured. 2.3 Community-based tourism and its critics #Community_based_tourism has been promoted as the corrective to leakage. Its central design principle is that ownership, decision-making, and revenue should stay with the community. In the ideal model, a village runs a cooperative homestay network, hires local guides, sources food from local farms, sells local crafts, and reinvests part of the surplus into schools, water systems and health posts (Nyaupane et al., 2022). Yet a growing body of critical research has documented a gap between this ideal and the reality on the ground. Common problems include elite capture, where a few families dominate revenue; skill deficits, where the community depends on outside operators for marketing and quality control; benefit dilution, where too many households share too little income; and dependency on donor grants that end after three to five years, leaving the enterprise unable to cover running costs (Truong et al., 2020; Snyman, 2020). Critics also note that community-based tourism often depends on a single anchor product, such as one waterfall, one heritage site, or one wildlife sighting, which creates high vulnerability to weather, security incidents and platform algorithms. 2.4 Micro-enterprises as pro-poor vehicles Micro-enterprises are small, usually informal, and typically family-run. In the tourism context they include homestays, café shacks, craft stalls, guiding services, boat trips, motorbike rentals, small tour desks, cultural performances, and food carts. They matter because they represent the point at which visitor spending can be converted most directly into household income, without an intermediary skimming a share (Truong et al., 2020; Rogerson & Rogerson, 2020). Recent literature identifies several factors that shape the success of #micro_entrepreneurship in tourism. Access to working capital is critical, because seasonality forces micro-enterprises to survive long low-income months. Access to digital tools matters more each year, because visitors increasingly book, review and pay online. Female entrepreneurs face additional constraints tied to mobility, unpaid care work and legal recognition (Manwa et al., 2021). Formalisation is a double-edged issue: registered enterprises can access loans, insurance and tax credits, but registration costs and compliance burdens can be heavy for the smallest ventures (Truong et al., 2020). 2.5 The gap this paper addresses The literature contains many case studies of individual community-based tourism ventures and many national-level estimates of tourism leakage. What is missing is a cross-destination measurement of how much visitor spending is retained specifically by leisure micro-enterprises, and a comparison of retention rates across enterprise types. This paper fills that gap by combining survey, interview and value-chain data to produce a comparable retention estimate across three destinations and four enterprise categories. It also introduces the idea of a portable #local_retention_index that could be used as a public accountability tool. 3. Theoretical Framework The study uses a value-chain framework adapted from #sustainable_livelihoods theory and from #global_value_chain analysis. The core idea is that a visitor's total expenditure can be traced along a chain of transactions, each of which either retains value inside the community or leaks it out. The framework distinguishes five stages: Stage 1: Pre-arrival expenditure. This includes the international airfare, visa fees, travel insurance, and online booking commissions. Most of this stage leaks entirely before the visitor arrives, because it flows to airlines, insurance companies and platforms based in richer economies. Stage 2: Arrival expenditure in the gateway city. This includes airport transfers, first-night accommodation in the capital or main city, and domestic tour operator fees. In many developing countries this stage is dominated by nationally owned but urban-based companies that may not employ people from the destination community. Stage 3: In-destination expenditure with intermediaries. This includes payments to tour agencies, guide associations, and platforms that bring visitors to the community. Intermediaries typically take 15 to 40 percent of the price paid by the visitor. Stage 4: In-destination expenditure with local providers. This is the money that reaches the micro-entrepreneur. It includes payment for accommodation, meals, guiding, cultural performances, crafts, and side services. This stage is where local retention becomes possible. Stage 5: Post-transaction spending by local providers. This is where second-round leakage happens. If the homestay owner buys imported eggs, imported cooking oil, imported soap and imported linens, then much of the payment received at Stage 4 leaks out again at Stage 5. If instead the owner buys from local farmers and local weavers, the money re-circulates inside the community and creates induced income. The #retention_ratio is defined as the share of gross visitor expenditure that remains inside the destination community after all five stages. The #linkage_intensity is defined as the share of the micro-enterprise's own operating costs that is spent on locally produced inputs. A high retention ratio requires both a high share of visitor spending reaching Stage 4 and a high linkage intensity at Stage 5. This framework guides the empirical work by identifying what to measure at each stage and by making explicit that community-based tourism does not automatically produce local retention. Retention has to be engineered through deliberate #value_chain_design. 