When AI Thinks for Us: Learning Science in the Age of LLMs

One of the most robust findings in the learning sciences is that what makes studying feel easy often makes it ineffective. Durable retention rests on 'desirable difficulties': self-testing rather than rereading, spacing review, generating an answer before checking it. The classic testing effect of Roediger and Karpicke shows that mere retrieval from memory, with no additional restudy, strongly improves long-term retention ^[1]. Large language models (LLMs) excel at removing exactly that effort: they produce the answer, the outline, the summary. The question is therefore not whether AI helps you produce, but whether it helps you learn.

Cognitive offloading, a double-edged sword

'Cognitive offloading' means handing a mental task to an external tool. Useful for freeing working memory, it becomes problematic when it replaces the processes that build understanding. Several 2024-2026 studies document 'metacognitive laziness': learners who rely on an LLM mostly interact with the tool, copy-paste more, and regulate their own learning less ^[2]. A measurement scale for the phenomenon has recently been proposed and validated ^[3].

The result is nuanced and often mis-cited: in the same study, a 'tutor' version of the model with pedagogical guardrails (progressive hints, refusal to give the full solution, prompts to reason) did not cause this degradation ^[4]. In other words, the effect depends less on the AI itself than on how the interaction is designed.

Designing for effort, not against it

Cognitive science offers a design grid. A learning-aligned AI tutor should favour active retrieval (ask questions before answering), meter its hints (progressive fading), space recall, and surface reasoning steps rather than the answer alone. UNESCO's guidance argues for a human-centred approach in which the tool remains subordinate to pedagogical goals ^[5].

• Retrieval before generation: have the learner attempt an answer before any help.
• Graduated hints: reveal information in steps, never the full solution up front.
• Spacing and interleaving: reschedule concepts, mix problem types.
• Externalising thinking: ask the learner to justify, rephrase, predict — not just confirm.

What AI can genuinely improve

AI is not doomed to erode learning. Well designed, it makes time-expensive principles scalable: immediate feedback, retrieval questions generated on the fly, difficulty adaptation. Stanford's 2025 AI Index highlights the rapid expansion of these uses while stressing the scarcity of large-scale, rigorous efficacy evidence ^[6]. Methodological caution remains warranted: many studies are short, on small samples, and measure immediate performance rather than retention.

Deep learning requires a desirable discomfort. The real challenge of educational AI is to preserve useful effort while removing useless effort.

Synthesis of desirable difficulties

For those who build educational systems — including European and Swiss actors attentive to privacy and data control — the stakes are twofold: instrument metacognition (make learners aware of what they know) and avoid the fluency trap, the illusion of mastery a well-turned but unearned answer creates. Technology should serve desirable difficulty, not abolish it.

Three cognitive principles worth instrumenting

Beyond retrieval, three robust principles of the learning sciences deserve to be instrumented by AI rather than short-circuited. The first is spacing: revisiting a concept at increasing intervals anchors it far better than massed cramming. A system can schedule these reviews automatically, something a lone teacher struggles to do for thirty students. The second is interleaving: mixing problem types rather than handling them in blocks, which improves transfer.

The third is the generation effect: producing an answer, even a wrong one, before receiving the correction strengthens learning more than passively reading the solution. This is exactly the opposite of what an unguarded generative assistant offers, handing over the answer up front. Designing against that reflex — making learners attempt before helping — turns the tool from a shortcut into a lever ^[4].

Then there is metacognition, the steering of one's own learning that studies flag as the first casualty of offloading ^[2]. AI can instead support it: asking the learner to estimate their confidence, spot what they did not understand, plan their revision. Making learners aware of their state of knowledge may be the most promising use — and the most faithful to a human-centred vision ^[5].

Reconciling effort and technology

The debate too often pits two caricatures against each other: AI that saves education and AI that ruins it. The more interesting reality is that the same tools can serve either, depending on their design. An assistant that hands over the answer short-circuits learning; an assistant that asks a question at the right moment strengthens it. The difference comes down to a fine understanding of cognitive mechanisms.

The promising future is one of evidence-based instructional engineering, where learning-science principles — retrieval, spacing, generation, metacognition — are deliberately built into the tools. Far from opposing effort and technology, the aim is to put technology in the service of useful effort. It is a demanding programme, but the only one that guarantees AI actually makes people learn, not just produce.