Overconfidence: Why You Think You Know More Than You Do

Ask a room of people to give a range they are 90% confident contains some quantity — the length of the Nile, the year a company was founded — and the true answer falls inside their range far less than 90% of the time. Classic studies found hit rates closer to 30–60%. The ranges people give are far too narrow, and they are too narrow in a way that is consistent, measurable and largely unaffected by being told about it.

That gap between confidence and accuracy is overconfidence, and it is one of the more robust findings in judgement research.

It is not one thing

Moore and Healy's contribution was to separate three effects that get lumped together and that behave differently:

Overestimation. Thinking you will perform better than you do. Strongest for hard tasks, and it frequently reverses on easy ones — people underestimate themselves on tasks that turn out to be simple.

Overplacement. Believing you are better than others. This is the "93% of drivers rate themselves above average" result, and it holds for easy, familiar tasks and reverses for hard ones — almost nobody thinks they are an above-average unicyclist.

Overprecision. Excessive certainty that your beliefs are correct — the too-narrow confidence interval. This is the most persistent of the three, the hardest to train away, and the one that matters most in practice.

Keeping them separate is useful because the advice differs. Overplacement mostly needs comparison data. Overprecision needs calibration.

On Dunning-Kruger

Worth stating carefully, since it is the most-cited and most-mangled idea here.

The popular version — incompetent people are confident, experts are humble — is not quite what the data shows. The original finding is that self-assessment correlates weakly with actual performance across the whole range: low performers overestimate substantially, and high performers tend to underestimate themselves. There is also a long-running argument that a good deal of the effect is a statistical artefact of regression to the mean plus the arithmetic of the measure.

The safe summary: people are poor judges of their own ability, in both directions, and low performers are worse. That is enough to be worth acting on, and it does not license the popular sneer.

Why knowing about it does not help

Reading this page will not make you better calibrated. Bias-awareness training has a weak record; you can know overconfidence exists and remain exactly as overconfident, because the feeling of certainty is generated by fluency rather than by evidence — the same mechanism behind the illusion of competence.

What does work is feedback with a scorekeeping mechanism. The Good Judgment Project's forecasting work found that ordinary people could become measurably better calibrated, and the ingredients were unglamorous: make explicit probabilistic predictions, record them, score them against outcomes, and repeat. Weather forecasters are among the best-calibrated professionals for exactly this reason — they make a numeric prediction daily and reality grades it within 24 hours.

Most professional judgement has no such loop. You make a call, the outcome arrives months later mixed with a dozen other factors, and your memory quietly revises what you originally predicted so it matches what happened. That last part — hindsight bias — is what makes the feedback loop close on its own without a written record.

The habit that supplies the feedback

This is the practical payoff, and it is a note-taking habit rather than a psychological technique.

Write your predictions down, with a number and a date.

  • "70% this ships by March"
  • "I think the client's real objection is budget, not scope — 60%"
  • "This candidate will be a strong hire — 80%"

Then revisit them. Once a quarter is enough.

Two things happen, both valuable. You find out whether your 70% events happen about 70% of the time, which is the only way to discover you are systematically overconfident. And you get the confidently-wrong calls — the highest-value learning material there is, because errors made with high confidence are the ones that correct permanently once feedback arrives. See the hypercorrection effect.

Without the written record, none of this works, because you will misremember what you predicted. That is not a character flaw; it is how reconstruction works — see false memory.

Other things that genuinely help:

  • Consider the opposite. Actively argue why you might be wrong. One of the few debiasing techniques with reasonable support.
  • Widen your ranges past comfort. If a range feels uncomfortably wide, it is probably about right.
  • Ask what would change your mind, before the evidence arrives.
  • Use the outside view. How long have projects like this usually taken, rather than how long this one should take?

Where this fits with notes

A prediction log is a small addition to a work journal and it is the part that improves your judgement rather than your record. Most professionals go a whole career without ever finding out whether their confidence is calibrated, because nothing in ordinary work forces the comparison.

The capture problem is the usual one: predictions are made in meetings, in conversations, on the way out of a room — not at a desk with a spreadsheet open. If recording one takes more than a few seconds it will not happen, which is what Clair Mind is designed around. The scoring can wait; the prediction cannot.

More in metacognition, the illusion of competence, and memory and focus.

More in Memory & Focus

75 more notes on this branch.