The Dunning-Kruger effect isn’t what everyone thinks it is

Justin Brown, Co-Publisher, Silicon Canals, 5 August 2026. A popular corrective: what Kruger and Dunning measured in 1999, what the internet turned it into, and the twenty-odd years of argument about whether the effect is real.

What the 1999 study found

Four studies on humour, reasoning and grammar. Participants in the bottom quartile averaged around the 12th percentile in actual performance and put themselves at roughly the 62nd — a gap of about 50 percentile points. Top performers were miscalibrated too, in the other direction: they underestimated where they stood against everyone else.

The proposed explanation was metacognitive. The skills needed to do a task well overlap with the skills needed to judge whether you did it well, so the people worst placed to perform are also worst placed to notice.

What it did not find

The article’s useful half is the list of things the paper never said.

It never tracked anyone over time. Performance and self-estimate were measured at a single sitting. Every version of the story that describes a beginner’s confidence rising and then collapsing is describing a study that was not run.
The “Mount Stupid” curve is not in the paper. That peak-then-valley graph, the form in which most people meet the effect, was drawn later by someone else.
It did not conclude that confident people are stupid, or that able people are naturally modest.

The case against the effect

Three challenges, in the order the article gives them:

Krueger & Mueller (2002): a general better-than-average tendency plus ordinary statistical regression could produce the same asymmetry without any metacognitive deficit.
Burson, Larrick & Klayman (2006): across twelve tasks, difficulty flips the pattern. On hard tasks the top performers can be the worse judges of their relative standing, which a competence-tracks-self-knowledge account does not predict.
A 2022 analysis in Frontiers in Psychology: noisy self-judgments on a bounded scoring scale reproduce the classic Dunning-Kruger shape on their own. Someone scoring near the floor has room to err in one direction only.

That third one is a different kind of objection from the first two and worth keeping separate: it does not say the effect fails to appear, it says the effect appears in data with no effect in it.

And the case for it

A 2021 replication published in Nature, roughly 4,000 participants in each of two studies, came down on the metacognitive side for grammar and logic tasks.

A separate five-study paper (1,189 participants) found intellectual humility tracking curiosity and openness to learning rather than raw cognitive ability. That cuts against the folk version from the other end, since the folk version has the smart people being humble.

What this source does and doesn’t give you

It names six studies with years, and authors for three of them. That is far better than does-language-determine-thought managed on a comparable subject, and it is why this sits at T3.

The withholding matters anyway. The 2022 Frontiers analysis gets no author. The 2021 replication is credited to Nature with no authors and no title, and Nature has several journals whose names begin that way, so the strongest evidence in the piece is the item a reader can least easily go and check. The five-study humility paper gets neither author nor venue. Every number here is reported at second hand.

The publisher states that AI tools assist drafting, research and editing while editorial judgment stays human, and that factual claims are sourced and linked. The URL slug and the headline-as-full-summary format are consistent with that workflow. Nothing in the piece is checkably wrong; nothing in it has been checked here either.

How it held up against the papers (2026-08-05)

Both primaries were fetched the same day, at the curator’s instruction: kruger-dunning-1999 and magnus-peresetsky-2022. The scorecard for this source is better than its tier implies, with one real hole.

Accurate: every number checks out — the 12th and 62nd percentiles, the single sitting, the absence of “Mount Stupid,” the twelve tasks in Burson et al., the ~4,000 participants per study. The Krueger & Mueller and Frontiers attributions are correct.

Slightly overstated in the retelling. “Noisy judgments and bounded scoring scales can reproduce the classic shape” invites the reading that the effect was generated from noise alone. The actual paper fits a censored model to 665 real exam predictions and argues the fit needs no psychological parameter. The page that took the looser reading from here has been corrected.

The hole. By presenting the statistical case as the counterweight to the 1999 finding, this piece leaves the impression that Kruger and Dunning had not considered it. They had, in print, twice — and Study 4’s training manipulation was built specifically to answer it. Nothing here is false; the omission is what makes the dispute look more settled than it is.

The unnamed 2021 Nature replication remains unverified, since the source gave no author or title to chase.

Connections

The spoke’s index note after the 2026-08-04 batch asked for sources from outside one essayist’s reading list. This is one, and it lands where the corpus was thinnest — the experimental end, arguing about a specific published result.

It also did the thing a good secondary source is for: it named enough studies precisely enough that the primaries could be found and read the same day, which is how this spoke got its first two T1s.

dunning-kruger-effect · heuristics-and-biases · replication-crisis · silicon-canals · justin-brown · thinking-fast-and-slow · synthesis