Plain-language answer
Bayesian updating is a rule for revising how strongly you believe something, given new evidence, that combines two ingredients: how likely you thought it was before seeing the evidence (the prior), and how much more likely the evidence makes it relative to the alternatives (the evidence’s diagnostic strength). The rule traces to a paper communicated after mathematician Thomas Bayes’s death in 1763, on updating an estimate of probability in light of new observations.(Bayes & Price, 1763)
Why it matters
A common error in reasoning about coincidences is treating a striking piece of evidence as though it should move belief all the way to certainty, in either direction — “this coincidence is too unlikely to be chance, so something more must be going on” or “extraordinary claims are always wrong, so I won’t update my view at all.” Bayesian updating replaces both overreactions with a proportional one: strong evidence should move belief a lot, weak evidence should move it a little, and — crucially — evidence that fits multiple competing explanations equally well should not move belief toward any one of them over the others.
Worked example: why the prior matters as much as the evidence
Two people witness the same striking coincidence. One already assigns a high prior probability to hidden causal or paranormal connections; the other assigns a very low one. Bayesian updating explains why they can reasonably reach different conclusions from identical evidence: the same evidence combines with different starting priors to produce different final beliefs, and neither person is behaving irrationally purely by disagreeing — the disagreement traces back to a difference in prior belief, which the new evidence updates but does not fully overwrite in either direction.
This is easiest to see clearly when a conditional-probability problem is framed as natural frequencies rather than abstract percentages: given a low base rate for a rare condition, even fairly reliable evidence for it still leaves the (Gigerenzer & Hoffrage, 1995) posterior probability far from certainty, because the prior rarity of the condition continues to weigh heavily in the final answer. Conditional probability works through this natural-frequency framing in detail.
Common misconception
“That’s such strong evidence, it should completely settle the question” skips past the role of the prior. Even genuinely strong evidence updates a belief in proportion to how much more likely that evidence is under one hypothesis than its competitors — it rarely, on its own, is strong enough to overcome an extremely low prior instantly, which is part of the logic behind how science tests unusual claims requiring more evidence for more extraordinary claims.
Limits and open questions
Bayesian updating is only as good as the numbers fed into it: a prior that is little more than a guess, or a likelihood estimate for the evidence under each hypothesis that is poorly specified, produces a “posterior” number with the same false precision problem as any other made-up probability. Its main practical value for everyday reasoning about coincidences is often the qualitative discipline it enforces — update in proportion to the evidence, and take the prior seriously — rather than a literal numerical calculation.
Related
- Conditional probability covers the natural-frequency framing that makes Bayesian reasoning easier to do correctly.
- Base rates covers the prior information that Bayesian updating starts from before any new evidence arrives.
Key takeaways
- Bayesian updating revises belief in proportion to how much more likely new evidence is under one explanation than its competitors — not to instant certainty in either direction.
- Two people can reasonably reach different conclusions from identical evidence if they started with different prior beliefs; the disagreement is about the prior, not necessarily about the new evidence.
- The historical origin of this idea traces to a paper on updating probability estimates, communicated posthumously for Thomas Bayes in 1763 — one of the oldest continuously relevant ideas in statistics.