Unskilled and Unaware of It

Kruger, J. & Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 77(6), 1121–1134. Both at cornell-university at the time.

This wiki’s first peer-reviewed source. Read in full, from the paper rather than about it, which closes the standing gap named in synthesis since the spin-out.

The claim

People overestimate themselves in many domains, and the unskilled do it worst because of a dual burden: their incompetence produces the poor performance and removes the metacognitive means of noticing it. The headline figure is the one everyone quotes: bottom-quartile participants scored at the 12th percentile and placed themselves at the 62nd.

The paper opens with McArthur Wheeler, who robbed two Pittsburgh banks in 1995 with lemon juice on his face, believing it would make him invisible to security cameras.

Four predictions, four studies

The structure people skip. Kruger and Dunning did not report one correlation; they stated four predictions and built a study around each.

1. The incompetent will dramatically overestimate their ability against objective criteria.
2. They will be less able to recognize competence — their own or anyone else’s.
3. They will fail to correct themselves through social comparison, because they can’t read others’ performances either.
4. Making them competent will make them aware — insight arrives with skill, not before it.

In Study 2 the bottom quartile did not merely misrank themselves, they placed themselves above average in absolute terms, and overestimated their raw number of correct answers by nearly 50%.

The two studies that do the real work

Study 3, phase 2 — recognizing competence in others. Bottom- and top-quartile performers came back weeks later, graded five peers’ tests, then re-rated themselves. The bottom quartile could not tell good performances from bad, and seeing their peers’ superior work did not move their self-assessment at all. The top quartile did revise — upward.

That asymmetry produced the paper’s second finding, which has its own section, The Burden of Expertise. Top performers underestimated their relative standing while judging their absolute score well. Kruger and Dunning read that as the false-consensus effect: having found the task easy, the able assumed everyone else had too. So the two errors have opposite sources — the incompetent are wrong about themselves; the competent are wrong about other people.

Study 4 — the training manipulation, and the reason it exists. 140 participants took a ten-item Wason test; 70 then got a short logic-training packet and 70 got a filler task of equal length. The trained bottom quartile (n = 19) moved from an estimated 55th percentile of ability to the 44th, from the 51st to the 32nd on test score, and from “I got 5.3 right” to “I got 1.0 right.” The untrained bottom quartile (n = 18) did not budge. Metacognitive accuracy in the trained bottom quartile rose from 3.5 to 9.3 out of 10, statistically level with the untrained top quartile’s 9.9.

They knew about the regression objection, and Study 4 is the answer to it

Worth stating plainly, because dunning-kruger-misread presents the statistical critique as though it landed on authors who had not considered it. They considered it in the paper, twice.

After Study 1: “At first blush, the reader may point to the regression effect.” Their first reply is that regression predicts symmetric miscalibration and the data are not symmetric — bottom-quartile participants were nearly four times more miscalibrated than top-quartile ones.

Their real reply is the design of Study 4, and they say so: manipulating competence after the test is something a regression account cannot accommodate, because regression predicts no change in self-appraisal from a training packet that arrives once the score is already fixed. In their words, the training result is “inconsistent with a simple regression effect interpretation.”

Whether that answers the bounded-scale version of the objection is a separate question, and magnus-peresetsky-2022 is where this wiki takes it up. It does not answer it, because that paper never addresses Study 4.

What the paper does not contain

Confirmed by reading it, not inferred:

• No time axis and no learning curve. Every study measures performance and self-estimate in one sitting. Study 4’s training is an experimental manipulation across about ten minutes, not a record of anyone acquiring a skill over months.
• No “Mount Stupid.” Four figures, all of them percentile-against-quartile plots. No peak, no valley, no confidence trajectory.
• No claim about intelligence. The subject throughout is skill in a specific domain and the ability to judge one’s own work in it.

Limits worth carrying

Every participant is a Cornell undergraduate taking a short test for course credit, quartile cells run from 11 to 37 people, and the humor study depends on a criterion the authors themselves call shaky — the professional comedians rating the jokes agreed only imperfectly with each other. The paper is candid about that last one, which is why Studies 2 and 3 moved to domains with definite right answers.

The section Incompetence and the Failure of Feedback is the most speculative part and reads that way: the authors ask why life does not teach the unskilled better, and answer with politeness norms and the scarcity of corrective feedback, citing others rather than testing it.

Also on the page, and pleasant given this wiki’s standing note that Amos Tversky remains its most deferred entity: Richard Nisbett, quoted by personal communication, on Tversky — “The quicker you realize that Amos is smarter than you, the smarter you yourself must be.”

Related

dunning-kruger-effect · magnus-peresetsky-2022 · dunning-kruger-misread · heuristics-and-biases · david-dunning · justin-kruger · cornell-university · replication-crisis · synthesis