Research Article
Generative AI as meaning-mediating infrastructure in education: a critical integrative review of inclusion, exclusion, and semantic transduction
Synthesis: Watson, Morgner and Brezovec (2026) ask how generative AI reshapes educational inequality and conclude that these systems matter less through isolated outputs than as meaning-mediating infrastructure. Through what they call semantic transduction, large language models reformat prompts, rubrics, Feedback, and disciplinary genres into communicatively plausible forms, changing who can be recognized as a legitimate participant. Treating inclusion and exclusion as selective communicative processes rather than opposites or properties of individual ability, they trace three recurring sites: access and capability divides, misrecognition through templates of good performance, and normative drift in pedagogy, authorship, and assessment. Consequences fall unevenly on disabled learners, neurodivergent students, multilingual writers, and groups already marked by the digital divide. Individual AI literacy is necessary but insufficient; the authors propose an infrastructural governance agenda spanning policy transparency, procurement, participation, redress, and audit.
Critical integrative review: the article reports no dataset of its own, estimates no effect sizes, and quantifies no prevalence. Every figure and finding described below is attributed to the sources it synthesizes.
Key Findings
- Reframe generative AI as infrastructure, not a tool. What matters is not any single answer but repeated interactional patterns normalized in classrooms, homes, and assessment cultures, which alter what counts as adequate preparation, legitimate assistance, credible authorship, and fair evaluation.
- Semantic transduction names the mechanism. AI reformats an educational demand — a prompt, brief, rubric, exam question, genre, or policy — into a communicatively plausible form, which can make a contribution easier to process or harder, never simply neutral.
- Three sites recur across the literature. Access and capability divides; misrecognition through templates of what a good answer, essay, or student sounds like; and normative drift in pedagogy, authorship, and assessment as platforms become defaults.
- Individual AI literacy is necessary but insufficient. Verification, transparency, and contestation depend on conditions set by platform design, procurement, policy, and accountability regimes, so responsibility must be distributed.
- Misrecognition can reproduce dominant norms. If models draw disproportionately on Anglo-European academic styles, culturally specific framings survive only after translation into mainstream forms, and detectors documented as biased against non-native English writing turn the same polish into suspicion.
- The most exclusionary pattern is a pincer, not a ban. Institutions implicitly raise expectations about pace, polish, and idiomatic fluency while simultaneously tightening authenticity regimes, so attempts to level the field can produce leveling-down effects.
- The goal is repairable partial inclusion. Selection is unavoidable in education; the task is to keep selections observable, contestable, and revisable before they harden into unanswerable infrastructure.
From access to address
The review begins by refusing the two dominant narratives: AI as a neutral access tool that levels the playing field, and AI as a simple amplifier of disadvantage. Both compress inequality into outcomes measured by scores and progression, or into access measured by devices, connectivity, and platform availability. Digital inequality research has long shown these accounts are incomplete, because inequality is also shaped by differences in use, skill, capability, institutional mediation, and the social organization of support. Watson and colleagues add a communicative emphasis: educational inequality is produced through the everyday processes by which learners are addressed, categorized, and recognized. Communication does not describe a learner who is already present; it creates the address through which a person can become relevant as a learner, author, candidate, or case.
This reframing draws on Luhmannian systems theory and its distinction between inclusion and exclusion, which replaces the older language of integration. A learner becomes included when educational communication can address them as someone whose contributions, questions, performances, or needs matter; a learner is excluded, or only partially included, when the communicative forms of the institution offer no such address. A deaf learner is not excluded by deafness alone, the authors note, but by settings that presuppose hearing as the normal route to participation. Exclusion therefore includes being formally enrolled yet routinely misrecognized, being addressed only through deficit categories, and being unable to contest the categories through which one is processed. From this angle, introducing large language models does not resolve the inclusion/exclusion paradox; it reorganizes it.
Semantic transduction: how demands get reformatted
The paper's conceptual contribution is the term semantic transduction, defined as the socio-technical process by which generative AI reformats an educational demand into a communicatively plausible form. The output may be a plan, summary, model paragraph, redrafted explanation, simulated response, policy interpretation, or set of alternative formulations. Unlike ordinary translation, transduction changes the learner's relation to the demand so that the resulting communication becomes more or less recognizable within a system's expectations; unlike scaffolding, it is not normatively oriented to development and can support learning, bypass it, standardize expression, or misrecognize learners depending on how it is embedded.
