Mind The Friction

Studying With AI Is Too Easy. Let’s Bring The Effort Back

Learning With A Model, Not From It

Every evening after work, she opens a chat and spends an hour in it. She’s getting into UX research — a field she’s long wanted to break into. The model explains the terminology, walks her through cases, answers her follow-up questions, offers examples from the industry. After a month she feels she’s got the subject down: she knows how usability testing differs from in-depth interviews, she’s seen breakdowns of a dozen real projects, she can hold a conversation about it. A month and a half in, at a job interview, she finds she can’t confidently explain the difference between two basic methods without a hint. Not because she asked badly, but because the model answered too well.

This is a considerable simplification, of course, but it’s roughly the situation language-model users often find themselves in. According to a 2025 survey from Elon University, 51% of them name informal personal learning as their main reason for turning to an LLM. And many of the people who open ChatGPT, Claude, or DeepSeek to put questions to the model — well-structured ones, following a plan they’ve worked out in advance — believe they’re learning. As for whether they actually are: by mid-2026 enough data has accumulated to answer that question definitely — and, unfortunately, not encouragingly.

The Paradox of Ease

André Barcaui, of the Federal University of Rio de Janeiro, ran a randomized controlled study with a hundred and twenty students. One group studied material on artificial intelligence with ChatGPT; the other used traditional methods, without the model. Forty-five days later, with no warning, both groups were given a retention test. The result: the ChatGPT group scored 57.5% correct, the control group 68.5%. Eleven percentage points of difference — a medium effect size (d = 0.68), but a tangible one in practice: it’s the gap between “I remember the logic of the argument” and “I remember reading something.”

The cause is not that the students copied answers or cut corners. Barcaui explains the result through a principle cognitive science described long before language models came along: desirable difficulties. The idea is simple — for information to lodge in long-term memory, the brain needs effort: recalling, putting things in your own words, getting it wrong and trying again. When the model strips that effort away by handing over a finished, well-structured answer, the subjective sense of understanding appears, but the memory trace doesn’t set. Barcaui calls this “borrowed competence”: the knowledge is there, in a way, as long as the model is nearby, and it disperses once it’s gone. And experience with the model offers no protection — the correlation between time spent using ChatGPT and test scores turned out to be weak and statistically insignificant.

Alongside this sits another finding, no less unwelcome. Rahul R. Divekar and colleagues at Bentley University (2025) asked twenty participants to study the same material two ways: through dialogue with a language model and through ordinary internet search. The conversations with the model were noticeably richer — participants asked more questions, reflected, dug into specifics, clarified definitions. But when the knowledge actually retained was measured, no difference between the groups showed up. The authors call this the “interaction paradox”: the subjective sense of a conversation’s depth does not convert into depth of understanding.

Two studies, different designs, one conclusion: what feels like learning is not necessarily learning. And it isn’t about the quality of the model. It’s that the model, by answering well, removes precisely the friction that makes knowledge stay in your head.

Why “Explain It to Me” Doesn’t Work

Most people who use a model for self-education ask it to explain — an entirely natural move. The trouble is that the quality of the explanation does not determine the learning outcome.

Shaz Furniturewala and co-authors (2026) at the National University of Singapore show that the decisive factor in dialogue-based learning isn’t the model’s eloquence but the user’s engagement. Knowledge grows through your own cognitive activity — through the effort of formulating, through the mistake, through the second attempt — not through the quality of the text you read. A finished, brilliant, exhaustive answer from the model can even do harm, if it’s so complete that there’s nothing left to think about.

In parallel, Talita de Souza and colleagues (2026) at the Federal Institute of São Paulo show another facet of the same problem: the prompt “explain it like a teacher” is too vague to set effective learning in motion. The model does not choose a teaching strategy for you. Write “explain reinforcement theory to me” and the model will — clearly, with examples, with structure. But that mode differs little from reading a good textbook: you consume the finished product without producing anything of your own.