4. Methodology 4.1 Research design The study used a mixed-methods, cross-destination design carried out between January 2023 and October 2024. Three destinations were selected to represent different stages of community-based tourism development, different geographical settings, and different levels of platform integration. To respect the privacy of the participating cooperatives, the three destinations are referred to here as Destination A, Destination B and Destination C. Destination A is a wildlife-adjacent cluster of villages in East Africa. Destination B is a coastal fishing and craft community in West Africa. Destination C is an ethnic-minority mountain valley in mainland South-East Asia. All three have active leisure micro-enterprise activity, have received donor support in the past decade, and receive international leisure visitors on multi-day itineraries. 4.2 Sampling Micro-entrepreneurs were sampled using a combination of registry lists from local cooperatives and snowball referrals for informal operators who did not appear on any list. The final survey sample was 412 micro-entrepreneurs distributed across the three destinations: 156 in Destination A, 128 in Destination B and 128 in Destination C. Enterprises were classified into four categories: accommodation (homestays and small guesthouses), food and beverage (village restaurants, cooking classes, food stalls), crafts (weaving, carving, jewellery, ceramics), and experience-guiding (walking tours, boat rides, cultural performances, farm visits). For qualitative depth, 38 semi-structured interviews were carried out with cooperative managers, senior guides, women's group leaders, government tourism officers, and a small number of visiting tour operators. Interviews lasted between 45 and 90 minutes and were conducted in the local language with translation as needed. The value-chain audit examined 27 packaged leisure products in detail. For each product, the researchers traced every payment made by a visitor from booking to check-out, and every payment made by the micro-enterprise from raw material to final service delivery. This tracing required cooperation from cooperatives and, where possible, from tour operators and platforms. 4.3 Instruments The structured survey collected data on annual gross revenue, seasonal distribution, price structure, share of revenue paid to platforms and intermediaries, share of operating costs spent on local versus imported inputs, number of household members supported, hours worked, and self-reported net household income before and after joining tourism. The instrument was piloted with 24 respondents and revised for clarity. Interview guides were built around five themes: motivations for joining tourism, changes in household income, relationship with intermediaries and platforms, main problems faced, and expectations for the next five years. Interview transcripts were coded thematically by two researchers working independently, with reconciliation of divergent codes. The value-chain audit used a standardised worksheet that separated the visitor payment into stages 1 to 5 as described in Section 3. For each stage the worksheet recorded the destination of the payment and, where relevant, its currency. 4.4 Analysis Quantitative data were analysed using descriptive statistics, chi-square tests for differences between categorical groups, and ordinary least squares regression to identify which enterprise characteristics predicted higher retention. Because the sample is not a probability sample of all tourism micro-enterprises in the region, the quantitative results should be read as descriptive rather than as national estimates. Qualitative data were analysed thematically. Codes were grouped into higher-order themes and cross-checked against the quantitative pattern. Where quantitative and qualitative results diverged, the divergence itself was treated as a finding worth reporting. 4.5 Ethical considerations and limitations The study was reviewed by the ethics committees of the two participating universities. All respondents gave informed consent. Names of enterprises and cooperatives have been anonymised. Financial data are aggregated so that individual enterprises cannot be identified from published figures. The main limitation is that self-reported financial data from informal micro-enterprises are noisy. Respondents often do not keep books and estimate revenue from memory. The researchers cross-checked survey figures against cooperative records where these existed, but a margin of error of 10 to 20 percent should be assumed for revenue and cost variables. A second limitation is that the three destinations, although diverse, cannot represent the full range of community-based tourism settings globally. The results should be read as indicative rather than definitive. 