Educating minds with generative AI

Synthesis: Di Paolo, Clark and Wachter argue that framing Generative AI as a tool for efficiency and Personalized Learning misses what the technology actually does to schooling. Working from the 4E account of cognition and from extended and hybrid views of mind (Distributed Cognition), they treat schools as hybrid, multi-technological cognitive ecologies in which tools, materials, norms and routines jointly shape how Learners attend, reason and remember. On this view Large Language Models (LLMs) systems are not an added aid but an epistemic infrastructure: unlike books, globes or Logo, they are active, persistent, generalist and functionally autonomous, and they consolidate explanation, Feedback, evaluation, tutoring and even proto-emotional support inside a single interface. The authors identify two longstanding misalignments that GenAI amplifies rather than creates: a pedagogical gap between learning-science accounts of learning and the disembodied, individual, screen-bound logic baked into most educational AI, and a goal gap between measurable performance and developmental aims such as agency, collaboration and ethical formation. Their constructive proposal is to treat GenAI as a diagnostic catalyst and redesign schooling as a deliberately designed multitechnological cognitive ecology, in which AI is one tool among many, chosen for its pedagogical affordances rather than adopted by default.

Key Findings

1. Schools are cognitive ecologies, not neutral settings. Formal education has always been hybrid and multi-technological: books, blackboards, desks, bells and games are functionally integrated into cognitive activity, so the skills cultivated in schools are inseparable from the tools through which they are developed. Educational technologies therefore scaffold which epistemic practices are normalised, and their effects should be assessed at the level of how learning environments are organised at scale, how agency is distributed, and which forms of knowledge are made visible or invisible over time.
2. GenAI is qualitatively different from earlier educational artefacts. The authors insist it is not a matter of moral hierarchy between technologies: books, Comenius's Orbis Pictus, Papert's Logo and Montessori materials are multifunctional but bounded by their materiality and their author's intent. Conversational AI is active or semi-active, continuously updated after deployment, generalist, low-cost to access and increasingly autonomous, and it can subsume instruction, error correction, adaptive Scaffolding, evaluation and encouragement within one system.
3. GenAI behaves like an epistemic infrastructure. Physical and institutional infrastructures (libraries, laboratories, ritual practices) organise how knowledge is produced, circulated, evaluated and legitimised. AI-powered teaching management systems deliberately consolidate curriculum planning, instructional support, learner assessment and progress reporting into one algorithmically mediated environment, imposing a mandatory participation structure on teachers, administrators, secretaries, principals and parents.
4. The pedagogical gap is a design problem, not a property of AI. Current systems reproduce a logic that is disembodied, abstract, individual and linguistic, at odds with research treating learning as Active Learning, Situated Learning, embodied and socially mediated. The gap shows up in interface design: the prompt-response loop trains prompt tuning, and empirical evidence cited by the authors indicates that students without clear guidance or sufficient AI Literacy rarely move beyond it, producing interactions of limited educational value and increased reliance on system outputs.
5. The goal gap concerns what education is for. Educational theory and policy emphasise collaborative capacities, ethical formation and Critical Thinking, yet systems remain organised around efficiency, standardisation and content coverage; technologies adapted to those logics reproduce and amplify them. Recent initiatives claiming to compress traditional curricula into a few hours a day through extensive AI use are read as exposing, rather than answering, the deeper question of whether the primary aim of schooling should be academic achievement alone.
6. Confabulation is both a risk and a pedagogical resource. Learners who by definition lack background knowledge struggle to distinguish accurate from spurious content delivered with equal authority, and sustained reliance may normalise passive consumption over critical scrutiny. But awareness of hallucination can make verification, source-checking and collaborative fact evaluation integral rather than optional, and the authors draw three implications: embed verification across subjects and modalities, redesign systems to scaffold epistemic agency rather than position the learner as a solitary recipient, and confront the fact that schools can no longer assume institutional authority over what is true and worth knowing.
7. The evidence the paper leans on is explicitly preliminary. It cites studies suggesting that pedagogically aligned AI implementations can support learning (refs. 14-16) alongside emerging evidence that intensive, unguided LLM use may erode the metacognitive capacities and epistemic ownership that underpin effective learning (ref. 17). It frames the enhancement-versus-deskilling debate as unresolved, and states plainly in its data availability section that no datasets were generated or analysed.
8. Friction and dissent are treated as learning infrastructure. The authors argue that one-to-one Intelligent Tutoring can displace productive sources of resistance: a peer's alternative or even incorrect answer that catalyses conceptual clarification, the missing note or broken pencil that introduces material resistance, the scheduled assessment that imposes delay and revision. Highly agreeable systems risk replacing dialogic friction with continuous affirmation and preference alignment, while the learning sciences identify effort, struggle and uncertainty resolution as key drivers of conceptual change.
9. The proposal is a multitechnological cognitive ecology. No single technology should dominate: low-tech, embodied and collaborative tools are combined with digital and AI resources according to pedagogical aims rather than market logic. The framing carries an eco-sustainable dimension, since AI systems are among the most resource-intensive technologies ever deployed in education, and implies a principle of deliberate use: combine tools by affordance, share resources across learners and generations, and treat AI as one component among many rather than a default infrastructure.

