A teacher asks ChatGPT to generate a lesson plan on the causes of World War I. The plan is competent, well-structured, and ready in seconds. The teacher reviews it, makes a minor adjustment, and moves on. Something important has been lost in this transaction — not the lesson plan, which is adequate, but the cognitive work of creating it: the weighing of student needs, the sequencing of ideas, the integration of content knowledge with pedagogical judgment. The AI performed the epistemic labor, and the teacher consumed its output. Across education, this pattern is repeating at scale, raising a question that goes beyond efficiency: what happens to human knowledge-making capacity when AI systems perform the cognitive work that builds it?
Beyond Tools: AI as Epistemic Infrastructure
Chen (2025) argues that this question cannot be answered by treating AI as a tool. Tools augment human capabilities while leaving the user's cognitive processes intact — a calculator enhances arithmetic performance without changing how the user thinks about numbers. AI systems in education operate differently. They do not merely assist with epistemic tasks; they restructure the conditions under which epistemic agency can be exercised.
Drawing on theories of technological mediation and distributed cognition, Chen proposes analyzing AI systems as epistemic infrastructure — systems that shape what questions appear reasonable, what forms of engagement seem possible, and what kinds of knowledge practices are sustained or abandoned. His framework identifies three conditions that epistemic infrastructure must satisfy to preserve human agency: it must afford skilled epistemic actions (not just produce outputs), support epistemic sensitivity (awareness of what one knows and does not know), and foster virtuous habit formation (practices that build expertise over time).
Applying this framework to AI lesson planning and feedback tools reveals a pattern he calls epistemic substitution: the AI efficiently handles teaching tasks while performing the cognitive operations — evaluation, synthesis, judgment — that would otherwise develop the teacher's professional expertise. The AI is useful for the immediate task but corrosive to the long-term development of the skilled judgment that makes teaching more than content delivery.
The Empirical Evidence for Cognitive Displacement
Yan, Pammer-Schindler, and Mills (2025), editing a special section of the British Journal of Educational Technology that has attracted 25 citations, assemble emerging empirical evidence on GenAI's cognitive and metacognitive effects on learners. Their synthesis moves the discussion beyond speculation into data.
The evidence confirms what Chen's theoretical framework predicts. GenAI delivers measurable efficiency gains — students complete tasks faster, produce more polished outputs, and report less frustration. But these gains mask a displacement effect. When AI handles the effortful cognitive processes that drive learning — formulating questions, evaluating evidence, constructing arguments, monitoring comprehension — learners engage in less of the deep processing that builds durable understanding. The efficiency gain comes at the cost of the struggle that produces learning.
The metacognitive dimension is particularly concerning. Effective learning requires accurate self-monitoring: knowing what you understand, recognizing when you are confused, and adjusting strategy accordingly. When AI provides fluent, confident answers, learners' ability to detect their own misunderstandings deteriorates. The AI's competence creates an illusion of the learner's competence — a form of metacognitive miscalibration that is difficult to detect precisely because the learner does not know what they do not know.
Designing for Self-Regulation
Xu, and colleagues (2025), also in the British Journal of Educational Technology, investigate practical strategies for enhancing self-regulated learning in GenAI environments. Their work addresses the question that follows from the diagnosis: if AI erodes metacognition, can educational design restore it?
Their approach involves structuring AI interactions to preserve the cognitive work that drives learning. Rather than providing complete answers, AI systems can be designed to scaffold the learning process — asking probing questions, prompting reflection before delivering information, requiring learners to formulate their own approach before seeing the AI's suggestion. The goal is to insert metacognitive checkpoints into the human-AI interaction: moments where the learner must actively monitor their understanding before proceeding.
The underlying principle is that self-regulated learning requires encountering difficulty in a structured way. Effortless AI assistance eliminates the productive struggle that builds metacognitive capacity. Structured AI assistance preserves the struggle while providing support at precisely the points where learners would otherwise disengage. The design challenge is calibrating this balance — enough difficulty to drive learning, enough support to prevent frustration.
The Epistemic Agency Stakes
The implications extend beyond pedagogy. If a generation of students learns to think by consuming AI-generated outputs rather than by constructing knowledge themselves, the long-term effects on epistemic agency — the capacity to independently form, evaluate, and revise beliefs — could be substantial. Education is not just about acquiring information; it is about developing the cognitive machinery to produce, evaluate, and apply knowledge independently. When AI performs this machinery's function, the machinery may not develop.
This does not mean AI should be excluded from education. The evidence suggests that AI collaboration improves performance on creative and analytical tasks. The question is not whether to use AI but how to design its integration so that it develops rather than displaces the metacognitive capacities it requires. The teacher who uses AI to generate three competing lesson plan options and then evaluates them against student assessment data is developing professional judgment. The teacher who accepts the AI's recommendation without evaluation is not. The difference lies in the design of the interaction, not the presence of the technology.