Picture a pub quiz where each team must write down, before answering, how sure they are: 60%, 80%, 95%. Now imagine the quizmaster offers a second go at any question where a team's stated confidence turns out to be shaky. Which teams take it up, and does the act of committing to a number change how they play the first attempt? That is, roughly, the question a growing body of behavioural research has been circling: when people rate their own confidence before committing to an answer, what happens to their willingness to try again?

The counterintuitive finding: pre-declared confidence increases retries

The headline result, replicated across several experimental formats, is that asking for a confidence score before a guess roughly doubles the rate at which people opt for a second attempt when the first is wrong. This runs against a naive prediction. You might expect that committing to "I'm 90% sure" would make people more defensive about being seen to fail, and therefore less likely to retry. The opposite tends to happen.

The mechanism appears to be metacognitive. Forcing a numeric confidence judgement pulls attention onto the quality of the evidence behind a guess rather than the guess itself. Participants who rate themselves at 55% have, in effect, already told themselves the answer is provisional. When it turns out wrong, the retry feels like a continuation of an acknowledged uncertainty rather than a fresh admission of incompetence. The sting is smaller because the fall was priced in.

There is a useful parallel here with Kahneman and Tversky's work on loss aversion. Losses loom larger than equivalent gains, and a wrong answer delivered with 95% certainty is experienced as a genuine loss — of face, of self-image as a competent decision-maker. A wrong answer delivered with 60% certainty is closer to a break-even. The confidence score, in other words, reframes the outcome before it happens.

Why the pre-commitment changes the first attempt too

The second-attempt effect is the eye-catching number, but the more interesting shift may be upstream. Studies using confidence-weighted scoring — where you gain more for a correct high-confidence answer and lose more for a wrong one — consistently show that people become better calibrated within a session. They stop clustering at 80% and start spreading their ratings across the range.

The calibration feedback loop

This is a textbook example of a feedback loop doing useful work. Each confidence rating is a small prediction about your own accuracy. When the outcome arrives, you get a signal: was I overconfident, underconfident, or about right? Over a run of twenty or thirty items, most people develop a rough internal sense of whether their "fairly sure" is really 70% or closer to 50%.

The behaviourist framing is instructive but incomplete. Variable-ratio reinforcement — the schedule where rewards arrive after an unpredictable number of responses — is famously powerful at sustaining behaviour, and it is the standard explanation for why uncertain-reward activities are so sticky. But confidence scoring adds something the pure reinforcement account misses: it makes the uncertainty itself the object of attention. You are not just responding to an unpredictable reward; you are modelling your own model. That reflective layer is what seems to drive the retry behaviour.

Risk-taking narrows when the number is on the page

There is a second, quieter effect. When people must attach a number to their confidence, they tend to become more conservative in what they will attempt at high confidence. The proportion of 95%+ ratings drops, and accuracy at that band rises. In plain terms: writing the number down makes people honest, and honesty makes them slightly more cautious at the top end and considerably more willing to engage at the bottom end.

For anyone designing decision environments — quiz formats, training tools, forecasting exercises, even performance reviews — that is a meaningful trade. You lose some bravado and gain a lot of usable signal about where someone's knowledge actually runs out.

A concrete example worth knowing

The clearest demonstration comes from the calibration literature that grew out of weather forecasting in the mid-twentieth century. Meteorologists who were required to state a probability of precipitation — "70% chance of rain tomorrow" — became markedly better calibrated over years of feedback than forecasters who gave qualitative terms like "likely" or "probable". The numeric commitment created an auditable record, and the record created learning.

The same structure appears in the well-documented "hard-easy effect": people are overconfident on difficult items and underconfident on easy ones, and the gap shrinks when confidence is elicited numerically and scored. Lichtenstein, Fischhoff and Phillips' classic work on calibration established much of this, and it has held up remarkably well across domains from medical diagnosis to trivia.

The retry finding is a natural extension. Once you have an auditable confidence record, a failed attempt is data rather than defeat. And data invites another go.

What this means for competitive and skill-based formats

In competitive settings where players choose their own difficulty, the confidence-first design has an obvious appeal: it lets people opt into challenge without feeling reckless. A player who rates themselves 50% and then takes a second attempt is behaving rationally and, importantly, feels rational doing it. The format legitimises the retry.

There is a caution, though. Confidence scores are only useful if they are honest, and honesty degrades when the stakes of the number itself become too high. If a low confidence rating carries a social cost — looking unsure in front of peers — people will inflate. The design lesson is to keep the confidence rating low-stakes and private where possible, and to reward calibration rather than high confidence. Scoring people on how accurate their confidence ratings are, rather than how high, flips the incentive in the right direction.

Designing for the retry, not the rescue

A second attempt offered unconditionally is a rescue. A second attempt offered after a declared confidence rating is a decision. The difference matters for how the experience is remembered. Rescues feel like charity; decisions feel like agency. If the goal is to build genuine tolerance for uncertainty — and in a world of probabilistic information, that is a real skill — the confidence-first format is doing something more valuable than it first appears.

Where to take this next

The obvious open question is durability. Does the doubled retry rate persist after the novelty of rating confidence wears off, or does it decay into box-ticking? Early indications from forecasting training programmes suggest the effect holds when feedback is timely and specific, and fades when it is vague. That points toward a design principle rather than a trick: pair the confidence score with a clear, immediate resolution, and let the loop do its work.

For anyone building assessment tools, training exercises, or competitive formats, the practical move is to test a confidence-first variant against your current design and measure two things — retry rates and calibration drift over time. The second metric is the one that tells you whether you are building a skill or just a habit.