A multilingual student pasting an assignment brief into a model illustrates both directions. The model can supply a possible structure, key terms, paragraph moves, and alternative framings, lowering the cost of participation by making expectations legible. The same process can steer the student toward mainstream argumentative templates, standardized academic tone, and forms of evidence the model supports more easily than the student's own repertoire. Meaning, on this account, is not a substance stored in data; it is a medium of selection. Generative AI does not transport meaning intact from one mind to another but participates in meaning-making by producing selections that are plausible within a situation — which is why communicative plausibility and epistemic validity come apart, and why verification practices decide whether plausible output becomes reusable knowledge.
Three sites where inclusion and exclusion are reorganized
Access and capability divides remain foundational. Access covers devices and connectivity but also institutional permission, paid subscriptions, language resources, time, safe spaces for experimentation, and trust that AI use will not be treated automatically as misconduct. Capability covers prompting, iterating, evaluating outputs, and integrating suggestions into situated tasks — capacities that reward iterative refinement and metacognitive control and that are unevenly distributed. What matters is not only who can use AI but who can be recognized as legitimately using it; learners under tighter surveillance regimes may be excluded through suspicion and sanction even when their purposes match those of peers.
Misrecognition arises when models supply ready-made templates of good performance. Recognition in schooling is closely tied to language and genre, so AI-generated templates can function as a new form of cultural capital, re-entrenching dominant norms by making them easier to reproduce and harder to contest. In decolonizing and social-justice contexts, inclusion can become conditional on assimilation to recognizable templates. Parallel risks appear in disability and neurodiversity contexts, where AI can support alternative expression and self-advocacy while also stabilizing narrow norms of what an articulate or organized learner sounds like. Normative drift completes the picture: feedback tone, task design, criteria of originality, expectations of pace, and assumptions about polish shift toward what AI makes easy to produce and evaluate, often without any explicit decision to exclude.
Governance as infrastructure
The practical implication is that AI literacy cannot carry the burden alone. Watson and colleagues propose four layers of distributed responsibility. Policy and transparency require legible published rules that distinguish formative support, accessibility support, drafting, translation, feedback, source use, and final assessment evidence instead of treating all AI use as one category. Design and procurement require institutions to treat these systems as pedagogical and governance infrastructure, weighing accessibility, language coverage, data practices, feedback tone, audit logs, bias testing, privacy, age-appropriateness, and teacher oversight. Participation and redress require channels through which students, teachers, and families can report harms and contest classifications, particularly around detection and integrity, where signals should not be treated as conclusive evidence. Audit and iteration require ongoing evaluation of how expectations about polish, pace, originality, legitimate help, and verification burdens shift, and whether policies differentially affect multilingual, disabled, first-generation, caring, and low-connectivity learners.
Underpinning this is an autopoietic ecological orientation: AI-mediated education is a coupled ecology of classroom routines, household capacities, platform interfaces, assessment programs, policy rules, and vendor infrastructures. Inclusion and exclusion are rarely produced by one actor, technology, or rule. 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 before exclusions harden.
What this means for practice
- Instructors. Treat genre as an object of learning rather than something to outsource: use AI to make expectations legible, then have students compare model formulations with their own reasoning, and configure it as a partner for thinking aloud — testing alternative solution paths and rehearsing competing interpretations — not as a supplier of polished final prose.
- Instructional designers. Build scaffolded pathways that keep the learner performing the cognitive work, and specify process evidence (drafts, revisions, attribution statements) alongside any product.
- Educational technology developers. Treat accessibility, language coverage, bias testing, auditability, and feedback tone as design requirements, and document the communicative styles a system recognizes best.
- Administrators. Publish program-level guidance that separates permitted formative, accessibility, translation, drafting, and feedback uses from final assessment evidence; require human review with appeal routes rather than detection-only decisions, and monitor whether expectations about pace, polish, and originality rise faster than the support available to meet them.
- Researchers. Test semantic transduction across subjects, languages, assessment forms, and disability contexts, and audit whether detection and feedback systems misrecognize particular linguistic, cultural, or neurodivergent styles.
Limitations
- As a critical integrative review it does not estimate effect sizes, quantify prevalence, or rank empirical findings with systematic-review procedures; its contribution is conceptual and diagnostic.
- The synthesis proceeded abductively from purposive theoretical sampling, keyword searching, and citation chaining rather than a protocol with reported screening counts, so coverage of a fast-moving field is selective.
- Searching ran through Google Scholar and selected publisher databases plus policy repositories, and inclusion favored English-language sources and work available after late 2022.
- Adjacent literatures such as human-computer interaction, automated writing evaluation, learning analytics, and disability studies were included selectively and are not reviewed exhaustively.
- The framework is offered as a vocabulary for future empirical, comparative, and design-oriented work rather than as a validated model, and the review itself supplies no new data.
Citation
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, 41, 10159–10173.