AI makes learning feel easy—but the knowledge doesn't stick

The author describes a recurring pattern: after an hour of learning with an AI chatbot, the material feels obvious and well-understood. Days later, when attempting to use that knowledge, the author immediately returns to the chatbot—a sign that little actually stuck. The author explores why, focusing on a psychological disconnect: the feeling of understanding and actual understanding are two different things, and AI is uniquely effective at maximizing the first while starving the second.

The mechanism lies in how the brain decides it has learned something. Humans do not have an internal exam; instead, they rely on a feeling called "ease"—how smoothly and obviously material seems as it is encountered. But ease is a misleading signal. Re-reading notes feels productive because it runs smoother on the second pass, yet it does almost nothing for retention. Conversely, struggling to recall something cold feels slow and inefficient, yet it is one of the most effective learning tools available. Psychologists call these mismatches "desirable difficulties": the conditions that make learning feel harder tend to make it last, while those that feel easy tend to let it fade.

AI chatbots systematically remove these desirable difficulties. They hand over answers before the student has begun retrieving them from memory, generate better words than the student could produce (eliminating the work of formulating ideas), and clear confusion on contact (removing the discomfort of sitting in confusion until it resolves). The result is that AI strips out the exact friction inside which learning was hidden, while producing a stronger sense of ease than any other tool the author has used. Everything feels smooth; little of it is earned.

This has been measured. A randomized trial with about a thousand high-school students (9th–11th graders across ~50 classes in Turkey) tested the effect. One group practiced math with a plain chatbot. While the chatbot was present, that group scored 48% higher than students with no AI. Then the researchers removed the chatbot for the exam, and the same students scored 17% worse than those who had never used it. The crutch had felt like a leg. Most striking: the chatbot users thought they had learned just as much as everyone else. Perceived learning went up; real learning went down; and there was no internal signal to tell them apart.

However, the article emphasizes that AI is not inherently the problem. A second trial pointed the opposite direction. College students taught by an AI tutor beat a live, active-learning class by a wide margin. But this tutor was not a plain chatbot. Instructors had built it to require students to retrieve, generate, and work through problems themselves, with difficulty deliberately put back in. Same technology, opposite design, opposite result. The variable that predicted learning was never whether AI was involved; it was whether the difficulty survived contact with it.

The author offers practical countermeasures. After the AI explains something, close the tab and rebuild the idea from nothing—whatever cannot be rebuilt was never truly learned. Make the AI ask the questions instead of answering yours. Make it withhold the solution and just mark your attempt. Treat a suspiciously smooth session as a warning, not a result. The fundamental insight is that friction is the part easiest to hand off and also the part that is the learning. When you give away the struggle, you are not offloading the overhead around learning; you are offloading the learning itself.

The article identifies a fundamental mismatch between how human learning actually works and how AI chatbots make it feel. Psychologists, citing research from Robert Bjork's lab, have long known that the conditions making learning feel easy tend to let knowledge fade, while the struggles and difficulties that feel inefficient are often the most effective for retention—a principle called "desirable difficulties." AI chatbots, by design, eliminate exactly these difficulties: they hand over answers before the student has to retrieve them from memory, they generate polished explanations that bypass the work of formulating ideas in one's own words, and they dissolve confusion instantly rather than letting it resolve through struggle.

The data from the high-school math trial makes this concrete: students felt they had learned equally well, but the tool masked a real learning deficit. The smoothness of the experience was itself the problem—it suppressed the internal signals that would normally warn a learner that retention is failing. However, the article notes that a second trial pointed the opposite way, showing that the same AI technology, when deliberately redesigned to preserve struggle and require active retrieval, outperformed traditional instruction. This suggests the risk is not AI itself but the default design of most chatbots, which optimize for ease rather than learning outcomes.