Schools as cognitive ecologies

The theoretical frame is explicit and philosophical rather than psychological. The authors adopt the 4E account of cognition, on which mind is embodied, embedded, enacted and in some cases extended, and they argue that education is not merely an application of that account but its extension: a fifth dimension, in which cognition is also educated. Because schooling organises environments, practices and norms over developmental timescales, it does not simply support cognition but structures how cognition develops across individuals and generations, which the authors place within Learning Theories research.

This makes educational technologies non-neutral. They foreground some forms of engagement and normalise some epistemic practices, and so function as scaffolds that shape how future cohorts learn. The paper runs a historical comparison to make the point concrete. The Orbis Pictus taught vocabulary, language and world knowledge at once, Logo merged programming with epistemic reflection around debugging, Abbott's Flatland could work as a lesson in geometry and social critique, and Montessori materials such as the Stamp Game and Puzzle Maps were deliberately self-correcting and multisensory. Each was multifunctional by design, but each carried a ceiling set by its materiality, and the learner had to step outside the object for elaboration, clarification or assessment.

Generative AI, by contrast, is placed on a different register because it participates in epistemic processes rather than merely supporting them. It can plan and execute sequences of actions with limited supervision, it is reconfigured after deployment, and it performs functions until recently reserved for human cognition or constant oversight: problem framing, explanation generation, evaluation, information selection, and forms of mental and emotional support. The authors also stress that these systems arrive as co-agents reflecting the values and constraints of the Stakeholders who develop and optimise them: despite multiple efforts at Bias Mitigation, they still embed demographic, gender, religious, linguistic and pedagogical biases, and remain open to post-deployment modification that alters what is generated, emphasised or suppressed. The risk they name is epistemic substitution: educational practice drifting toward what is easiest to generate, evaluate and scale, shifting the sources of epistemic authority that shape public knowledge.

What generative AI changes about thinking and knowing

The interface is where the argument becomes concrete. Most GenAI tools rely on a narrow interaction paradigm in which the learner formulates a query, the system answers, and the exchange ends or iterates along the same line; the system is cast as the source of a coherent and authoritative answer. The authors read this as instructionist pedagogy in a new medium, decomposing knowledge into discrete units and evaluating progress by correctness and fluency, and they note that promising variants designed as cognitive partners using Socratic questioning and retrieval practice still typically conceive of the interaction as one-to-one, leaving the social and collaborative side of Collaborative Learning unaddressed.

