The Feeling of Understanding

The Dunning-Kruger effect is one of the few findings from experimental psychology that made it into ordinary conversation almost intact: people who are bad at something tend to be the worst judges of how bad they are, while people who are genuinely skilled tend to underrate themselves slightly, humbled by knowing what good work actually requires. It's a tidy, reassuring asymmetry — skill and self-knowledge drift apart in a predictable direction, and the direction at least makes sense. New research on AI-assisted reasoning finds that the asymmetry doesn't survive contact with a chatbot. It isn't that AI makes the overconfident more overconfident. It's that AI makes almost everyone overconfident by roughly the same amount, erasing the one pattern the effect was named for.

What the Aalto studies measured

Daniela da Silva Fernandes and Robin Welsch, at Aalto University, ran two preregistered experiments built to isolate exactly this question. In the first, 246 participants worked through twenty LSAT-style logical reasoning questions with a ChatGPT window open beside them, required to consult it at least once per question. They scored noticeably better than an unassisted comparison group — then estimated they'd gotten about seventeen of twenty right, when they'd actually averaged closer to thirteen. A second study, with 452 participants split into AI-assisted and unaided groups, added a real incentive: a monetary bonus for an accurate self-estimate. The bonus didn't close the gap. The unaided group showed the familiar Dunning-Kruger slope, weaker performers overestimating more than stronger ones. The AI-assisted group overestimated almost uniformly, high performers and low performers alike, and those who rated themselves as more AI-literate were, if anything, more overconfident, not less.

The chat logs explain some of why. Most participants typed a question into ChatGPT once, read the answer, and moved on without checking it against their own reasoning — what the researchers describe as a pattern of passive acceptance rather than active verification. Measured directly, their metacognitive sensitivity was close to flat: participants were about as confident when they were wrong as when they were right, which is the technical definition of not being able to tell the difference from the inside.

A name for what's happening, and company in the literature

A synthesis published this year by independent researcher Christopher Koch, reviewing the Aalto findings alongside a cluster of related studies, argues that generative AI introduces a variable the classic Dunning-Kruger framework was never built to handle: a fluent, confident, externally produced output that sits between a person's actual understanding and their self-assessment of it. People calibrate confidence by judging the quality of what's in front of them, and a well-written AI answer is, by design, high quality regardless of whether the person reading it actually grasped the reasoning behind it. The studies Koch gathers make the same point from different angles. Steyvers and colleagues found that longer, more elaborate AI explanations raised people's confidence without improving their ability to tell a correct answer from an incorrect one — more words, not more accuracy, drove the feeling of understanding. Li and colleagues found that users' confidence tracked the AI's own expressed confidence so closely that the effect persisted even after the AI was taken away, as if a borrowed certainty could be internalized and carried out of the room. Bastani and colleagues, studying AI math tutoring, found large gains while the tool was present and a seventeen-percent performance drop on a later test taken without it — though a tutor redesigned to require students to explain their reasoning first largely closed that gap, a design detail worth noticing precisely because it's the exception.

What doesn't follow, and what does

None of this is an argument that AI assistance makes people worse reasoners in the moment, or that the LSAT-style tasks these studies used generalize cleanly to every kind of thinking. The AI-assisted groups in the Aalto studies did perform better; the complaint isn't that the help didn't work, it's that the help came bundled with a self-assessment that didn't track it. It's also worth taking seriously that this isn't quite the ordinary Dunning-Kruger story continuing under a new name — the mechanism here is closer to borrowing someone else's fluency and mistaking the loan for your own balance, which is a narrower and stranger claim than "people are overconfident."

What the research converges on is a specific, almost mechanical failure point: the feeling of understanding something and the fact of understanding it have always been two different things, loosely correlated by the ordinary friction of having to work a problem out yourself. A fluent answer removes exactly the part of that process — struggling, backtracking, noticing your own confusion — that used to generate the signal telling you how much you actually followed. Take that friction out, and confidence stops tracking competence not because people got worse at thinking, but because the one instrument they'd always used to check their own thinking stopped receiving a reading at all.

Sources

1. AI makes you smarter but none the wiser: The disconnect between performance and metacognition — Computers in Human Behavior (Peer-Reviewed)
2. Users of generative AI struggle to accurately assess their own competence — PsyPost (Journalism)
3. AI use makes us overestimate our cognitive performance — Aalto University (Research Communication)
4. Beyond the Steeper Curve: AI-Mediated Metacognitive Decoupling and the Limits of the Dunning-Kruger Metaphor — arXiv (Preprint)