AIをどう使うかで、学びの残り方は変わるRandomized controlled trials show that passively delegating to AI reduces…

Many developers know the feeling: after a productive day building software with an AI assistant, they close their laptop only to realize they can barely recall the code they just shipped. A growing body of empirical research suggests this sensation is not just fatigue or imagination. Randomized controlled trials indicate that when the goal is learning, leaning on AI can quietly erode how much a person actually retains, which matters for anyone who cares about long-term skill and not just short-term output.

The central finding reported across these studies is narrow but pointed. For tasks whose explicit purpose is learning, using AI tends to reduce the amount learned, and it does not reliably speed the work up either. In other words, the intuitive trade-off many people assume they are making, sacrificing depth of understanding in exchange for velocity, may not hold. In these controlled settings the expected productivity dividend often failed to appear, leaving learners with both a weaker grasp of the material and no meaningful time savings to show for it.

What makes the research more nuanced, and more actionable, is that identical access to AI produced dramatically different outcomes depending on how people used it. Participants who passively handed tasks off to the model, effectively delegating the thinking, scored on average between 24 and 39 percent. Those who engaged more actively, treating the AI as something to interrogate, challenge, and verify, scored between 65 and 86 percent. For comparison, the group that used no AI at all landed around 67 percent. The high-engagement AI users therefore matched or exceeded the no-AI baseline, while the passive delegators fell far below it.

The practical takeaway is that the decisive variable appears to be not whether AI was used, but how the user engaged with it. Dialogue and verification, asking the model to explain its reasoning, questioning its assumptions, testing its output, and reconciling it against one's own understanding, seem to preserve or even enhance learning. Passive acceptance of generated answers, by contrast, is associated with the steepest declines. This aligns with a well-established idea in learning science often described as the generation effect or desirable difficulty, where the effort of retrieving and constructing knowledge is what makes it stick, and where offloading that effort can hollow out the result.

This distinction is especially relevant given how deeply AI coding tools have embedded themselves into everyday workflows. Editors and assistants such as Cursor, GitHub Copilot, and similar tools now generate substantial portions of the code many developers produce, and their fluency makes passive acceptance the path of least resistance. The convenience is real, but the research implies that the same tool can either support or undermine a developer's growth depending entirely on the mode of interaction. A tool that autocompletes an entire function is doing the generative work that would otherwise build the user's mental model.

It is worth being precise about the boundaries of these results. The studies specifically examined tasks framed around learning, so the conclusions apply most directly to situations where building durable understanding is the goal, such as studying a new framework, onboarding to an unfamiliar codebase, or mastering a concept. They do not necessarily generalize to routine production work where the aim is simply to ship a well-understood feature quickly, and where retention of every detail is less important. Readers should also treat any single set of percentages as indicative rather than definitive, since effect sizes in this kind of research can vary with task design and population.

For teams and individuals, the implication is a shift in habit rather than a rejection of the technology. When the objective is to learn, the evidence favors an approach that keeps the human in an active, questioning role: prompting the AI to justify its choices, manually reworking or rewriting generated code, verifying claims against documentation, and pausing to articulate why a solution works. Used that way, AI appears capable of matching or beating the no-tool baseline. Used as a substitute for thinking, it is likely to leave learners worse off, both in what they know and, perhaps counterintuitively, in the time they spend. The choice of engagement, more than the choice of tool, seems to determine what remains.