Everyone has seen the chart. A confidence line, a competence line, and an embarrassing gap on the left where the people who score worst rate themselves as nearly the best. It's the Dunning-Kruger effect, and it has become internet shorthand for the fool who's certain he's a genius.
I went back to the original numbers expecting them to wilt. They mostly held. In 1999 Justin Kruger and David Dunning ran Cornell undergraduates through tests of logic, grammar, and humor, then asked each person to guess their own rank. The bottom quarter scored around the 12th percentile and guessed they sat near the 62nd. People at the bottom of the actual ranking believed they had outscored roughly two-thirds of the room.
That description is real, and it replicates across decades of studies. What I didn't expect was that the chart built on top of it, the picture that makes the effect look like a law of physics, would come out almost identical if you fed it numbers from a random number generator.
The score is hiding on both axes
Look at how the picture gets made. The horizontal axis is your actual score. The vertical axis, in the version everyone shares, is the distance between what you guessed and what you got: self-assessment minus score. Now follow a low scorer across that plot. A low score is a small number, and it's sitting right there inside self-assessment minus score, dragging the gap upward at the bottom and shrinking it toward the top. The downward slope is baked in. It exists before a single person is tested.
Statisticians have a blunt name for this. You are correlating a variable with a negated copy of itself, so the relationship is guaranteed in advance. In 2016 and 2017, a team led by Edward Nuhfer made it impossible to ignore. They ran the standard self-assessment analysis on pure random numbers and out came the canonical Dunning-Kruger scissors: the least able random values overrated themselves, the most able underrated themselves, the famous crossing lines appearing from data that held no people and no psychology at all.
Random numbers have no metacognition. The chart produced the effect anyway.
The same algebra explains the flattering second act of the legend, where the true experts modestly underrate themselves. Run the subtraction near the top of the scale and a high score forces a small or negative gap. The humble master and the deluded beginner turn out to be one line bending under one equation.
What actually survives
So is the whole thing a mirage? Here's where I had to slow down, because it's just a statistical artifact is its own confident little story, and the cleanup has its own messy history.
The honest test is to stop subtracting and plot the two real quantities against each other: what people guessed versus what they scored. Do that and you usually get a genuine, positive line. People who know more do rate themselves higher. The line is just shallow, because almost everyone parks their self-estimate a bit above the middle. So the weakest land far above their true rank and the strongest a little below, for a reason that has nothing to do with a private blindness: hardly anyone, at any skill level, rates themselves below average. The better-than-average effect is close to universal. The low performers only look dramatic because the truth about them sits farther from the spot where everyone guesses.
In 2020 Gilles Gignac and Marcin Zajenkowski put this to 929 adults, comparing self-rated intelligence against Raven's Progressive Matrices. Using tests that keep the score off both axes, they found people mispredicted themselves by about the same amount across the entire ability range, a near-zero correlation of r = -0.05. The classic story needs the bottom to miss by more. It didn't. Their title still hedged, "(mostly) a statistical artefact," and the hedge earns its keep: a 2023 reply in the same journal argued one of their central results came from a recoding choice rather than the data. The debunking has a debunking. It stays turtles for a while.
What I'm left holding is smaller and steadier than the meme. Low performers probably do overrate themselves a little, which is mildly interesting and roughly what you'd guess. The grand version, that incompetence comes wrapped in a private blindness the rest of us are spared, is the part the evidence won't carry. And the single most shared proof of that grand version, the chart, is the weakest link in the chain, because it would draw itself out of television static.
There's a tired joke that falling for the Dunning-Kruger chart is itself a Dunning-Kruger moment. It's funnier than it is fair. The lesson I'd rather keep is duller and more portable. A figure can be real and empty at the same time: drawn from honest data, perfectly reproducible, shared a million times, and still mostly a self-portrait of the procedure that drew it. We believed the curve because it was clean. The cleanness was the tell.