The stronger claim is about distribution. In a conventional classroom, teaching and learning are spread across people, artefacts and temporal rhythms: teachers question and explain, peers challenge and negotiate, materials resist through incompleteness, assessments impose delay, and the bell structures shared attention. GenAI compresses explanation, feedback, evaluation and content generation into a single interactional loop that smooths the frictions and ambiguities the authors treat as pivotal to learning. Immediate feedback may help, but displacing peers, disagreement and collective sense-making shifts practice toward individualised competencies centred on prompting, filtering and evaluating outputs, with acknowledged implications for Motivation and belonging that remain largely unstudied.

Agreeability compounds this. The authors worry that fluent, accommodating systems habituate students to avoid effortful engagement and sustained Problem Solving, reducing encounters with resistant others whose perspectives cannot be personalised away, and they connect this to learning-science findings on Desirable Difficulties, uncertainty resolution and productive failure. Their counterfactual is a system that reintroduces desirable difficulty, but they concede that even Socratic or retrieval-based designs keep the learner as a solitary interlocutor.

Two misalignments: the pedagogical gap and the goal gap

The pedagogical gap is defined as the misalignment between 4E-informed understandings of learning and the pedagogical assumptions encoded in current AI systems, nested in a wider disconnection between the learning sciences and technology design. The authors are careful not to indict AI as such: the gap reflects the dominance of disembodied frameworks in educational technology more generally, and closing it requires pedagogical, institutional and cognitive reflection rather than technical innovation. Their summary of the point is Weidlich and colleagues' reminder that the efficacy of learning is determined by the instructional method, not the medium through which it is delivered. They note exceptions (GenAI embedded in mixed reality, or in embodied collaborative frameworks for learners with special educational needs) but regard them as structurally constrained, and they argue that lesson-planning and Learning Design tools developed within directive, teacher-centred models reproduce highly structured formats that limit learner choice and collaboration, and can overlook diverse cognitive profiles and the environmental supports Neurodiversity research highlights.

The goal gap concerns aims rather than mechanisms. Education is officially about collaborative capacity, ethical formation and critical thinking, yet systems stay organised around inherited logics of efficiency, standardisation and coverage, the logics embedded in Educational AI Policy, so technologies introduced without interrogating aims get adapted to that misalignment and amplify it. The example the authors use is the claim that AI can compress traditional curricula into a few hours per day: if that is possible, they say, it exposes unresolved questions about whether achievement should be schooling's primary aim, whether the time conventionally spent has been worthwhile, and whether metrics have obscured what education is for. Once capacities are reduced to what can be articulated, assessed and scaled, socialisation and individual agency risk being marginalised, alongside Metacognition, epistemic responsibility and tolerance of uncertainty. They also criticise the STEM-centric orientation of AI research in education for crowding out interpretive understanding, ethical reasoning, social participation and Creativity that STEM knowledge itself depends on when applied reflexively.

Confabulation, epistemic authority, and the multitechnological ecology

The third concern is hallucination or confabulation: incorrect or fabricated content presented fluently and with unwarranted confidence, persisting despite mitigation and requiring external verification. In educational settings this is asymmetric, because learners lack the background knowledge for independent verification and cannot easily separate accurate from spurious content delivered in the same authoritative register. The authors see a genuine advantage too: awareness of confabulation can make verification, source-checking and collaborative fact evaluation integral to Assessment rather than optional.

The deeper implication is institutional. Knowledge in formal education has been stabilised through teachers, textbooks and curricula that defined what counted as relevant and reliable, and GenAI introduces an opaque, probabilistic source whose outputs do not map onto established standards of authority. Schools must therefore reconfigure how Trust, responsibility and accountability are distributed across teachers, learners and tools, and independent verification should become a core component of learning rather than an error-mitigation add-on.

The constructive conclusion is a multitechnological cognitive ecology: a deliberately designed configuration of material, social and digital resources in which no single technology dominates and tools are combined according to pedagogical aims. The authors pair this with a principle of deliberate use, including sharing resources across learners and generations and accounting for the energy, water and infrastructural costs of AI, which they describe as among the most resource-intensive technologies deployed in education. Their closing framing of the technology is deflationary: AI is neither saviour nor threat, but a stress test for education that exposes the limits of existing models, and one tool among many.

