On this page

Synthesis: Videla, Ross and Aros (2026) argue that the challenge artificial intelligence poses to education is a philosophical and onto-epistemological problem rather than a technical one. AI fits contemporary schooling so readily, they contend, because it extends a cognitivist and instructional model that treats learning as information processing, teaching as transmission, and educational quality as optimization. Drawing on enactivism and situated action, the authors relocate learning in embodied participation, structural coupling, and participatory sense-making, and conclude that educational automation is an ontological illusion: the risk is not that teachers are replaced but that education is progressively redefined around whatever computational systems can represent, predict, and optimize. Against adaptive learning platforms and large language model tutors they propose AI for Embodiment, an orientation in which technology expands perception, coordination, and human agency rather than substituting for them; against automated assessment that rewards correct responses, they defend the productive friction that opaque systems erode and the automation bias that follows from it.

Conceptual position paper: the authors report no study and no data of their own. The opening teacher vignette and the two STEAM design projects are illustrative, and figures cited from other work are attributed to the studies that produced them.

Key Findings

  1. Automation continues cognitivist schooling rather than breaking from it. Adaptive tutoring, automated assessment, predictive analytics, and generative feedback extend a logic that already treated learning as information processing, teaching as transmission, and quality as prediction and control.
  2. Large language models predict tokens, not understanding. A coherent answer about a café or Newton's laws is the output of successive probabilistic selections over candidate tokens, produced without embodied participation in the situation it describes.
  3. Frictionless learning is the problem, not the promise. Systems that anticipate and minimize epistemic difficulty realize what Biesta calls learnification, displacing the ethical, normative, and relational purposes of education.
  4. Enactivism relocates learning in structural coupling. Cognition emerges from the dynamics between organism and environment, so knowledge takes shape through the coordination of gestures, bodies, expressions, and social relationships rather than through internal representation.
  5. Radical and extensive enactivism remove representation from basic cognition. Content appears at cultural and linguistic levels as normatively stabilized patterns of action, leaving no room for a computational substitute at the level of lived experience.
  6. Technology becomes an ecological participant rather than a cognitive tool. Extensive enactivism makes AI one component in networks of bodies, materials, and practices, so the design question shifts from executing tasks efficiently to expanding possibilities for perceiving and acting.
  7. The irreducible core is productive friction. Hesitation, breakdown, and material resistance become occasions for reflective reorganization, which is why education cultivates ways of inhabiting a world before it transmits information about it.

Where the automation ideal comes from

The aspiration to automate education predates contemporary AI. Videla and his colleagues trace it to mid-twentieth-century cybernetics, where learning was recast as the regulation of information flows and feedback loops, and forward through the computational theory of mind, which treated cognition as the manipulation of symbols according to formal rules. Generative AI is the latest development of that representational tradition. Its fit with schooling is therefore not a rupture but a continuity: cognition as internal computation, the body as a functional supplement, and learning as a logical sequence of input, processing, storage, and output. Within such a frame, learning becomes ontologically fragmentable and optimizable, and substituting algorithmic judgment for pedagogical judgment looks feasible and desirable.

The authors argue this mentalist ontology has been challenged since at least the 1990s. Post-cognitivist work in embodied, enactive, embedded, extended, and ecological traditions shows that purely internalist explanations cannot account for cognitive phenomena. Ryle's earlier charge of a category mistake — confusing dispositions with performances — already dismantled the assumptions that thought precedes action and that mind and body are separable. The consequence for education is direct: if learning is not the acquisition of representations, then systems built to predict the next token inherit the limits of the model rather than the capacities of the learner. Maturana's warning is quoted with approval: attributing intelligence to artificial systems confuses observed regularities between outputs and observers' expectations with embodied, historically situated cognition.

Enactivism against the representational picture

Enactivism begins from the biology of cognition: knowing emerges through situated bodily action that brings forth a meaningful world. Applied to education, this redefines learning as the dynamic reorganization of sensorimotor activity rather than the internalization of content. Motor activity, affectivity, and the temporality of experience move from the periphery to the center of the account. Radical enactivism, developed by Hutto and Myin, rejects mental representation and semantic content at the level of basic cognition; content appears later as normatively stabilized patterns sustained by shared narrative and symbolic practice. Extensive enactivism then offers an alternative to both the extended mind thesis and distributed cognition: cognition is constitutively worldly from the outset, entangled with body, action, and environment through dynamic relational couplings.

