What is the Dunning-Kruger effect?

The Dunning-Kruger effect is the tendency of people with less competence at a task to overestimate their own performance, because what is missing is not only the skill but the yardstick to judge it: recognizing a wrong answer requires much of the same competence needed to avoid giving one, and someone who lacks it does not know they lack it. The study that gave the phenomenon its name, by David Dunning and Justin Kruger, published in 1999 in the Journal of Personality and Social Psychology, measured students on tests of grammar, logic, and humor and asked them to estimate their own percentile. Bottom-quartile scorers averaged the 12th percentile but estimated themselves around the 62nd, while the strongest performers slightly underestimated themselves. The chart that went viral, a curve climbing to a "peak of Mount Stupid" before crashing, does not appear in that paper: it is a later addition from popular science writing, not a measured result.

See also: Automation bias · Automation bias is the tendency of whoever receives an automated output to trust it more than it deserves.

What the study actually says, and what got added to it

Kruger and Dunning's "Unskilled and unaware of it" (1999) explain the gap with a metacognitive hypothesis, not arrogance. What the paper actually reports is actual performance by quartile beside an almost flat self-estimate, not a "mountain" curve. The sharpest critique, from Nuhfer et al. (2016, Numeracy) and later Gignac and Zajenkowski (2020, Intelligence), shows that much of that pattern reproduces with random data, through two statistical effects unrelated to skill: regression to the mean, and the better-than-average effect, by which most people rate themselves above the median. The original finding stays measured, but it narrows from "the incompetent always delude themselves" to "part of the gap is a statistical artifact": self-reported confidence alone stays a weak instrument for judging competence.

An enterprise example

In enterprise AI adoption the error shows up when a team scores its own AI readiness or security posture with no external yardstick: the people with the least data governance experience tend to judge themselves ready precisely because they have not yet hit the cases that would reveal the gaps. A manufacturer that calls itself "AI ready" because it runs a modern ERP, without ever measuring the quality of its master data, finds the gap only at the first model trained on that data, not before. A serious AI maturity assessment measures instead of asking for self-scoring: incident logs, error rates, data completeness, not an open-ended questionnaire filled out by the people it is meant to judge.

Why it matters for decision makers

The practical risk is not philosophical: an internal self-assessment with no outside check tends to be optimistic exactly where the opposite is needed, and recent research extends the problem to AI tool use. A 2025 study in Computers in Human Behavior (Fernandes et al., "AI makes you smarter but none the wiser") had participants solve logical reasoning tests with a chat assistant: participants scored about three points out of twenty above a norm population, but overestimated their own score by about four points. Anyone deciding on teams, vendors, or tools should treat self-reported competence as a weak signal, always paired with a verifiable external measure: an independent technical test, a data audit, a comparison against a base rate, never the confidence alone with which someone describes themselves.