A conceptual argument rather than an empirical study

This is a Perspective, and its evidential basis is correspondingly thin by design. The authors state that no datasets were generated or analysed, so every empirical claim in the paper is a citation to other work rather than a finding of their own, and they describe the enhancement-versus-deskilling literature as preliminary and the debate as unresolved. Claims about what GenAI does to classroom distribution, social friction or epistemic authority are argued conceptually, with illustrative cases such as AI-managed teaching platforms, lesson-plan generators and compressed-curriculum schools, not demonstrated empirically.

The proposal itself is normative and design-oriented. The multitechnological cognitive ecology and the closing of the two gaps are offered as a reframing plus a direction for research and practice, with no implementation, evaluation or measurement attached; the paper's own summary states that the analysis hinges on treating GenAI as a diagnostic opportunity rather than on new evidence. The specific products and systems named will date as capabilities and deployment policies change, though the authors argue the underlying principles on which they criticise those systems will not. The ecological costs of AI are named as a constraint on any ecological design, but not quantified here.

Connected Concepts

• Distributed Cognition — the extended and hybrid accounts of cognition on which the whole argument rests
• Embodied Learning — the embodied, situated learning the authors say current educational AI undercuts
• Cognitive Offloading — the delegation and erosion risk from unguided, intensive LLM use
• Critical Thinking — the capacity the authors say unguided use may erode and the goal gap marginalises
• Learning Theories — learning-science findings on effort, struggle and conceptual change, set against AI design
• Scaffolding — what educational technologies do to learning, and what AI now does more autonomously
• Generative AI — the technology recast here as epistemic infrastructure rather than a helpful tool
• Large Language Models (LLMs) — large language models as the active, persistent, generalist systems at issue
• Hallucination Risk — confabulation as both an educational risk and a verification pedagogy
• AI Sycophancy — agreeable systems replacing dialogic friction with continuous affirmation
• Desirable Difficulties — effort and productive struggle as the conditions AI interfaces tend to smooth away
• AI Literacy — the guidance whose absence pushes students into prompt tuning

Connected Articles

• Artificial intelligence, cognitive offloading and implications for education — Cognitive offloading and its implications for education, a key source for the erosion concern
• Meta-Cognitive Insights into Cognitive Offloading: Mechanisms, Interventions, and Educational Implications — Review evidence on offloading and metacognition that bears on the enhancement-versus-deskilling debate
• Towards a philosophy of ensemble cognition: Reconceptualising agency and mind in AI-mediated educational environments — Philosophical work on cognition and AI in education in the same conceptual register
• "If You Can't Dance Your Program, You Can't Write It": Challenges and Implications for AI in Education — Embodied AI in education, the alternative pathway the paper calls for
• Thinking Through AI: Advancing Cognitive and Collaborative Research for AI in Education — Thinking through AI: distributed cognition as an educational frame
• Generativism: Toward a Learning Theory for the Age of Generative Artificial Intelligence — A learning-theory attempt to account for generative AI in instruction
• Generative AI (GenAI) as a mindtool that supports generative learning (GL) — Generative AI framed as a mind tool rather than a delivery mechanism
• The critical-thinking paradox in generative AI-integrated learning: distinguishing efficiency from cognitive depth — a differentiated framework and testable propositions — The paradox of GenAI supporting and undercutting critical thinking
• Rewriting the Curriculum: A Systematic Review of Generative AI-Driven Pedagogical Change and Emerging Systems of Learning in Higher Education — Curriculum-level rethinking of pedagogy under generative AI

Citation

Di Paolo, L. D., Clark, A., & Wachter, T. (2026). Educating minds with generative AI. Communications Psychology, 4, 128.