For pedagogy this shifts the unit of design. Agency is understood as a way of being implicated in a lived world, in which an organism regulates its interactions according to norms that emerge from its own organization and are shaped through social, cultural, and ethical participation. Educational agency is therefore inseparable from the intrinsic normativity of living systems and their capacity to assume responsibility for the consequences of action. Artificial intelligence may participate in educational assemblages by expanding perception, coordination, and action; the formative dimensions of agency still depend on embodied, ethical, political, and relational participation. The authors read this through Dewey's claim that education is grounded in experience, Merleau-Ponty's embodied perception, Heidegger's ready-to-hand engagement, Gibson's education of attention, and Malafouris's material engagement theory, in which brains, bodies, gestures, tools, and material culture continuously reshape one another.

From instruction to embodiment

The paper's concrete illustration is deliberately mundane. Asked for the best café in San Francisco, a model does not draw on visited cafés, tasted coffee, or compared atmospheres; it computes a probability distribution over thousands of candidate tokens and recursively selects the most likely continuation. The same architecture explains AI tutors that explain Newton's laws, solve calculus problems, generate code, or comment on essays: the responses are statistically probable token predictions, and the hesitation before manipulating a graph, the exploratory movements of debugging, and the gradual reorganization of understanding with teachers and peers remain outside the system's experiential horizon. Because generative models learn from statistical regularities in vast corpora rather than embodied participation, the apparent intelligence corresponds to regularities between outputs and observers' expectations.

That mismatch matters pedagogically as well as technically. Instructional AI aligns content ever more precisely with an estimated cognitive state and promises seamless trajectories in which difficulty is anticipated and minimized. Videla and colleagues read this as learnification and performance optimization, and they locate the corrective in design rather than in rejection of the technology. An enactive pedagogy asks how AI can expand learners' possibilities for perceiving, acting, collaborating, and making sense. Embodied design research shows conceptual understanding developing through the progressive reorganization of perceptual and sensorimotor coordinations during guided action, so the educational contribution of AI lies in shaping the ecological conditions under which learners discover concepts by acting. Two design-based STEAM projects, building fog collectors and an AI-powered waste bin, are offered as converging illustration: creativity emerged through iterative cycles of design, prototyping, failure, negotiation, and redesign, with AI functioning as one component of a broader sociomaterial ecology. What remains is a pedagogical orientation — what Abrahamson and colleagues describe as empathetic listening to student sense-making — rather than a new technical architecture.

What this means for practice

  • Instructors. Keep productive friction in the task — hesitation, debugging, handwriting, material resistance — and evaluate any AI-supported activity by which embodied action the learner still performs rather than which output the system produces.
  • Instructional designers. Specify the perception, coordination, and collaboration the learner must enact, and treat speed and seamlessness as risks to be designed against rather than defaults to optimize.
  • Educational technology developers. Build tools that expand opportunities for perception and joint action, and instrument the ways they change what learners notice rather than only the answers they return.
  • Administrators. Judge procurement and adoption by whether a system widens participation in meaningful practice or narrows education to what can be measured and predicted.
  • Researchers. Operationalize enactivist principles in diverse settings, especially under-resourced ones where automation is promoted as the answer to scarcity, and test them longitudinally and through design-based studies.

Limitations

  • The paper is primarily theoretical and interpretive; the empirical cases are illustrative rather than exhaustive evidence, and the authors say so themselves.
  • The central construct — non-automatable human experience — is argued philosophically and is not operationalized into a measure that could be tested against alternatives.
  • The two STEAM projects and the single teacher interview are thin support for claims about schooling at scale, and neither is reported with the design detail a reader would need to judge transfer.
  • The argument rests on a particular post-cognitivist reading of enactivism; rival accounts of embodied and extended cognition are acknowledged but not engaged in depth.
  • No longitudinal evidence addresses whether AI-supported, materially grounded pedagogy develops the human capacities the paper prizes over time.

Citation

Videla, R., Ross, W., & Aros, M. B. (2026). Why Education Cannot Be Automated: Enactive Pedagogy, AI, and the Crisis of Meaning. Accepted manuscript; the supplied PDF prints no venue or DOI.

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.