What the Study Found
- The classic Dunning-Kruger pattern is largely a statistical artefact of sorting people by luck-laden test scores, the authors argue.
- Reanalysing about 4,000 online test-takers, corrected slopes ran slightly above one (1.15 and 1.13), so overconfidence rose with ability.
- The reversal held even assuming half the variation in people’s self-estimates was random noise, a deliberately extreme test.
- A signalling model frames confidence as a costly, hard-to-fake advertisement of hidden ability, paired with hidden loss aversion.
Here’s the idea that launched a thousand memes: the least competent people are the most sure of themselves. It has a name, the Dunning-Kruger effect, and for decades it has been the internet’s favourite way to explain a bad boss, a confident fool, a stranger holding forth online about a subject they plainly know nothing about. Now two economists say the pattern is mostly a mirage. When you clean up the statistics, they argue, it flips: the people who overrate themselves the most tend to be the ones who are actually good.
The claim comes from Chris Dawson at the University of Bath and David de Meza at the London School of Economics, writing in the peer-reviewed journal Psychological Review. Their target is not the idea that people misjudge themselves, which nobody disputes, but the shape of the misjudgement.
To see what they’re talking about, you have to look at how the original effect is measured. Take a test, then ask people how well they think they did, and sort everyone by their actual score. The low scorers, on average, wildly overestimate. The high scorers are close to the mark, sometimes even a touch modest. Plot it and the story looks clear: incompetence breeds confidence, competence breeds humility. That’s the picture hundreds of studies have reproduced, off the back of Kruger and Dunning’s original 1999 paper, which has since been cited in more than 8,500 scholarly publications, and it is the picture Dawson and de Meza reckon is built on sand.
Why? Because a test score is a noisy thing. Some of it is skill; some of it is luck, a kind question here, a badly worded one there.
And that noise, they argue, does something sneaky when you sort people by their scores. Anyone who lands at the very bottom is disproportionately likely to have been unlucky, which means their true ability is higher than the score says, which in turn makes their own estimate look wildly inflated by comparison. The mirror image happens at the top. Exceptional scores are padded by good fortune, so those people look almost prophetic about their own ability when really they just caught a break. Sort by an outcome that contains luck, in other words, and you manufacture exactly the pattern Dunning and Kruger reported.
De Meza puts it in terms of volatility: performance swings around, buffeted by luck and by the phrasing of questions, and slicing people up purely by test score mathematically distorts the data until a false psychological phenomenon appears out of nowhere. It’s a statistical illusion, Dawson says, adding that the popular image of the incompetent person brimming with confidence “tells us less about human psychology than it does about the hidden traps in statistical analysis.”
Dawson and de Meza aren’t the first to level this charge, and it strengthens their case that they’re not. A line of critics has been picking at the effect’s statistics for a decade: Edward Nuhfer and colleagues argued in papers in 2016 and 2017 that much of the pattern is an artifact of regression to the mean and of autocorrelation, the trap of letting the test score contaminate both axes of the chart, and Gilles Gignac and Marcin Zajenkowski made a similar case in 2020, showing the effect shrinks sharply once continuous-variable methods replace the original quartile buckets. What Dawson and de Meza add is not the accusation but a specific correction, and a claim about what is left once the artifact is removed.
What Happens When You Correct the Statistics
So the pair went back to the data. Not new data, mind: they reanalysed an existing public dataset from a 2021 study by Jansen and colleagues in Nature Human Behaviour, in which roughly 4,000 people recruited online sat two sets of 20 multiple-choice questions, one on grammar and one on logic, and then guessed their own scores. The standard approach sorts by test score and reads off the gap. Dawson and de Meza instead used a regression method, called reduced major axis regression, that treats the wobble in people’s self-estimates and the wobble in their test scores even-handedly, rather than pinning all the randomness on one side. Do that, and the classic downward slope inverts. The relationship between how good people actually were and how much they overrated themselves came out slightly above one, at about 1.15 for grammar and 1.13 for logic, meaning overconfidence grew, gently, with ability.
They also checked how fragile the result was. Even when they assumed, implausibly, that random noise accounted for half the variation in people’s self-estimates, the corrected pattern held.
The authors are careful to frame the evidence as consistent with their account rather than proof of it, and it rests on a single dataset of online test-takers whose guesses were never incentivised for accuracy. That’s a long way from every domain in which humans rate themselves. It is also worth saying that the wider replication record is mixed rather than settled in either direction: in the domain of creativity alone, one 2024 study found the classic effect held up under simple methods but not under the more advanced statistical tests, while another group, testing the same year, reported finding it after all.
