What the Dunning-Kruger Effect Really Says (and What It Does Not)

You have probably seen the cartoon. A curve shoots up to a peak labelled something like "Mount Stupid", crashes into a valley of despair and then climbs slowly towards real expertise. It is shared as a summary of the Dunning-Kruger effect, usually to make a point about somebody else.

That curve is not in the original paper. The real study is more modest, more interesting, and, as it turns out, harder to interpret than the meme suggests.

What the 1999 study actually found

In 1999 Justin Kruger and David Dunning, then at Cornell University, published four studies with a memorable title: "Unskilled and unaware of it". Students took short tests of humour, English grammar and logical reasoning, and then estimated how well they had done compared with their peers.

The headline result concerned the bottom quarter of performers. Their scores put them, on average, around the 12th percentile. Their own estimates put them around the 62nd. In other words, the people who did worst believed they were comfortably above average.

Kruger and Dunning offered an explanation they called a dual burden. The skills needed to do a task well are often the same skills needed to recognise a good answer. If you do not understand grammar, you also struggle to see that your sentence is wrong. They tested this directly: when participants were given a short training in the logic task, their ability to judge their own performance improved, and the weakest ones became better at recognising the limits of what they had done.

It is worth being clear about what this sample was. These were undergraduates at one American university, doing short laboratory tests. That does not make the result wrong, but it is a narrow base for a law of human nature.

The problem hiding in the graph

The famous figure in the paper shows four groups, from worst to best performers, with two lines: actual percentile and estimated percentile. The actual line rises steeply. The estimated line is much flatter, so the gap is large at the bottom and much smaller at the top.

Critics soon pointed out that this shape appears even if nobody lacks special insight. In 2002 Joachim Krueger and Ross Mueller argued that two ordinary forces could produce most of it. The first is the better-than-average effect: most people rate themselves somewhat above average on most things. The second is regression to the mean: whenever two measures are imperfectly correlated, people who score at the extremes on one tend to be closer to the middle on the other. Put people who score worst on a test at the bottom, and their self-estimates will almost inevitably sit higher.

In 2016 Edward Nuhfer and colleagues made the point in a striking way. They generated random numbers to stand in for test scores and self-assessments, and showed that common ways of plotting self-assessment data, including the quartile graph, produce patterns that invite misinterpretation even when the numbers mean nothing at all. In a follow-up with real data from 1,154 people rating their own science literacy, they concluded that people's self-assessments generally do reflect a competence they can demonstrate. Experts judged themselves more accurately than novices, but most people were not wildly off.

Testing it properly

If the classic graph cannot settle the question, what can? In 2020 Gilles Gignac and Marcin Zajenkowski proposed two statistical tests that the familiar artefacts cannot easily fake. If low performers really are worse at judging themselves, their self-estimates should be more scattered around the truth, and the link between real and estimated ability should bend rather than run in a straight line.

They tested 929 adults from the general community, who estimated their own intelligence and then completed a demanding reasoning test. Neither prediction held. The scatter was not significantly larger at the bottom, and the relationship was essentially linear. Their conclusion was careful: the effect may exist for some skills, but it is likely much smaller than previously reported.

The debate did not end there. In 2021 Rachel Jansen, Anna Rafferty and Thomas Griffiths ran a large replication of the original design, with about 4,000 participants in each of two studies, and compared formal models of how self-assessment works. Their results supported a version of the original idea: in grammar and logical reasoning, low performers did seem less able to tell whether each individual answer was correct.

So the fair summary is this. Part of the classic pattern is a statistical artefact. Part of it may be real. The real part is smaller and narrower than the internet version, and it applies to specific skills rather than to a class of people who are simply ignorant and overconfident.

What this means for you

The most useful lesson is not about other people. Almost everyone places themselves somewhat above average, and the Dunning-Kruger debate is really a debate about how much worse that bias gets when feedback is poor. That gives you something practical to work with.

• Get outside feedback on skills that matter. Your own sense of how well you did is least reliable in areas where you are still learning. A test with a right answer, or someone more experienced, is more informative than your gut.
• Check your answers one at a time. The 2021 study suggests the problem sits in judging individual answers. Asking "how would I know if this were wrong?" is a small habit with a real payoff.
• Do not use it as an insult. Calling someone a case of Dunning-Kruger is itself a confident judgement on thin evidence, which is a little ironic.
• Notice your own style. Our intellectual humility test looks at how open you are to being wrong, our self-other test compares how you see yourself with how others might, and our need for cognition test looks at how much you enjoy thinking problems through.

Confidence is not the enemy. Unchecked confidence in areas where you rarely receive feedback is where the trouble tends to start, whoever you are.