Generative artificial intelligence is quietly rewriting the rules of who counts as a student, and a major new review argues that the technology’s deepest effects on educational inequality have little to do with access to information at all. In a critical integrative review published in the journal AI & Society, Steven Watson of the University of Cambridge, Christian Morgner of the University of Portsmouth and Erik Brezovec of the University of Zagreb contend that large language models such as ChatGPT act less as neutral tutoring tools than as “meaning-mediating infrastructure” — socio-technical systems that reshape the very communicative forms through which learners become recognisable, supported and judged in education. The authors introduce a new concept, semantic transduction, to describe how these systems reformat prompts, rubrics, feedback and disciplinary genres into communicatively plausible forms, and they warn that this process can either widen participation or entrench exclusion depending on how it is embedded across classrooms, households, platforms and policy regimes.
The review positions itself against two dominant and opposing narratives that have framed public debate since generative AI burst into classrooms in late 2022. The first is optimistic: AI will “level the playing field” by making high-quality explanation, feedback and tutoring universally available, a view echoed in sector guidance from UNESCO and the OECD. The second is alarmist: AI will deepen inequality by privileging those with better devices, stronger digital skills and more supportive home environments, while shifting new burdens of verification and judgement onto learners and educators. The authors accept that both narratives capture real dynamics but argue that both compress inequality into an outcomes problem, measured in scores and progression, or an access problem, measured in devices and connectivity. Decades of digital inequality research, they note, have shown that differences in use, skill, institutional mediation and the social organisation of support matter just as much — and something still more fundamental is being missed.
That missing dimension, the review argues, is communication itself. Drawing on Niklas Luhmann’s sociological systems theory, the authors contend that educational inequality is produced through the everyday communicative processes by which learners are addressed, categorised and recognised as participants. Communication does not merely describe a learner who is already educationally present; it creates the address through which a person becomes relevant as a learner, author, candidate, support recipient or suspected cheater. Inclusion, in this framing, is not a settled policy achievement but an ongoing communicative accomplishment. A deaf learner, for example, is not excluded by deafness alone but by educational communication that presupposes hearing as the normal route to participation; in a setting organised through writing, captions and visual materials, the same bodily condition is configured differently. When the dominant forms of communication are inaccessible or narrowly coded, exclusion is readily attributed to individual ability even though it is partly manufactured by the structure of address itself.
To capture what generative AI actually does to these communicative structures, the authors develop the concept of semantic transduction. Large language models generate text by modelling statistical regularities across vast corpora and predicting likely continuations given a prompt and interaction history. They do not understand in a human sense, yet they produce fluent, contextually responsive and rhetorically convincing output that can participate in educational communication as a quasi-interaction partner — suggesting next steps, modelling genres, simulating feedback and rephrasing instructions in real time. Semantic transduction names the socio-technical process by which such systems reformat an educational demand, whether a prompt, assignment brief, rubric, exam question or policy rule, into a communicatively plausible form. Crucially, what changes is not merely wording but the admissibility of a contribution within a recognitive setting: something becomes easier, harder or different to process as an educational contribution. The term deliberately differs from translation, scaffolding and genre modelling because it can support learning, bypass learning, standardise expression or misrecognise learners depending on how it is embedded.
A concrete example from the review illustrates the double edge. A multilingual student receives an assignment brief asking for a critical discussion of a historical event. The barrier is not simply lack of information but the difficulty of translating partial understanding into an institutionally recognised genre. Pasting the brief into a large language model yields a possible structure, key terms, paragraph moves and alternative framings — semantic transduction lowering the cost of participation. Yet the same process can narrow possibilities by steering the learner towards mainstream argumentative templates, standardised academic tone and forms of evidence more easily supported by the model than by the student’s own cultural or linguistic repertoire. Recognition risks becoming conditional on assimilation. The authors stress that plausibility is not validity: language models can produce authoritative-sounding prose without evidence, invent citations and present contested issues as settled, so their educational value is inseparable from practices of verification — and those practices are unevenly distributed.
Synthesising scholarship from inclusive education, digital inequality, critical edtech, science and technology studies, philosophy of technology and systems theory, the review identifies three recurring sites where AI reorganises inclusion and exclusion. The first is access and capability divides. Access now includes not only devices and connectivity but institutional permission, paid subscriptions, language resources, time for safe experimentation and trust that AI use will not be automatically treated as misconduct. Capability includes the interactional skill of formulating prompts, iterating and evaluating outputs — capacities that reward metacognitive control and are often bolstered by family resources and school cultures. From an inclusion perspective, the sharpest question is not who can use AI but who can be recognised as legitimately using it: in some settings AI use is framed as cheating, in others as normal productivity, so learners under tighter surveillance may be excluded through suspicion even when their purposes are identical.
The second site is misrecognition through templates of good performance. When learners rely on AI-generated exemplars, or when educators increasingly expect them, educational communication can drift towards surface conformity. Because recognition in schooling is closely tied to language and genre, AI templates can function as a new form of cultural capital, supplying the right tone, structure and rhetorical pacing for particular assessments. This can support learners otherwise excluded by unfamiliar genres, but it can also re-entrench dominant Anglo-European academic norms and nudge learners away from culturally specific framings or epistemic traditions from the Global South. Parallel risks arise in disability and neurodiversity contexts, where AI can enable alternative expression and self-advocacy while simultaneously stabilising narrow norms of what an articulate, organised or “appropriate” learner sounds like. The documented bias of AI text detectors against non-native English writers — findings published in the journal Patterns showing that GPT detectors systematically misflag non-native writing — gives this concern particular urgency.
