Bolt 34 — How Fast Should You Update What You Believe About Yourself?

Bolt Series · Human Performance Calibration · Article 34

There are two ways to be badly calibrated

A student has believed for years:

“I’m bad at Mathematics.”

Then the evidence changes.

78%.

82%.

85%.

Still the student says:

“No. I’m still bad at Mathematics.”

The self-model is updating too slowly.

Now imagine another student.

They usually score around 65%.

One unusually favourable paper comes back at 92%.

The student announces:

“I’ve mastered the subject.”

The self-model is updating too quickly.

Good calibration needs something in between.

Change your belief when the evidence changes — but change it by an amount the evidence has earned.

In learning science, the update itself can be studied

Recent research on confidence and self-beliefs shows that these estimates are not fixed.

They move in response to feedback.

A 2025 study modelled how people changed confidence after trial-by-trial prediction errors. Confidence rose after repeated feedback that was better than expected and fell after repeated feedback that was worse than expected, even when objective performance itself was held unchanged by the experimental manipulation.

This is a useful distinction.

Performance and the estimate of performance can move separately.

The update process is itself part of calibration.

But people do not update like neutral calculators

New evidence enters an existing self-model.

And existing self-models can resist change.

A recent study of self-belief formation and revision found that initial expectations shaped the beliefs people formed about their own ability. As confidence grew, those beliefs became increasingly resistant to later contradictory feedback.

Other research on self-relevant feedback has found a tendency to integrate feedback that fits existing self-views more readily than feedback that conflicts with them.

This creates a serious calibration problem.

Once a person has decided:

“I am good at this.”

or:

“I am bad at this.”

new evidence may no longer receive equal treatment.

Confidence should affect updating — but not freeze it

If a learner has twenty high-quality performances supporting one conclusion, one contradictory result should not have the same effect as if the learner had only one previous observation.

Strong prior evidence should create some stability.

But stability is different from immunity.

A 2024 study found that higher confidence was associated with reduced feedback processing and less subsequent belief updating in a probabilistic learning task.

That makes functional sense up to a point.

If every piece of weak evidence could overturn a strongly supported belief, the system would become unstable.

But if confidence shuts feedback out completely, the system becomes uncorrectable.

A good self-model should be stable enough to resist noise and flexible enough to respond to signal.

Source reliability matters

Suppose two people give you conflicting feedback.

One has observed you carefully across twenty performances.

The other saw one attempt from across the room.

Those pieces of feedback should not receive equal weight merely because each arrived as one opinion.

Research on belief updating under conflicting information shows that people can appropriately incorporate source reliability into revision, discounting information from less trustworthy or corrected sources.

For education, the source is not only a person.

The source can be:

a carefully designed examination;
a casual worksheet;
a teacher who has observed the learner for a year;
a new tutor after one session;
the learner’s own confidence;
a repeated pattern across tasks;
one unusually good or bad day.

Good updating asks what each source deserves to contribute.

Repetition should increase the update

One surprising score says:

“Look again.”

Five similar scores across varied tasks say something stronger:

“Your previous model is probably outdated.”

Repeated evidence matters because it reduces the chance that we are reacting to noise.

But repetition should be meaningful.

Five copies of the same heavily scaffolded task may tell us less than three varied independent tasks that test the same underlying capability.

The size of the surprise matters too

Suppose a learner expects 80 and gets 78.

Small prediction error.

The current model probably needs little change.

Now suppose the learner expects 90 and gets 38 on a well-matched independent task.

That is a much larger mismatch.

The evidence is demanding a more serious investigation.

The right response is still not:

“You are now a 38% person.”

It is:

“A high-quality observation landed very far from our prediction. The model needs substantial testing and probably substantial revision.”

Your state can change the update rate too

Even the machinery that updates beliefs is state-dependent.

A 2026 study found that total sleep deprivation disrupted the normal relationship between initial confidence and belief revision after social feedback. Participants changed beliefs more readily and showed larger confidence shifts when sleep deprived.

We should not turn that laboratory finding into a classroom diagnostic rule.

But it gives us an important general lesson.

The person evaluating the evidence is also a changing system.

Major self-judgements made while exhausted, distressed or highly activated may deserve later re-examination.

A practical update rule for learners

When new evidence arrives, ask:

Relevance: Does this task actually measure the capability I am updating?
Independence: How much of the performance did I produce myself?
Reliability: How trustworthy is this source?
Repetition: Has this happened more than once?
Variation: Does the pattern survive different tasks and conditions?
Magnitude: How far was the result from what I predicted?
Mechanism: Can I explain why the mismatch occurred?

Then update proportionately.

Not zero.

Not everything.

What the evidence earned.

Why this matters for education

Children are constantly building stories about themselves from educational events.

One failure can become:

“I cannot do this.”

One success can become:

“I do not need to practise anymore.”

And an old school label can survive long after performance has changed.

Education should teach a better update discipline.

Do not protect an old self-belief from good evidence. Do not let one noisy event rewrite an entire self-model. Learn how much each receipt deserves to move you.

That is what a calibrated person does with surprise.

They update.

At the right speed.