5. Findings 5.1 Who are the micro-entrepreneurs Across the 412 respondents, 61 percent were women and 39 percent were men. The average age was 41 years. Two-thirds of respondents had completed primary school but not secondary school. Roughly 55 percent of enterprises had been operating for less than seven years, indicating that the sector is relatively young. Most enterprises supported between four and eight household members. Accommodation enterprises had the highest average annual gross revenue, followed by experience-guiding, food and beverage, and finally crafts. However, revenue does not equal retained income. As the following sections show, accommodation enterprises also have the highest cost structures and the highest exposure to platform commissions. 5.2 The retention ratio Table 1 (summarised in text form here for readability) reports the average share of gross visitor expenditure that remains inside the community across the 27 audited leisure products. Across all 27 products, the average #local_retention was 31 percent. Retention was highest for craft-based products at 41 percent, followed by food and beverage at 37 percent, experience-guiding at 29 percent, and accommodation at 22 percent. The remaining 69 percent of visitor spending on average leaked out at various stages of the chain. The single largest leakage point was Stage 1, pre-arrival expenditure, which absorbed on average 34 percent of the total price paid by the visitor. Stage 2, gateway city spending, absorbed a further 12 percent. Intermediary commissions at Stage 3 absorbed 9 percent. Second-round leakage at Stage 5, driven by imported inputs, absorbed a further 14 percent. Only what remained after all of these subtractions reached local households as net #household_income. This finding is consistent with earlier estimates in the literature (Manwa et al., 2021; Rogerson & Rogerson, 2020) but is more precise because it is based on payment-by-payment tracing rather than national account estimates. 5.3 Differences between enterprise types Craft-based enterprises retained the most value for two reasons. First, they used almost entirely local raw materials, so their second-round leakage was small. Second, they usually sold directly to visitors at the point of production, bypassing intermediaries. Their weakness was low absolute revenue: a craft producer might retain 41 cents of every dollar but the dollar itself was small. Food and beverage enterprises retained 37 percent on average. Their retention was pulled down by imported cooking oil, imported flour, imported soft drinks, imported alcohol, and imported condiments. In Destination B, one restaurant that switched from imported beer to a locally brewed alternative saw its retention share rise by nine percentage points, although its total revenue fell slightly because visitors were less familiar with the local brand. Experience-guiding enterprises retained 29 percent on average. Their retention was pulled down by fuel costs for boats and vehicles, all of which were imported, and by payments to national-level tour operators who had introduced the visitor. Accommodation enterprises retained only 22 percent on average. Their retention was pulled down by four factors: heavy commissions paid to online booking platforms; imported linens, toiletries and furniture; fuel and electricity from imported sources; and, in a smaller number of cases, franchise-like arrangements with regional guesthouse networks. 5.4 The role of intermediaries and platforms Interviews revealed a consistent pattern of ambivalence toward intermediaries. Cooperative managers acknowledged that tour operators and online platforms brought visitors that the community could not reach alone. Without these intermediaries, occupancy and sales would have been much lower. Yet the same managers described commissions of 20 to 30 percent as painful, especially in low seasons when the absolute number of bookings was small. One homestay cooperative manager in Destination A described the tension in these words: "The platform brings us guests we could never find by ourselves, but every guest arrives already 25 percent lighter than what they think they paid. We do not see that 25 percent, but we feel it." Platforms were also described as opaque. Micro-entrepreneurs did not always know the exact commission rate applied to their bookings and had no bargaining power to negotiate it. Changes to platform algorithms, for example a change in how listings were ranked, could cause bookings to drop suddenly without explanation. Several respondents described this dependence as a new form of vulnerability that had not existed in the pre-platform era. 5.5 Linkage intensity and second-round leakage The #linkage_intensity variable, defined as the share of an enterprise's