Why Feeling Sure Is Not the Same as Being Right

You answer a trivia question in two seconds. You feel sure. You are wrong. Anyone who plays quizzes has been there, and it is a good way into an idea people call the Dunning-Kruger curve.

What the Dunning-Kruger curve claims

The classic finding dates to 1999. Kruger and Dunning reported that the less people know, the more they tend to overestimate their knowledge. Even so, low performers still rate themselves below people who know more, as Steven Novella summarises. The pattern has been reported in groups as different as medical residents, tournament chess players, debate teams and hospital lab technicians, according to The Psychologist at the BPS.

The "curve" is the chart behind all this. Players are sorted by actual score, and their self-ratings are plotted against that score. The bottom group sits well above the line. The top group sits a little below it. It is a tidy picture, and tidy pictures travel fast.

Why confidence feels like accuracy, and why that story is disputed

The original explanation is simple. The skills needed to give a right answer are often the skills needed to spot a wrong one. Without them, you may not see your own mistakes. The BPS piece describes this as faulty self-evaluation caused by deficient cognition.

That explanation is now contested. In The Psychologist, Robert McIntosh and Sergio Della Sala argued that much of the effect is a statistical artefact. David Dunning replied that their discussion covers only a small part of more than 20 years of research. Both sides are on record at the BPS.

A 2026 study in Psychological Review, by researchers at the University of Bath and the London School of Economics, pushes harder. Its authors say the pattern comes from flawed analysis. When it is corrected, the most able people turn out to be the most overconfident. They also call the popular shorthand, that low performers are "unskilled and unaware of it", an internet simplification. See the University of Bath announcement and the Phys.org coverage. The coverage we could access gives no effect sizes, so the size of the correction is unclear.

One piece of the artefact argument is easy to grasp. It is regression to the mean. A very low score is often partly bad luck with the questions, so the person's self-estimate looks too high by comparison. A very high score is often partly good luck, so the estimate looks too low. Novella adds that random data can produce a similar graph, though a more symmetrical and less pronounced one.

So the safe reading is modest. Feeling sure is weak evidence of being right. Whether that is mainly a skill gap or mainly noise in the scoring is still argued over.

Testing your own calibration with puzzles

You can run a small check at home. Answer a set of questions. Next to each answer, write a confidence rating from 1 to 5. Then score the set and compare. Did your 5s turn out right more often than your 2s?

Here is a simple way to run it:

1. Answer each question before you look anything up.
2. Rate your confidence right away, before you see the result.
3. Mark each answer right or wrong afterward.

After a few days, read the log with these checks:

1. Count how often your top-confidence answers were right.
2. Count how often your lowest-confidence answers were right anyway.
3. Look at the trend across days, not any single day.

Aim for at least 30 to 50 rated answers before you read anything into the result. That is a rule of thumb, not a figure from the research. Ten answers can swing on luck alone.

A ready-made question set

Use questions with a single checkable answer. Here are some examples:

• Which is longer, the Nile or the Amazon?
• How many sides does a heptagon have?
• Is 91 a prime number?
• Which year did the first Moon landing take place?

Rate each one from 1 (pure guess) to 5 (certain) before you check.

A worked example

Imagine 40 answers over a week. These numbers are invented to show the method. Your 5-rated answers: 12, of which 9 were right, or 75 percent. Your 3-rated answers: 16, of which 8 were right, or 50 percent. Your 1-rated answers: 12, of which 3 were right, or 25 percent. Accuracy rises with confidence, so your sense of certainty is tracking something real. If the 5s were right only half the time, that gap is the thing to watch. Also look at the 1s. If they are right more often than you expect, you may be underrating yourself.

Limits of the check, and what the research asks of you

Your log has the same weakness that critics find in the classic studies. The BPS piece describes double-dipping. The typical study uses one task score both to sort people into skill groups and to judge how accurate their self-assessments are. That can produce the pattern by itself. At home, the equivalent is judging yourself on a single day and then sorting your days into good and bad by the same results.

Luck matters too. A badly phrased question can sink a confident answer. Bath and LSE researchers point out that a low score is often partly bad luck. So do not read one wrong 5 as proof of anything. Look for a pattern across many rows.

Two questions are worth asking of any article on this topic, including this one. Does the study use the same score to group people and to judge their accuracy? Does the analysis account for randomness in the scores? An article that presents the curve as settled, with no mention of the 2026 reanalysis or the artefact critique, is leaving out half the story.

The sources we could read do not explain how to apply these findings in everyday decisions. They also do not report how the effect holds up for casual puzzles. Your log will not settle the academic dispute. It will tell you something narrower and more useful: on these questions, this week, how well your confidence matched your results.