The instructive counterexample is a randomized experiment by the LearnLM team (Google DeepMind together with the education platform Eedi, 2025). An AI tutor working by the Socratic method delivers results comparable to a live tutor. But — and this is the crux — the teaching strategy was specified in the prompt with a high degree of concreteness: the model asks one diagnostic question at a time, withholds the answer until the student has tried, adapts the difficulty to the responses. “Be a teacher” doesn’t deliver that. A set of procedural rules by which the model runs the dialogue does.

What We Know and What We Don’t

A caveat is needed here. Nearly all the experimental data described above comes from students — young people studying a specific course within formal education. The only large-scale study devoted to informal adult self-learners — work by Nađa Terzimehić and colleagues at the Technical University of Munich (2025) — is a survey of seven hundred and seventy-six people, not an experiment. It shows that 88% of respondents already use language models in everyday learning, and that young adults lead in adopting them, but it says nothing about how effective that learning is.

Does this mean the conclusions don’t apply to an adult who, after work, digs into a new subject? Not necessarily. Desirable difficulties, the testing effect, the protégé effect — these are basic mechanisms of memory and learning, described long before language models and tied neither to age nor to format. The model doesn’t abolish them. It only makes them far easier to avoid.

NotebookLM, which students in particular often reach for as a learning tool, combines three ways of working with the same material and, taken together, neatly shows the difference between simply consuming information and learning actively. The Audio Overview feature generates a podcast from your documents: two voices discuss your material. It comes out fairly lively, but from a cognitive-science standpoint it’s passive consumption, little different from sitting through a lecture. The Flashcards mode (with links back to the source) is closer to active recall — one of the most reliable learning mechanisms — since you try to answer before the prompt appears. Finally, the interactive mode, which lets you interrupt the “podcast” with a question, is one more step toward a desirable difficulty: your question has to be formulated, which means applying cognitive effort of your own.

Two Modes That Work

From the research published by mid-2026, two strategies for learning with a model emerge, each with experimental backing.

The first is the Socratic mode: the model asks questions, you answer. Here the wording of the prompt decides everything. “Help me get my head around X” is useless — the model will simply start explaining. A working prompt is procedural: “I want to understand [topic]. Ask me one question at a time. Don’t give me the answer until I’ve tried myself. If I get it wrong, ask a leading question rather than correcting me directly.” The difference looks cosmetic, but it’s fundamental: in the second case the cognitive effort stays on your side — and that, as we’ve established, is what converts into memory.

The second strategy is the reverse mode: you explain to the model, not the other way around. Xinming Yang, Haasil Pujara, and Jun Li at the City University of New York (2025) showed that students who taught a model to solve problems posted a statistically significant improvement over those who learned from the model in the usual way. The principle isn’t new — cognitive science knows it as the protégé effect: in explaining something to someone else, a person restructures their own knowledge, uncovers gaps, and strengthens the links between concepts. The language model here is a patient, endlessly attentive “student” that never tires of asking again. The prompt might look like this: “Pretend you know nothing about [topic]. I’ll explain it to you. Ask clarifying questions if something isn’t clear, and point it out if I contradict myself.”

What unites both strategies is one thing: they don’t remove the effort, they redirect it. You do the work — the model supports it.

Fresh confirmation arrived right at the start of June 2026. In a quasi-experiment presented at the ICCE conference, students in a statistics course were split into three groups: no model, free access to a model, and regulated access — that is, governed by a set of explicit rules: prioritize reasoning over the finished answer, ask for step-by-step hints rather than the solution, check the result yourself. On tests taken without the model’s help, the group with rules outperformed the group with free access. The authors’ conclusion is stated plainly: access to a language model is, on its own, an incomplete educational intervention; the difference lies in exactly how you work with it.

Exactly how turns out to be a short list of simple rules. You don’t have to keep them in your head and recite them to the model every time you decide to learn something new — you can set them up once. Both modes, Socratic and reverse, the author of this piece has gathered into a small skill that drops into Claude, Codex, Cursor, and Gemini; for an ordinary chat window there’s the same set as a prompt, which you can paste as your opening message or save as a Custom GPT or Gem. You can grab it here, instructions included.