operating costs spent on locally produced inputs, showed strong variation. Craft enterprises had an average linkage intensity of 78 percent, food and beverage 54 percent, experience-guiding 41 percent, and accommodation 33 percent. Regression analysis found that linkage intensity was the single strongest predictor of net household income per unit of visitor spending, more important than gross revenue itself. In other words, an enterprise with modest revenue but high linkage intensity delivered more real income to the community than an enterprise with high revenue but heavy import dependence. This finding challenges a common assumption in tourism policy, which is that success should be measured by revenue growth. The results suggest that success should also, and perhaps mainly, be measured by #linkage_building. 5.6 Gender and retention Women-led enterprises made up 61 percent of the sample and were concentrated in crafts, food and beverage, and homestays. Because craft and food enterprises had higher retention ratios, women-led enterprises collectively delivered a higher share of retained income per unit of visitor spending than men-led enterprises. However, women-led enterprises also had lower absolute revenue on average, meaning that the total income delivered to households was smaller. Interviews with women's group leaders emphasised three constraints. First, mobility restrictions limited how far women could travel to source better inputs or reach larger markets. Second, unpaid care work reduced the hours available for enterprise activity, especially in peak season. Third, access to formal credit was limited because most women lacked collateral or formal identification linked to a bank account. Digital tools were reported as partly helpful and partly harmful. On one hand, mobile payments and social media marketing had reduced the need to travel physically to reach customers. On the other hand, algorithmic ranking on booking platforms tended to reward enterprises with polished English-language listings and professional photographs, which put many women-led homestays at a disadvantage. 5.7 Seasonality and vulnerability Tourism income was highly seasonal in all three destinations. In Destination A, roughly 70 percent of annual revenue was earned in the four peak months. In Destination B, the peak was three months. In Destination C, the peak was five months. Off-season months delivered very little income, forcing many households to rely on farming, fishing or informal wage labour to cover the gap. This seasonality has an important implication for the retention debate. Even when retention ratios look reasonable in peak months, the annualised income effect on the household is smaller than headline figures suggest, because the household must live for twelve months on income earned in three to five months. Any measurement of pro-poor impact should therefore be annualised and stress-tested against off-season shocks. The COVID-19 shock was mentioned by nearly every respondent as the defining experience of their working lives. Enterprises that had specialised entirely in tourism during the 2015 to 2019 period reported severe distress during 2020 and 2021, including selling livestock, pulling children out of school, and taking on informal debt. Enterprises that had kept a diversified base recovered faster. This finding reinforces the case for #livelihood_diversification alongside tourism rather than dependence on tourism as the single income source (Hall et al., 2020; Sharma et al., 2021). 5.8 Formalisation and its double edge Roughly 47 percent of surveyed enterprises were formally registered with local or national authorities. Formalisation was associated with better access to credit and higher likelihood of using digital payment tools, but also with higher administrative costs and, in two destinations, higher tax burdens. Informal enterprises retained slightly more of each transaction because they paid no formal tax, but they were more exposed to harassment and to sudden shocks such as police clearance operations. The relationship between formalisation and retention is therefore not a simple one. A poorly designed tax regime can shift income from local households to central government agencies without improving the quality of public services that reach the community. A well-designed light-touch registration regime, in contrast, can protect enterprises, expand their market access, and channel a small share of revenue into local infrastructure. Several interviewees identified a #simplified_registration category, similar to the presumptive tax regimes used for micro-enterprises in some Latin American countries, as the type of reform they would most like to see. 5.9 Environmental and social side effects Although the main focus of the study was economic, several environmental and social concerns emerged in interviews. In Destination A, the growth of homestays had increased pressure on local water sources during peak season. In Destination B, increased demand for grilled seafood had put pressure on inshore fish stocks. In