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 it does to schooling. Working from 4E cognition and extended views of mind (Distributed Cognition), they treat schools as hybrid cognitive ecologies where tools, norms and routines shape how Learners attend and remember. Large Language Models (LLMs) systems are therefore not an added aid but an epistemic infrastructure: unlike books or Logo they are active, persistent, generalist and autonomous, consolidating explanation, Feedback and evaluation in one interface. Two misalignments are amplified rather than created: a pedagogical gap between learning-science accounts of learning and the screen-bound logic of most educational AI, and a goal gap between measurable performance and aims such as agency, collaboration and ethical formation. Their proposal: treat GenAI as a diagnostic catalyst and redesign schooling as a multitechnological cognitive ecology, AI one tool among many.

Key Findings

1. Schools are cognitive ecologies, not neutral settings. Schooling is always multi-technological: the skills it cultivates are inseparable from the tools carrying them, and technologies normalize particular epistemic practices.
2. GenAI differs qualitatively from earlier artifacts. Books, Logo and Montessori materials are multifunctional but bounded by their materiality; Conversational AI is active, updated after deployment, generalist, autonomous, and folds Scaffolding, instruction and evaluation into one system.
3. GenAI behaves like an epistemic infrastructure. Teaching-management systems consolidate curriculum planning, instructional support, assessment and reporting, imposing one mediated environment on teachers, administrators and parents.
4. The pedagogical gap is a design problem, not an AI property. Systems reproduce a disembodied, screen-bound logic at odds with Active Learning, Situated Learning and embodied research, so students without AI Literacy rarely move past prompt tuning.
5. The goal gap concerns what education is for. Policy emphasizes collaboration, ethical formation and Critical Thinking, yet systems stay organized around efficiency and coverage, marginalizing Metacognition and Learner Agency.
6. Confabulation is both a risk and a resource. Learners struggle to tell accurate from spurious content, yet awareness of hallucination can make verification and fact evaluation integral, not optional.
7. The proposal is a multitechnological cognitive ecology. No technology should dominate: low-tech, embodied and collaborative tools combine with AI by pedagogical aim, though AI is among education's most resource-intensive tools.

Schools as cognitive ecologies

They adopt 4E cognition (mind embodied, embedded, enacted, sometimes extended) plus a fifth dimension: schooling shapes how cognition develops across developmental timescales (Learning Theories, philosophical). Technologies are therefore non-neutral: the Orbis Pictus, Logo, Flatland and Montessori materials were multifunctional by design but capped by their materiality, so learners stepped outside them for clarification or assessment. Generative AI sits on a different register, framing problems, generating explanations and evaluating outputs. It arrives as a co-agent reflecting its Stakeholders' values and, despite Bias Mitigation efforts, embedding demographic, linguistic and pedagogical biases, so practice risks drifting toward what is easiest to generate and scale.

What generative AI changes about thinking and knowing

Watch the interface: the learner queries, the system answers authoritatively, the exchange ends or iterates: instructionist pedagogy in a new medium. Variants built as cognitive partners using Socratic questioning or retrieval practice still conceive the interaction as one-to-one, leaving Collaborative Learning unaddressed. The stronger claim is about distribution: classrooms spread teaching across people and rhythms, with teachers explaining, peers challenging, materials resisting, assessments imposing delay. GenAI compresses explanation, Feedback and evaluation into one loop that smooths those frictions, shifting practice toward competencies centered on prompting, with unstudied effects on Motivation. One-to-one Intelligent Tutoring displaces that resistance, and agreeable systems habituate students away from effortful Problem Solving, against Desirable Difficulties and productive failure findings.

Two misalignments: the pedagogical gap and the goal gap

The pedagogical gap is the mismatch between 4E-informed learning and the assumptions encoded in current systems, against a wider gap between the Learning Sciences and design. That gap reflects disembodied frameworks across educational technology, and closing it needs Reflection rather than technical innovation, since method, not medium, determines efficacy. GenAI in mixed reality is structurally constrained, and tools built within teacher-centered Learning Design models limit learner choice and overlook Neurodiversity supports. The goal gap concerns aims: education is officially about collaboration, ethical formation and Critical Thinking, yet systems stay organized around efficiency and coverage under Educational AI Policy. Compressed-curriculum claims expose rather than answer whether academic achievement is schooling's primary aim, while STEM-centric AI research crowds out Creativity.

Confabulation, epistemic authority, and the multitechnological ecology

Confabulation is the third concern: incorrect content delivered fluently and confidently, and asymmetrically dangerous because learners cannot check it. They also see a resource: awareness of hallucination can make verification integral to Assessment rather than optional, and scaffold epistemic agency rather than cast the learner as a recipient. The implication is institutional: knowledge was stabilized by teachers, textbooks and curricula, so an opaque probabilistic source forces schools to rethink how Trust and accountability are distributed. The conclusion is a multitechnological cognitive ecology where no tool dominates and tools are combined by pedagogical aim, AI's energy and water costs included.

What this means for practice

• Instructors. Make verification a graded part of the task: learners cannot separate fluent confabulation from accurate content, so source-checking belongs in assessment.

• Instructors. Choose tools by pedagogical affordance and keep low-tech, embodied and collaborative options in the mix rather than defaulting to a conversational system.

• Designers. Protect the frictions that carry conceptual change, and do not let the prompt-response loop define the interaction.

• Researchers. Treat the enhancement-versus-deskilling question as open: the cited evidence is preliminary, leaving intensive unguided LLM use unresolved.

Limitations

• The paper generates no datasets, as its data availability statement says, so claims about classroom distribution and epistemic authority rest on illustrative cases.

• Its borrowed warrant is self-described as thin: pedagogically aligned AI implementations (refs. 14-16) against emerging evidence of eroded metacognitive capacity (ref. 17), leaving enhancement versus deskilling unresolved.

• The proposal is normative with nothing attached: no implementation, evaluation or measurement, its named systems will date, and the ecological cost that should bound the design is unquantified.

Citation

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