How can AI enable real learning?

In a randomised study of about 1,000 high school maths students in Türkiye, students using a standard ChatGPT scored 48% higher than classmates without it while they had it. When access was taken away, they scored 17% lower than students who never had it. A 2025 paper in Nature Reviews Psychology makes the underlying point: performance gains are not the same as learning.

Most AI tools were built for work, not to optimise learning. At work, the goal is to finish the task with the least effort. But in learning, that effort is the point: it is what builds the capability. The task gets finished, but the understanding never develops. Researchers call this “metacognitive laziness”: the learner stops planning, monitoring and self-evaluating, because the AI always has an answer.

Learning scientists Robert and Elizabeth Bjork call these “desirable difficulties”: things like recalling an answer from memory, spacing your practice out, or trying a problem before you see the solution. They make learning feel harder now, but they make it last. When AI removes that effort, it can quietly remove the learning the effort produced. A 2026 review of 67 studies names the same mechanism, epistemic friction: without it, “AI-generated fluency can bypass the reflective struggle central to deep learning.”

If the struggle is the learning, the first thing it depends on is the person doing it. Writing in The Atlantic, David Brooks argues that what will set people apart in an AI-saturated world is not how smart they are but their relationship to mental effort. His frame is need for cognition, the trait named and measured by John Cacioppo and colleagues, who reviewed more than 100 studies of it: at one pole, people who read dense books and play hard games for pleasure; at the other, the cognitive miser, who avoids effortful thought where possible. It correlates with intelligence, Brooks notes, but is not the same as it.

The evidence he marshals is largely about what happens when effort is offloaded. An MIT Media Lab team led by Nataliya Kosmyna measured brain connectivity dropping by up to 55% during ChatGPT use; Michael Gerlich (SBS Swiss Business School) found a significant negative correlation between frequent AI use and critical thinking; a Carnegie Mellon team led by Grace Liu found that after about ten minutes of AI-assisted problem solving, people who then lost the tool did worse than those who never had it. A study of endoscopists found precancerous-lesion detection fell from 28.4% to 22.4% once AI was withdrawn. From this he sketches three responses: the Productive Passenger, who lets AI think and stops noticing; the Reluctant Optimizer, who means to resist and gets pulled in; and the Mental Marathoner, who keeps the hard parts hard. His worry is a cognitive polarization between the two ends. His practical rule: ask AI for thinkers, not thinking, and treat it as a brilliant librarian, not an oracle.

Volition is not the only person-side variable. Stephen Fitzpatrick, a history teacher of thirty years, argues that the skill gating all of this is reading: AI output is fluent, confident and endless, and reading it with the skepticism it demands takes the vocabulary and background knowledge that only years of reading build. “The bottleneck to effective student AI use is mostly a problem of reading”, not writing, which gets the attention because writing is what we grade. Deep research makes the point sharper: finding sources is no longer the hard part; reading the report, and its sources, carefully is.

Marie Dollé, writing in French about the summer’s “brainmaxxing” apps, takes the same point one level deeper. Against the comforting idea that machines can hold the knowledge while we keep the judgment, she argues that judgment cannot form in a vacuum: “on ne problématise pas sans repères”, you cannot frame a problem without reference points already in your head. And she names a limit no better model fixes: what matters is often only knowable later, so no tool can pre-sort “the essential” for us. Consulting a fact on demand is not the same as having it in mind.

Daniel Susskind, an economist who has spent fifteen years studying AI and work, gives the same claim a mechanism. AI systems are what Geoffrey Hinton calls “idiot savants”: impressive on hard problems, wrong on easy ones, and nothing in the output tells you which one you are reading. So the basics are not a safe harbour, they are an instrument. Use AI critically rather than blindly, he argues, “keeping our basics sharp”, precisely so you can tell the savant from the idiot. Which is why he wants literacy and numeracy taught intensely even where AI already does them better, and calls it a no-regrets investment: one that pays off whatever the future turns out to be. The backdrop is not reassuring. Since 2009, literacy and numeracy have been falling among young people worldwide, according to the OECD’s PISA programme, and among adults too.

In a Harvard physics study, a purpose-built AI tutor designed with proper scaffolding beat in-person active learning by 0.73 to 1.3 standard deviations, two to three times the usual bar for a substantial effect in education research. In the Türkiye study, the standard ChatGPT left students worse off once it was removed, while a guardrailed tutor version avoided that loss entirely. Same models, opposite outcomes, depending on how they were designed.

In a controlled study in a university creative-coding course, students received identical AI-generated feedback on their work. Those told it came from a human teaching assistant ran their code more, wrote more code, and spent more time on later work. They rated the feedback equally helpful either way. The effect on effort was large (d = 0.88 to 1.56). The content was the same.

Education does three things at once: it builds knowledge and skills (qualification), it helps you find your place among others (socialization), and it helps you become someone who thinks independently and takes responsibility (subjectification). AI tools mostly reach the first. They rarely, if ever, address the other two, and those are where a teacher does their deepest work.

An LLM answers your question. Faster work, less learning.

An AI tutor asks questions back, no matter what you actually need. Often frustration and drop-out.

An AI learning companion (Dr Philippa Hardman calls it a “study mate”) remembers where you got stuck and pushes you towards the thinking you avoid. Capability that lasts.

Holmes (UNESCO) asks it directly: if generative AI is this powerful, do we still need to learn? His answer is yes, and the list grows: on top of what we wish to learn, we now need to learn AI’s profound limitations, its impacts on human rights, social justice and the environment, and “perhaps most importantly, [to] learn how to think… critically.” The World Economic Forum (WEF) keeps the balance: rote memorisation may matter less, but “the process of mastering knowledge continues to develop broader capabilities” such as grit, curiosity, communication and critical thinking, and assessment must evolve to capture them.

The tool: the guardrailed tutor and the learning companion are instructional design built into software: scaffolding, answers withheld, help that fades as you grow.

The classroom: co-design tools with teachers rather than deploying them on teachers, and set tasks that make learners compare, justify and revise what the AI produces. In the 67-study review, that scaffolding is what separated gains from cognitive offloading.

The system: AI only helps where the conditions are ready. As the WEF puts it, “learning outcomes will not be determined by technology itself, but by the conditions in which it is deployed”, and isolated fixes across policy, pedagogy and technology are unlikely to be sufficient.

In August 2026, the system layer got its first institution-scale test case. MIT’s committee on AI in teaching and learning refused patching (“this is not a moment for patches and duct tape”) and called for all three moves in concert: AI-aware redesign of every subject, an in-person social component in every subject, and permanent adaptation structures (a standing committee, per-school AI leads, a pilot fund), on a platform deliberately tied to no single AI vendor. One survey number shows what is at stake: MIT undergraduates report feeling more replaceable than capable (40% versus 34%).

Susskind adds the one move here with a precedent. In 1982 a British government report by Wilfred Cockcroft answered the electronic calculator not by banning it but by splitting mathematics in two: part of the time learning to work with a calculator, the rest learning to cope without one, and, the decisive part, examining both. That split is now the standard almost everywhere. He proposes the same structure for AI, in every subject, and calls it teach both, test both. His reason for putting the weight on assessment rather than detection is the part worth keeping: a teacher cannot know whether a student used AI alone in their bedroom, but nobody forgets sitting an exam they have only half prepared for.