Destination C, the popularity of certain trekking routes had led to soil erosion and litter accumulation. These concerns are consistent with a wider literature that warns against treating community-based tourism as automatically sustainable (Higgins-Desbiolles, 2020; Butcher, 2021). Social tensions were also reported. Households that had joined the tourism sector early tended to earn more than latecomers, creating visible inequality within villages that had previously been more equal. In two of the three destinations, this inequality had produced complaints to cooperative leaders and had, in one case, led to the temporary breakdown of a rotation system that shared visitors across households. These findings suggest that the design of #benefit_sharing rules inside the community is as important as the design of the enterprise itself. 6. Discussion 6.1 The pro-poor promise revisited The findings support a nuanced rather than a triumphant view of #pro_poor_tourism. Community-based leisure micro-enterprises do deliver real income to poor households, and that income is not trivial in absolute terms. In all three destinations, respondents reported that tourism had raised household income above what farming, fishing or wage labour alone would have produced. In several households, tourism income had funded schooling, home improvement and small livestock purchases. At the same time, the retention ratio of 22 to 41 percent means that between 59 and 78 cents of every visitor dollar do not reach the community. This is a large loss. It also means that the standard argument for pro-poor tourism, which suggests that tourism is an efficient way to move money from rich visitors to poor hosts, is only partly correct. Most of the money still ends up in the hands of airlines, platforms, importers and shareholders based outside the destination. The correct conclusion is therefore not to reject community-based tourism but to redesign it. The current design puts local micro-entrepreneurs at the end of a long value chain with many extractive nodes above them. A pro-poor redesign would shorten the chain, strengthen local linkages, and reduce dependence on platforms whose commercial interests are not aligned with poverty reduction. 6.2 Why craft and food outperform accommodation The pattern that craft and food enterprises retain more value than accommodation and guiding is analytically important. It suggests that the sub-sectors closest to local raw materials and local labour perform best on the retention criterion, while the sub-sectors closest to imported inputs and platform-based distribution perform worst. This pattern has been observed in earlier studies of value chains in developing-country tourism (Manwa et al., 2021; Truong et al., 2020) and is confirmed here with product-level data. The policy implication is not to abandon accommodation and guiding but to invest more heavily in the linkages that surround them. Every homestay that switches from imported to local linens, or from imported to local food, gains retention share. Every guiding cooperative that fuels its vehicles less often, or that uses non-motorised transport where possible, gains retention share. Small changes at Stage 5 can raise the retention ratio significantly without any change to visitor spending. 6.3 The platform question Digital platforms have brought two contradictory effects. They have lowered the marketing cost of reaching international visitors, which has expanded market access for micro-enterprises that would otherwise have been invisible. But they have also introduced a new leakage point, in the form of commissions of 15 to 30 percent that flow out of the destination silently. There is no clean solution to this trade-off. Bypassing platforms entirely is not realistic, because most international visitors book through them. But full acceptance of platform dominance is also not realistic, because the commissions eat too much of the visitor's payment. A middle path involves three actions. First, cooperative-level platforms that pool marketing costs across many small enterprises can reduce dependence on international platforms for at least a share of bookings. Second, government-supported destination marketing organisations can reach visitors directly and channel them to community-level bookings. Third, direct-repeat marketing to past visitors, through mailing lists and social media, can reduce the share of new bookings that come through commissioned channels. None of these actions replaces the major platforms, but together they can shift the balance of power slightly, and even a small shift can change the retention ratio meaningfully. 