Bayesian updating
Also known as: Belief updating, Bayesian revision
Rationally revising your confidence in a belief as new evidence comes in.
What it means
Bayesian updating is the normative procedure for revising a belief in light of new evidence by combining one's prior probability with how strongly the evidence favors that belief over its alternatives. Formally it applies Bayes' rule—posterior odds equal prior odds times the likelihood ratio—so the degree of revision depends jointly on how confident you were beforehand and how diagnostic the new data are. As a model of the ideal reasoner it specifies exactly how much, and in which direction, beliefs should move; as a description of real people it serves as the benchmark against which biases are measured. Humans systematically depart from it: they neglect base rates, are 'conservative' in updating too little when evidence is strong, over-update on vivid or recent data, and let prior attitudes distort how they read the evidence. Yet the framework remains central because it makes precise what good belief revision requires and pinpoints where intuition goes astray. It matters for forecasting, diagnosis, learning, and any domain where conclusions must track accumulating evidence.
What actually moves the belief
The direction and size of the update are set almost entirely by the likelihood ratio: how much more probable the new evidence is if the belief is true than if it is false. Evidence that is roughly as likely either way carries no information and should move nothing, however dramatic it feels; only diagnostic evidence shifts the odds. This is why two rational people who start with different priors can read the same data and still disagree, yet with enough independent, diagnostic evidence their conclusions converge. The prior is never erased, only outweighed. The most common practical error is treating a striking, memorable observation as strong evidence when it is in fact consistent with many competing explanations and so barely discriminates between them.
What the evidence shows
Phillips and Edwards's 1966 bag-and-chip experiments launched a literature finding that people revise too little relative to Bayes, a pattern named conservatism and later replicated in more than a hundred studies. But Kahneman and Tversky's work pointed the opposite way: people often ignore base rates and over-weight vivid individuating evidence. The two bodies of work looked contradictory for decades. Augenblick, Lazarus and Thaler (2025) reconciled them across lab tasks, sports betting and financial markets, finding a consistent rule: people overinfer from weak signals and underinfer from strong ones. Meanwhile Griffiths and Tenenbaum (2006) showed that for familiar everyday quantities, people's implicit priors track real-world statistics remarkably well. The human updater is neither reliably Bayesian nor reliably biased; performance depends on signal strength and task.
Where it shows up
In medical testing, a positive result for a rare condition rarely means the condition is likely, yet clinicians and patients routinely overweight it. In finance, post-earnings-announcement drift, where prices adjust too slowly to news, is conservatism operating at market scale. Intelligence analysis and professional forecasting drill explicit prior-then-update discipline precisely to counter these tendencies. In experimentation and machine learning, a well-chosen prior regularizes noisy data and prevents overreaction to a small sample. Courts wrestle with it too: the prosecutor's fallacy confuses the probability of the evidence given innocence with the probability of innocence given the evidence, mistaking a likelihood for a posterior. Any setting where conclusions must track accumulating, imperfect evidence inherits the same arithmetic and the same failure modes.
Using it in practice
The discipline is to separate two questions people habitually blur: how confident was I beforehand, and how diagnostic is this new evidence? Stating the prior explicitly, as a number or at least rough odds, stops a single vivid anecdote from hijacking the conclusion. Then ask the likelihood-ratio question directly: how likely is what I just saw if my belief were wrong? Weak or ambiguous evidence should nudge you; strong, hard-to-fake evidence should move you sharply. Finally, respect the ceiling. A posterior is only as good as its inputs, so a confidently wrong prior or a mis-estimated likelihood yields a confidently wrong answer that no amount of correct arithmetic downstream can rescue. Good updating is mostly honest bookkeeping about what you knew and what the data can actually distinguish.
Examples
A doctor who knows a disease is rare (low prior) and that a test sometimes gives false positives should remain fairly doubtful even after a positive result—yet many people, and many patients, leap to near-certainty.
One glowing review shifts your view of a restaurant more than it should, while a hundred quietly consistent ones barely move you; vividness beats diagnosticity in the update.
A forecaster who put a team's title odds at 5% should barely budge after a single good win, yet fans leap to near-certainty; the evidence is weak and the prior was low.
A spam filter that has learned most mail is legitimate needs strong spammy signals, not one suspicious word, before routing a message to junk; a single odd phrase barely shifts the odds.
A manager convinced a candidate is strong should treat one fumbled interview answer as weak evidence and adjust only slightly, but a fabricated credential is highly diagnostic and should move the judgment sharply.
First described in Thomas Bayes (posthumous, 1763); modern decision use mid-20th century.
Key references
- Augenblick, N., Lazarus, E., & Thaler, M. (2025). Overinference from Weak Signals and Underinference from Strong Signals. The Quarterly Journal of Economics, 140(1), 335-401. doi.org/10.1093/qje/qjae032
- Benjamin, D. J. (2019). Errors in Probabilistic Reasoning and Judgment Biases. In Handbook of Behavioral Economics: Applications and Foundations, Vol. 2 (pp. 69-186). Elsevier. doi.org/10.1016/bs.hesbe.2018.11.002
- Griffiths, T. L., & Tenenbaum, J. B. (2006). Optimal Predictions in Everyday Cognition. Psychological Science, 17(9), 767-773. doi.org/10.1111/j.1467-9280.2006.01780.x
- Phillips, L. D., & Edwards, W. (1966). Conservatism in a Simple Probability Inference Task. Journal of Experimental Psychology, 72(3), 346-354. doi.org/10.1037/h0023653