The third site is normative drift in pedagogy, authorship and assessment. As AI becomes embedded in ordinary workflows, feedback tone, task design, criteria of originality and assumptions about polish may all shift towards what AI makes easy to produce and easy to evaluate. This drift is rarely intentional. Teachers adopt AI to reduce workload; leaders introduce detection tools in the name of integrity; vendors add AI functions as default features. Yet together these pragmatic responses recalibrate what counts as normal, efficient and credible communication. If polished prose becomes the new baseline, learners who most benefit from supportive translation and drafting face both rising expectations and heightened suspicion — a “levelling down” effect in which attempts to equalise advantage those already privileged. Survey evidence underscores how fast the ground is moving: the 2026 HEPI/Kortext survey indicates very high levels of generative AI use among UK undergraduates, while training, policy clarity and confidence in appropriate use remain uneven.
Against these patterns, the review proposes “processual inclusion” — treating inclusion and exclusion not as binary outcomes but as ongoing accomplishments that can be observed early and renegotiated before trajectories harden. GenAI can genuinely widen participation when it helps learners translate ideas into recognised genres, serves as a dialogic partner for self-explanation, provides accessibility supports such as summaries, stepwise instructions and planning aids, or substitutes for paid tutoring that disadvantaged students cannot afford. A student with dyslexia may use AI to simplify dense text before returning to the original; a student with ADHD may break an extended task into stages; a multilingual learner may rehearse questions or draft an idea for revision. But these uses are inclusion-enhancing only when institutions also learn from the barriers that made the mediation necessary. The most exclusionary pattern, the authors warn, is not students using AI but institutions implicitly raising expectations about pace, polish and idiomatic fluency while simultaneously tightening authenticity regimes.
The policy conclusion is that individualised AI literacy, however necessary, is not enough. Framing AI literacy purely as personal competence — prompt better, check facts, avoid plagiarism — pushes responsibility for inclusion onto learners and teachers while the conditions of AI-mediated communication are set by platform design, procurement and institutional policy. Instead, the authors propose an infrastructural governance agenda across four layers: policy transparency, with legible rules distinguishing formative support, accessibility support, drafting, translation and assessment evidence rather than treating all AI use as one category; inclusive design and procurement, treating AI systems as pedagogical infrastructure whose accessibility, language coverage, data practices, bias and auditability are contractual concerns; participation and redress, ensuring that students, teachers and families have channels to report harms and contest classifications, and that detection signals are never treated as conclusive evidence of misconduct; and iterative audit and repair, monitoring whether AI policies differentially affect multilingual learners, disabled learners, first-generation students and those with limited access to paid tools. The EU AI Act’s designation of certain educational AI uses as high-risk signals, the authors note, growing legislative recognition of these stakes.
Ultimately, the review reframes the question educators and policymakers should be asking. The issue is not whether generative AI is inclusive or exclusive in the abstract, but how these socio-technical arrangements change the communicative conditions under which inclusion and exclusion are produced — and whether institutions build re-entry and repair into AI-mediated education before selections harden into unanswerable infrastructure. AI-mediated education, the authors argue, is a coupled ecology of classroom routines, household capacities, platform interfaces, assessment programmes and governance protocols, and no single actor controls it. The task is therefore not control in any strong sense but the cultivation of repairable infrastructures: arrangements able to absorb feedback, revise distinctions and redistribute burdens of verification before exclusions become permanent. Future research, they suggest, should examine how semantic transduction operates differently across subjects, languages, assessment forms and national policy regimes, and how societies might shape algorithmic infrastructures rather than merely adapting learners to them.
Subject of Research: The role of generative AI as meaning-mediating infrastructure in education, and how it reorganises inclusion, exclusion and educational inequality through communicative processes such as semantic transduction.
Subject of Research: Technology and Engineering
Article Title: Generative AI as meaning-mediating infrastructure in education: a critical integrative review of inclusion, exclusion, and semantic transduction
Article References: Watson, S., Morgner, C., & Brezovec, E. (2026). Generative AI as meaning-mediating infrastructure in education: a critical integrative review of inclusion, exclusion, and semantic transduction. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03221-4
Image Credits: AI Generated
DOI: 10.1007/s00146-026-03221-4
Keywords: Generative AI, Educational inequality, Inclusion/exclusion, Semantic transduction, Critical integrative review, Systems theory, AI governance, AI literacy, Meaning-mediating infrastructure, Assessment, Digital inequality
Cite Scienmag News
APA MLA Chicago
Denise Maddox. (September 5, 2026). Generative AI reshapes educational meaning: a critical review of inclusion and exclusion. Scienmag. https://scienmag.com/generative-ai-reshapes-educational-meaning-a-critical-review-of-inclusion-and-exclusion/
Denise Maddox. “Generative AI reshapes educational meaning: a critical review of inclusion and exclusion.” Scienmag, 5 September 2026, https://scienmag.com/generative-ai-reshapes-educational-meaning-a-critical-review-of-inclusion-and-exclusion/. Accessed 5 September 2026.
Denise Maddox. “Generative AI reshapes educational meaning: a critical review of inclusion and exclusion.” Scienmag. September 5, 2026. https://scienmag.com/generative-ai-reshapes-educational-meaning-a-critical-review-of-inclusion-and-exclusion/
Copy citation Download RIS
Tags: AI and educational policy implicationsAI and pedagogical transformationAI as meaning-mediating infrastructureAI-driven communication reform in classroomsAI-driven communication reshaping in classroomsAI’s role in academic feedback and assessmentChatGPT and educational accessibilityChatGPT and educational practicescritical review of AI’s influence on educational equitycritical review of AI’s role in educationdigital divides in AI-enabled learningeducational inequality and accesseducational inequality and inclusionGenerative AI in educationimpact of large language models on learninginclusion and exclusion in AI-supported educationlarge language models and their impactpolicy implications of AI in educationsemantic transduction in AIsemantic transduction in learningsocio-technical systems in education