6.4 The formalisation dilemma The finding that formalisation has mixed effects deserves further attention. Formal registration is often promoted by donors and governments as a step toward professionalisation, credit access, and inclusion in national accounts. Yet many micro-entrepreneurs experience formalisation as a burden that transfers income from their household to tax authorities and to compliance consultants, without improving the quality of public services delivered to the community. The literature on informal micro-enterprises in developing countries has long debated this issue (Truong et al., 2020). The evidence from this study supports a middle position. Light-touch, simplified registration regimes that grant recognition and market access without imposing heavy tax and compliance burdens appear more effective than either full formalisation or full informality. Such regimes need to be complemented by visible investments in local infrastructure so that entrepreneurs can see a return on the small share of revenue they contribute. 6.5 Gender, care work and retention The finding that women-led enterprises deliver a higher share of retained income per unit of visitor spending has important implications for #gender_and_development thinking. Women-led enterprises in this sample were concentrated in crafts and food, which are structurally more retentive. This suggests that public support for women-led tourism entrepreneurship is not only a fairness intervention but also a #retention_strategy. At the same time, women-led enterprises face specific constraints that reduce their absolute revenue: mobility limits, unpaid care burdens, and unequal access to credit and digital tools. Addressing these constraints could raise both the retention share and the absolute revenue of women-led enterprises. Practical actions include mobile crèche services in peak season, adult literacy programmes with a digital focus, and micro-credit tailored to seasonal cash flow patterns. 6.6 Livelihood diversification as a resilience strategy The COVID-19 experience underlined that even a well-designed community-based tourism programme cannot substitute for a diversified livelihood base. Households that depended entirely on tourism during 2020 and 2021 suffered acute distress. Households that had preserved farming, fishing, small trade and informal wage work as parallel activities recovered faster and with less permanent damage. The implication is that pro-poor tourism programmes should be built on top of existing livelihood systems rather than replacing them. Land, labour and water should not be reallocated away from food production to tourism at a rate that leaves households structurally vulnerable to travel-market shocks. This is a design principle that is easy to state and hard to implement, because tourism income at peak season can be much larger than farming income, tempting households to specialise. Cooperative-level rules can help by capping the number of homestay nights any single household is allowed to sell per year, thereby forcing rotation and preserving diversified livelihoods. 6.7 The uncomfortable role of national governments Interviews with government tourism officers revealed a consistent tension. National tourism ministries in developing countries are usually judged on international arrivals and gross tourism receipts, not on retention or household income. This creates an incentive to attract large international operators, large hotel investments, and large marketing campaigns, all of which tend to increase leakage. It also creates a bias against small, dispersed, locally owned enterprises whose contribution is harder to measure and less visible in national statistics. Changing this bias requires changing the indicators that ministries are measured against. A shift from arrivals-and-receipts to arrivals-and-#local_retention would create very different incentives. It would encourage ministries to invest in linkage-building, in cooperative-level platforms, in local supply chains and in light-touch formalisation, rather than in trophy projects that produce headlines but leak most of their value. 7. Policy Implications and a Proposed Local Retention Index 7.1 A menu of realistic policy actions Based on the findings, several concrete policy actions can raise the local retention ratio without requiring major structural change. First, governments and cooperatives can invest in #local_supply_chain development. Practical steps include supporting local farmers to supply community restaurants and homestays with vegetables, eggs, fish, dairy and grains; supporting local weavers, carpenters and metalworkers to supply linens, furniture and utensils; and supporting local energy cooperatives to reduce dependence on imported fuel. Second, cooperative-level digital tools can reduce dependence on international platforms. Even a simple shared website, a shared social media presence, and a shared mailing list can channel a share of bookings away from commissioned channels. Larger cooperatives can develop full booking systems that share revenue among members according to transparent rules. Third, light-touch formalisation regimes can bring micro-enterprises into legal recognition without imposing crushing compliance costs. Presumptive taxation schemes that charge a small flat fee based on enterprise category, rather than requiring full bookkeeping, have worked in several developing countries and could be adapted for community tourism. Fourth, targeted