There is also a gap between the paper and the press release announcing it, which billed the finding as turning Dunning-Kruger on its head and reached for the old shorthand about people too stupid to know they are stupid. The paper is more measured. It allows that poor performers may sometimes fail to see their own limitations, and presents the reversal as a statistical correction, not a slogan.
Why Capable People Still Bluster
Then comes the second, stranger half of the argument. If overconfidence is not the preserve of the clueless, why do capable people inflate their own abilities at all? Dawson and de Meza’s answer borrows from evolutionary biology and the economics of signalling. Confidence, they propose, is a costly advertisement of hidden ability. Bluster carries a price, because acting on inflated beliefs leads to bad decisions, and that price is lower for the able than for the hopeless. So sincere self-belief becomes a signal others can more or less trust, not because it never lies, but because the lie is expensive. Bolted onto this is loss aversion, our tendency to fear losses more than we relish equivalent gains, which reins in the worst decisions overconfidence would otherwise produce, all while staying hidden enough not to give the game away. Talking the talk, in short, without walking the walk.
If they’re right, a lot of well-meaning advice needs a rethink. Daniel Kahneman argued that a good organisation should stamp out both overconfidence and loss aversion. Dawson and de Meza suspect the two evolved together, as a package, each softening the other’s edges, and that pulling one out might do more harm than good. Whether the wider field buys the signalling story, or just the demolition of the old one, is now the interesting question. For a psychological effect that spent a generation as a punchline, being the wrong way round would be quite a twist.
- Study type: Theoretical paper (evolutionary signalling model) with an empirical reanalysis of existing data; peer-reviewed, published in Psychological Review (advance online publication)
- Sample size: Reanalysis of approximately 4,000 online participants from Jansen et al. (2021); the model itself has no sample
- Model: Costly-signalling game in which overconfidence signals ability and hidden loss aversion offsets its decision costs
- Inputs and assumptions: Grammar and logic test scores plus self-estimates; ordinary least squares and reduced major axis regression treating noise in both symmetrically
- Period covered: Reanalysis of a 2021 dataset; test items were 20 grammar and 20 logic multiple-choice questions
- Funding / conflicts of interest: No specific grant reported; authors declare no conflicts of interest. Open access under CC BY 4.0, University of Bath
- Data availability: Underlying data from Jansen et al. (2021) public on the Open Science Framework; analysis script publicly posted by the authors
- Preregistration: Not preregistered
- Main limitation: The empirical evidence is described by the authors as consistent with the model’s necessary conditions rather than proof, and rests on a single dataset of online test-takers whose self-estimates were not incentivised for accuracy
Reference
Dawson, C., & de Meza, D. (2026). Talking the talk, not walking the walk: The coevolution of overconfidence and loss aversion. Psychological Review. https://doi.org/10.1037/rev0000644
Frequently Asked Questions
Is the Dunning-Kruger effect actually real?
The Dunning-Kruger effect may not be real in the way it is usually understood. Two economists argue that the classic pattern, in which low performers are the most overconfident, is largely a statistical artefact of sorting people by noisy test scores. Once they correct for that noise, the pattern reverses, and the most capable people look like the most overconfident.
Why would sorting people by their test scores create a false pattern?
Sorting people by their test scores can create a false pattern because a score mixes real ability with luck. Whoever lands at the very bottom is disproportionately likely to have been unlucky, so their true ability is higher than the score suggests and their self-estimate looks wildly inflated by comparison. The reverse happens at the top, and the result is a manufactured link between low scores and high confidence.
Does this mean smart people are the most arrogant?
It does not quite mean smart people are the most arrogant. The reanalysis suggests overconfidence rises gently with ability rather than falling with it, but the authors stress the evidence is consistent with their account rather than proof, and it rests on a single dataset of online test-takers. It is a correction to a famous claim, not a new law of human nature.
Why would capable people overrate themselves at all?
Capable people may overrate themselves because confidence works as a costly signal of hidden ability. Acting on inflated beliefs leads to poor decisions, and that cost is lower for able people, so sincere self-belief becomes a signal others can more or less trust. The researchers pair this with loss aversion, which curbs the worst decisions overconfidence would otherwise cause.
Cite This Page