How much should a new fact change your mind? Not as much as it feels, and not as little either. The right update depends on two things: your prior — how likely the thing was before — and how diagnostic the evidence is. Get it wrong one way and you anchor, ignoring what you just learned. Get it wrong the other way and you over-react, forgetting how rare the thing was to begin with. Bayes' rule is the calibrated middle — and it explains why a positive result on a test for a rare condition still, usually, means you're fine.
Set the base rate (how common the condition is) and the quality of the test, then read off the answer. The grid of 1,000 people shows you why the number lands where it does — and the button lets a second independent test compound the evidence.
Faced with a fresh signal, people tend to make one of two opposite mistakes — and Bayesian updating is the fix for both. The difference is entirely in how each one treats your prior and how diagnostic the evidence really is.
| Response | The prior (base rate) | New evidence | Where you land | The failure |
|---|---|---|---|---|
| Anchoring | Clings to it | Barely registers it | Stuck at the prior | You miss real news and update too slowly |
| Over-reacting (base-rate neglect) | Forgets it | Taken at face value | Jumps to the evidence | You're fooled by false alarms and lurch to certainty |
| Bayesian updating | Starts from it | Weighted by how diagnostic it is | Between the two — rarely certain | The goal: calibrated |
Our Base Rate Machine tool shows why a base rate alone can make even a 99%-accurate test mostly wrong. This tool is the next move: taking that base rate as your prior and combining it with a fresh piece of evidence to land on a posterior — then updating again when the next piece arrives. Base rates are the starting line; updating is the race.
Your prior is what you believed before (often the base rate). The
posterior is what you should believe after seeing the evidence. Bayes' rule is
just the exchange rate between them:
posterior = (sens · prior) / (sens · prior + fpr · (1 − prior)).
P(condition) = 1%.
P(positive | condition) = 90%.
P(positive | healthy) = 9%.
P(condition | positive) = ?
Most people either freeze or blurt out "90%". The conditional probabilities give the mind
nothing to hold on to.
Picture 1,000 people. About 10 have it; 9 of them test positive. Of the 990 healthy, about 89 also test positive. So ~98 people test positive and only 9 truly have it — about 9%. Same maths, but the denominator is now in plain sight.
Frequencies beat percentages because they keep the whole population — the denominator — visible. Conditional probabilities quietly hide it, which is exactly where intuition breaks down. It's the same reason the grid above is more convincing than the formula.
For each scenario you get a prior and a piece of evidence. Your call: should the belief barely move, or shift a lot? Decide, then see what Bayes actually does — and why.
Once you see it, calibrated updating is everywhere a rare event triggers an imperfect signal. In each of these, the mistake is the same: reacting to the signal without weighing the base rate behind it.
The maths is the easy part. The hard part is doing it in the moment — resisting the pull to anchor when the news is inconvenient, and the pull to panic when one alarm goes off. That's the work we do with the leaders and teams we coach: turning a rule on a screen into calibrated judgement in hiring, screening, security, and experiments.
Talk to us about coachingBefore you react to any signal, ask how likely the thing was to begin with. A low prior means even strong evidence should leave you cautious — the false alarms outnumber the real ones.
A calibrated belief lands between your prior and the evidence, not at certainty. To move further, gather a second independent read — repeated evidence compounds surprisingly fast.
Strong, diagnostic evidence earns a big swing; weak or ambiguous evidence earns a small one. Match the size of the move to the quality of the signal — no more, no less.