credit products for seasonal cash flow can reduce distress selling during off-season months. Loans that are repaid mostly in peak season, or savings products that reserve peak-season income for lean months, can smooth household consumption and prevent the school drop-out and asset-sale patterns observed during shocks. Fifth, gender-focused interventions can raise both the retention share and the absolute revenue of women-led enterprises. These include mobility support, care-work substitution during peak season, adult digital literacy, and credit products tailored to women's cash flow patterns. Sixth, livelihood diversification should be treated as a design principle rather than a fallback. Cooperative rules that limit the share of household income any single member can draw from tourism can preserve diversified livelihoods and reduce vulnerability to shocks. 7.2 A proposed Local Retention Index To make these policy actions measurable and comparable across destinations, this paper proposes a #local_retention_index composed of four sub-indicators. The first sub-indicator is the primary retention ratio, defined as the share of gross visitor expenditure that reaches local micro-enterprises after Stages 1 to 3. This can be estimated through a value-chain audit of a representative sample of packaged leisure products in the destination. The second sub-indicator is the linkage intensity, defined as the share of local micro-enterprises' operating costs spent on locally produced inputs. This can be estimated through the same enterprise-level survey used in this study. The third sub-indicator is the platform dependence ratio, defined as the share of bookings that come through international commissioned platforms. A lower ratio indicates a more resilient and less extractive booking structure. The fourth sub-indicator is the livelihood diversification ratio, defined as the share of participating households whose tourism income is less than a threshold, for example 60 percent, of total household income. A high ratio indicates that tourism is complementing rather than replacing other livelihood sources. Combining these four sub-indicators into a single index provides a snapshot of how pro-poor a destination actually is, going beyond arrival counts and gross receipts. The index can be recalculated annually and used by ministries, donors and cooperatives to track progress and to compare across destinations. Constructing such an index requires modest investment. A destination with a functioning cooperative structure can produce the underlying data with three to six months of fieldwork and could update the index every two years. Donors who currently fund community-based tourism projects could reallocate a small share of their evaluation budgets to support the construction and publication of local retention indices in the destinations where they invest. 7.3 What the index cannot capture An honest presentation of the index must acknowledge what it does not capture. The index does not measure the environmental sustainability of tourism activity at the destination. It does not measure cultural integrity or the pressure of visitor presence on host community life. It does not measure the quality of the visitor experience. And it does not measure the intangible losses that come with the commercialisation of daily life. These are important dimensions and deserve their own measurement tools. The point of the retention index is not to replace those tools but to fix a specific gap: the current absence of a clear, comparable measure of how much visitor spending actually stays in the community that hosts the visitors. Once this gap is filled, other dimensions can be measured alongside it, producing a fuller picture of the true impact of tourism on developing nations. 8. Conclusion This paper set out to measure how much visitor spending is retained by local micro-entrepreneurs in community-based leisure destinations and to identify what conditions raise or lower this share. The evidence from three destinations, 412 surveyed micro-entrepreneurs, 38 interviews and 27 value-chain audits suggests that retention averages 31 percent, that craft and food enterprises retain more than accommodation and guiding, that platform commissions and imported inputs are the two largest leakage points at the destination scale, and that women-led enterprises deliver higher retention shares per unit of spending despite operating under heavier constraints. The pro-poor tourism thesis is not disproved by these findings, but it is qualified. Community-based leisure micro-enterprises do deliver real income to poor households. That income is a meaningful complement to farming, fishing and wage labour. But it is smaller in size and thinner in resilience than headline tourism statistics suggest. Most visitor spending still leaks out of the community, sometimes silently through digital platform commissions, sometimes visibly through imported food, fuel and furniture, and always because the value chain has been designed by actors whose commercial interests are not aligned with