Confidence tracks how easily an answer arrives, not how often it is right

Certainty is a signal, and it’s measuring something else

Confidence feels like an estimate of accuracy, arriving alongside the answer as though the mind had checked its work. It is better understood as a reading of how smoothly the answer was produced. When the pieces fit into a coherent story and no competing account is in view, certainty is high. When retrieval is effortful or two accounts compete, it drops.

Neither of those conditions has much to do with whether the answer is correct. A coherent story can be built from very little evidence — in fact, less evidence makes coherence easier, because there is less to reconcile. This is the uncomfortable core of the thing: the confidence attached to a judgement can be inversely related to the amount of information behind it.

The best-measured version of overconfidence

Researchers distinguish several kinds, and they behave differently. Overplacement — believing you are above average — is real in some domains and reverses in others, with people rating themselves below average at difficult and unfamiliar tasks. Overestimation of your own performance varies similarly with task difficulty. Neither is the simple universal effect that popular accounts describe.

Overprecision is the sturdier finding. Ask people for a range wide enough that they are highly confident the true value falls inside it, and the true value falls outside far more often than the stated confidence allows. This appears across many kinds of question and among professionals estimating quantities in their own field. The interval people offer is too narrow, and pushing them to widen it produces only modest improvement.

Why more expertise does not automatically fix it

Expertise improves accuracy in domains where the environment is regular and the feedback is prompt and unambiguous. Weather forecasters and some medical specialties are the standard examples of well-calibrated professional judgement, and what they share is that they find out what happened, quickly, in a form that cannot be argued with.

Where the outcome is delayed, noisy or open to reinterpretation, experience builds confidence without building accuracy, because every case can be absorbed into the existing account. The person has genuinely learned something — the vocabulary, the typical patterns, the way cases present — and the part that never got trained is the mapping from felt certainty to actual hit rate.

What calibration would actually require

To know whether your confidence means anything, you need a record: predictions made in advance, with numbers attached, resolved against what happened. Grouping them afterwards by stated confidence shows whether the things you were nearly certain about came true nearly always. Almost nobody has this record, which is why almost nobody knows their own calibration, including people who make confident judgements for a living.

The absence is not incidental. Without written predictions, hindsight quietly rewrites what you expected, so the raw material for calibration destroys itself within days. The record has to be made at the time or it doesn’t exist.

There is a further wrinkle worth stating: calibration is domain-specific. Somebody well calibrated about the weather next week isn’t thereby well calibrated about staffing costs or a legal outcome, because the feedback that trained the first judgement never touched the second.

Confidence in other people is read the same way

A speaker who states things fluently, without hedges, is judged more competent and more credible, and this holds even when the content is no better. That creates an unfortunate incentive: expressed certainty is rewarded socially, so people who hedge accurately are penalised for it, and the hedging drops out of public discussion over time.

The reader’s corrective is to ask for the shape of the claim rather than its volume. What would change your mind, how wide is your range, what is the strongest case against — these questions distinguish a considered position from a fluent one, and they are uncomfortable to ask precisely because fluency reads as authority.

Living with a badly calibrated instrument

The useful adjustment is not to feel less certain, which is not under voluntary control, but to widen the ranges you commit to. When asked for an interval, take the one that feels right and extend both ends further than seems necessary; the evidence on overprecision suggests the honest interval is wider than the comfortable one nearly every time.

And treat strong certainty about an unfamiliar subject as information about the subject, not about the answer. It usually means you are seeing a simple story with nothing contradicting it yet, which is the normal state of affairs shortly before finding out more.