On this page

Critiques of AI in education — the body of scholarship that questions not whether AI in education works but whether it should, and what it is quietly turning education into. These are mostly conceptual, normative and political arguments rather than empirical ones, drawn from philosophy of education, critical theory, science and technology studies, and disability and postcolonial scholarship. Their recurring targets are the cognitivist model of learning that AI extends rather than displaces, the reframing of teaching as optimization, the assumption that adoption is inevitable, and the concentration of educational decisions in a small number of private providers. The critiques do not share one alternative. What unites them is a refusal to treat the question as settled, and a demand that education set its own agenda rather than adapt to a technology built elsewhere.

Questions to Consider

  • A critic argues that AI fits schooling too well, because it extends an existing model of learning as information delivery. Where in your own institution does the technology change the task, and where does it only make the existing task faster?
  • "Human-centered AI" is the near-universal policy baseline. What does that phrase commit its users to, and what does it leave unexamined?
  • If a system reliably produces fluent, plausible work, what is lost when a learner no longer has to struggle through producing it? Is the loss a learning loss or a workload loss, and who bears each?
  • Who profits from AI in education, and who decides? What would it mean for a school or university to set the agenda rather than adapt to one?
  • Several of these critiques are philosophical and do not report data. What would count as evidence that a critique is right, and what would count as evidence that it is wrong?
  • Is "resisting AI" a position teachers and students can hold without being called obstructive? What would a defensible middle position look like?

Introduction

Most research on AI in education asks how well it works. A smaller, older and less easily counted body of work asks a different question: what is it doing to education while we measure it. Taken together they form a distinct body of scholarship, one that argues from premises rather than from results. It is not a single school of thought. It runs from analytic philosophy of education through critical pedagogy, post-humanist theory, science and technology studies, and scholarship on disability, language and the global South. Some of it is written by people who build AI systems; some by people who refuse to.

The critiques are harder to summarize than findings because they argue about premises rather than results. A study can be replicated; a claim that education is being redefined as the production of measurable outputs is not. The critiques are best arranged by their targets, and they disagree with each other as often as they agree.

Automation as continuity, not rupture

The most direct critique is that AI in education does not break with the past but perfects it. Videla, Ross and Aros argue that the appeal of AI as an educational agent comes less from the technology's capabilities than from its compatibility with cognitivist and instructional models that already treat learning as information processing, teaching as transmission and educational quality as optimization (Videla, Ross and Aros, 2026). On that reading, automation is an ontological illusion: it abstracts learning from the embodied, relational and lived conditions in which meaning, agency and responsibility appear. The risk is not that teachers get replaced but that the definition of education narrows to what a machine can administer.

This is a stronger claim than "teachers are being displaced". It says the technology arrives already fitted to a particular idea of schooling, so adoption is not a neutral choice among tools but a vote for that idea. A companion argument locates the problem in the design objective itself: general-purpose models are tuned for task completion, so the better they get at helping the more reliably they absorb the cognitive work through which capability develops (CAEAI, 2026).

Learnification and the human-centered consensus

A second cluster attacks the vocabulary rather than the machinery. Wang and Jin read policy documents from UNESCO, the OECD and the Montréal Declaration alongside Gert Biesta's account of "learnification" and the post-humanist critique of anthropocentrism, and conclude that the insistence AI in education be human-centered is a discourse with costs rather than a neutral ethical baseline (Wang and Jin, 2026). Once education is defined as learning, and learning as the acquisition of pre-specified outcomes, the interesting questions — what education is for, who decides, what counts as a good life — fall out of view. Similar arguments appear in work on critical pedagogy, which treats AI literacy as an exercise in power-knowledge rather than a set of competencies (AI Literacy: An Exercise in Power-Knowledge), and in calls to add "resisting AI" as a legitimate stance within critical AI literacy (Li and Mroziak, 2026).

Infrastructure, transduction and misrecognition

A third cluster shifts attention from the tool to the infrastructure around it. Watson, Morgner and Brezovec treat generative AI as meaning-mediating infrastructure and name its mechanism semantic transduction: the reformatting of an educational demand — a prompt, brief, rubric, exam question, genre or policy — into a communicatively plausible form (Watson, Morgner and Brezovec, 2026). Because transduction is selective, it can include or exclude. Three sites recur: access and capability divides, misrecognition through templates of what a good answer or a good student sounds like, and normative drift in what pedagogy is taken to mean. The exclusion is often a pincer rather than a ban — institutions raise expectations about pace, polish and idiomatic fluency while tightening authenticity regimes at the same time.

Agenda-setting, simulated innovation and experimental futuring

Other critics look at who sets the terms. Fajardo-Ramos, Chiappe and Martínez-Pérez argue that educational technology repeatedly produces "simulated innovation" — symbolic modernization without pedagogical direction — and ask whether education can set the agenda for AI rather than keep adapting to it (Fajardo-Ramos, Chiappe and Martínez-Pérez, 2026). Williamson and colleagues study the policy machinery rather than the classroom, naming "experimental futuring" the mode of anticipatory governance in which future-oriented talk about AI is turned into pilots, prototypes and procurement (Williamson et al., 2026). Tlili and colleagues set a techno-optimist and a critical-realist reading of open educational resources side by side, and decline to choose between them (Tlili et al., 2026).

Others press the civic case: that the dominant pattern of use trades away the practice democratic citizenship needs (Muniz Samayoa, 2026), and that the evidence base for agentic AI in higher education is not yet mature enough to support the governance claims made for it (Dey, 2026).

What the critiques share, and where they diverge

The shared move is to treat adoption as a political and philosophical decision rather than a technical one. The shared demand is that education set its own agenda. Where the critiques diverge is on what should replace the present arrangement: enactivist accounts want embodied participation restored, critical-pedagogy accounts want power made visible, infrastructure accounts want selection made contestable, and sustainability accounts want durable learning value placed ahead of efficiency (Komodromos, 2026).

These arguments also sit in tension with the field's own self-critique. The methodological problems documented on Limitations in AIEd Research — inflated meta-analytic effect sizes, construct incoherence, the speed at which findings age — are about the quality of the evidence. These arguments are about the purposes the evidence serves. A rigorous study of the wrong question does not answer them.

What the critiques do not settle

  • They are not a unified alternative. No shared program follows from them, and several prescriptions contradict each other.
  • They are mostly non-empirical. A critique can be cogent and still be wrong about how teachers and students actually behave.
  • They risk a symmetrical error. Treating all AI use as false consciousness is as crude as treating all adoption as progress.
  • They are written largely in the global North. The infrastructure and transduction critique is the partial exception, and its evidence base is uneven.
  • They do not tell practitioners what to do on Monday. Their value is in the questions they make askable, not in a checklist.

Common Misconceptions

  • "Critics of AI in education are against technology." Most of the work summarized here is not anti-technology. It is against a particular settlement in which the technology's design goals are adopted as education's goals.
  • "These are just opinion pieces." Several are peer-reviewed analyses of policy texts, citation records or infrastructures. Their claims are contestable, not merely expressive.
  • "A critique and a finding cannot both be right." A tool can raise measured learning and still narrow what education is for. The two claims operate at different levels.
  • "The critique is settled by better AI." The arguments about learnification, agenda-setting and provider dependence are not addressed by a more capable model.

Connected Concepts

Connected Articles

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.