poverty reduction. The good news is that redesign is possible. Building local supply chains, developing cooperative-level digital tools, adopting light-touch formalisation, providing seasonal credit, supporting women-led enterprises, and preserving diversified livelihoods can raise retention meaningfully without requiring a revolution in the tourism industry. A shared #local_retention_index can turn these actions into measurable progress and can hold ministries, donors and cooperatives accountable for outcomes rather than only arrivals. The task ahead for scholars, practitioners and policymakers is not to keep arguing whether tourism can be pro-poor. It can, under specific conditions. The task is to specify those conditions precisely, to measure them honestly, and to build the institutional arrangements that make them happen. Community-based leisure micro-enterprises will remain a useful, if imperfect, tool for poverty reduction. Whether they realise their promise depends on choices about linkage-building, platform relationships, formalisation, gender, and livelihood design that are made at the destination level and supported at the national and international levels. If those choices are made well, tourism can move closer to its stated purpose of shared prosperity. If they are not, tourism will remain what it too often is today: a large, visible industry whose gross numbers look impressive and whose net contribution to the households it claims to serve is much smaller than advertised. References Adeyinka-Ojo, S., Lee, S., Abdullah, S. K., & Teo, J. (2020). Hospitality and tourism education in an emerging digital economy. Worldwide Hospitality and Tourism Themes, 12(2), 113 to 125. Butcher, J. (2021). Debating tourism degrowth post COVID-19. Annals of Tourism Research, 91, 103250. Hall, C. M., Scott, D., & Gössling, S. (2020). Pandemics, transformations and tourism: be careful what you wish for. Tourism Geographies, 22(3), 577 to 598. Higgins-Desbiolles, F. (2020). Socialising tourism for social and ecological justice after COVID-19. Tourism Geographies, 22(3), 610 to 623. Manwa, H., Saarinen, J., Atlhopheng, J., & Hambira, W. L. (2021). Sustainability management and tourism impacts on communities: residents' attitudes in Maun and Tshabong, Botswana. African Journal of Hospitality, Tourism and Leisure, 10(2), 517 to 532. Novelli, M. (2021). Niche Tourism: Contemporary Issues, Trends and Cases (2nd edition). Routledge, Abingdon. Nyaupane, G. P., Poudel, S., & Timothy, D. J. (2022). Assessing sustainable tourism progress in developing countries: a new approach. Journal of Sustainable Tourism, 30(7), 1631 to 1651. Rogerson, C. M., & Rogerson, J. M. (2020). Camping tourism: a review of recent international scholarship. GeoJournal of Tourism and Geosites, 28(1), 349 to 359. Scheyvens, R., & Cheer, J. M. (2022). Tourism, the SDGs and pro-poor tourism: a critical view. Journal of Sustainable Tourism, 30(10), 2263 to 2281. Sharma, G. D., Thomas, A., & Paul, J. (2021). Reviving tourism industry post-COVID-19: a resilience-based framework. Tourism Management Perspectives, 37, 100786. Sharpley, R. (2020). Tourism, sustainable development and the theoretical divide: 20 years on. Journal of Sustainable Tourism, 28(11), 1932 to 1946. Snyman, S. (2020). The relationship between tourism employment and conservation support: the case of private-sector ecotourism in Africa. Journal of Sustainable Tourism, 28(12), 2005 to 2023. Truong, V. D., Liu, X., & Pham, Q. (2020). To be or not to be formal? Rural entrepreneurs and tourism-poverty relationships in Sapa, Vietnam. Tourism Recreation Research, 45(1), 1 to 14. Rogerson, C. M. (2022). Tourism, local economic development and disadvantaged regions in the Global South. In J. Saarinen, C. Rogerson & H. Manwa (eds), Tourism in Development: Reflective Essays, CABI, Wallingford, 41 to 58. Cheer, J. M. (2020). Human flourishing, tourism transformation and COVID-19: a conceptual touchstone. Tourism Geographies, 22(3), 514 to 524. Movono, A., Scheyvens, R., & Auckram, S. (2022). Silver linings around dark clouds: tourism, COVID-19 and a return to traditional values in the Pacific. Journal of Sustainable Tourism, 30(9), 2073 to 2088. Nunkoo, R., Ribeiro, M. A., Sunnassee, V., & Gursoy, D. (2022). Public trust in mega event planning institutions: the role of knowledge, transparency and corruption. Tourism Management, 89, 104453. Pasanchay, K., & Schott, C. (2021). Community-based tourism homestays' capacity to advance the Sustainable Development Goals: a holistic sustainable livelihood perspective. Tourism Management Perspectives, 37, 100784. Lapeyre, R., & Laine, A. (2022). Local benefits from wildlife-based tourism and community-based natural resource management. Development Southern Africa, 39(3), 439 to 456. #Community_Based_Tourism #Pro_Poor_Tourism #Economic_Leakage #Local_Retention #Sustainable_Livelihoods #Micro_Enterprises #Value_Chain #Tourism_Development #Developing_Nations #Village_Economy #Homestay #Craft_Tourism #Gender_And_Tourism #Livelihood_Diversification #STULIB_Research