--- title: "Bayesian Reasoning" description: "Bayesian reasoning updates the probability of a belief as evidence arrives, in proportion to how strongly that evidence distinguishes between competing explanations. It starts from a prior — how likely the claim was beforehand — and adjusts it, rather than treating each new piece of evidence as t..." last_updated: "2026-08-02" --- # Bayesian Reasoning Bayesian reasoning updates the probability of a belief as evidence arrives, in proportion to how strongly that evidence distinguishes between competing explanations. It starts from a prior — how likely the claim was beforehand — and adjusts it, rather than treating each new piece of evidence as though nothing were known already. Updating the probability assigned to a belief as new evidence arrives, in proportion to how strongly that evidence distinguishes between competing hypotheses. *Example: You think a coworker is upset; then you learn they're always quiet on Mondays, and you dial your confidence back accordingly.* ## What it looks like **In a diagnosis** > The test came back positive, so let's start treatment. *Why it fails:* A Bayesian reading asks how common the condition is and how often the test errs, then combines both. A positive result on a rare condition often leaves the probability low enough to warrant confirmation first. **In a product decision** > One user complained about the new flow, so we should roll it back. *Why it fails:* A single complaint barely shifts a prior built on thousands of successful sessions. The question is how much more likely that complaint is under 'the flow is broken' than under 'the flow is fine'. **In an argument about a rare event** > The witness is certain, so it must have happened. *Why it fails:* Certainty is a fact about the witness, not the event. Extraordinary claims start from a low prior, so testimony must be much more likely if true than if mistaken to move the conclusion far. ## Practice ### Question 1 What are the three ingredients of a Bayesian update? A. The hypothesis, the conclusion, and the sample size B. The prior probability, how likely the evidence is if the hypothesis is true, and how likely that same evidence is if it is false C. The base rate, the confidence interval, and the p-value D. The observation, the theory, and the expert opinion **Answer: B.** The third ingredient is the one people forget: evidence only supports a hypothesis to the extent it is *more* expected under that hypothesis than under the alternatives. Evidence equally likely either way carries no information, however dramatic it seems. ### Question 2 Why does evidence that fits your hypothesis sometimes provide almost no support for it? A. Because fitting evidence is always weak evidence B. Because if the evidence is equally likely under competing hypotheses, it fails to distinguish between them and moves the probability barely at all C. Because confirming evidence is a form of confirmation bias D. Because only disconfirming evidence counts **Answer: B.** Diagnosticity is what matters, not fit. A horoscope description fits you and also fits nearly everyone, so it distinguishes nothing. The useful question is not 'does this match my theory?' but 'would I expect to see this if my theory were wrong?' ### Question 3 You believe a coin is fair. It lands heads five times running. What does Bayesian reasoning say? A. The coin is definitely biased B. Shift somewhat toward bias, but the size of the shift depends on how strongly you believed it was fair beforehand — five heads has a 1-in-32 chance even for a fair coin C. Nothing changes; the coin was assumed fair D. The next flip is more likely to be tails **Answer: B.** The evidence is real but modest, and how far it moves you depends on your prior. A coin from your own pocket warrants a small shift; one handed to you by someone offering a bet warrants a larger one — same evidence, different starting point. Option D is the gambler's fallacy. ### Question 4 What does 'extraordinary claims require extraordinary evidence' mean in Bayesian terms? A. That unusual claims should simply be rejected B. That a very low prior needs correspondingly strong evidence — evidence far more likely if true than if false — to reach a high posterior probability C. That only scientists can evaluate unusual claims D. That extraordinary claims are usually false **Answer: B.** The slogan is a statement about arithmetic. Starting from a very low prior, only highly diagnostic evidence moves the probability substantially. It is not a licence to dismiss unusual claims, which is what option A makes it — it specifies what would be needed to accept one. ### Question 5 What most distinguishes Bayesian updating from ordinary belief revision? A. It requires a calculator B. It makes the prior explicit and asks how diagnostic the evidence is, rather than letting the most recent or most vivid information dominate C. It only applies to numerical questions D. It guarantees the correct answer **Answer: B.** The discipline is in naming the prior and asking about diagnosticity — both of which ordinary reasoning skips. You do not need to compute anything to benefit; asking 'how likely was this beforehand, and would I expect this evidence if I were wrong?' captures most of the value. ## Frequently asked ### Do I need to do maths to reason in a Bayesian way? No. Most of the value comes from two habits: stating how likely the claim was before the evidence arrived, and asking whether you would expect that same evidence if the claim were false. Formal calculation helps in structured settings but is not required for the reasoning discipline. ### What is a prior? The probability you assign to a hypothesis before considering the new evidence. It often comes from a base rate — how common the thing is in general — but can also come from prior experience or background knowledge. Choosing priors is where Bayesian reasoning is most contested. ### What is the difference between Bayesian reasoning and base-rate neglect? Base-rate neglect is the error; Bayesian reasoning is the corrective. Neglect drops the prior and reads the evidence alone. Bayesian updating keeps the prior in the calculation and adjusts it by how much the evidence distinguishes between the competing possibilities. ## Related - [Base Rate](https://www.criticalthinkingexercise.org/pages/glossary/base-rate.md) - [Falsifiability](https://www.criticalthinkingexercise.org/pages/glossary/falsifiability.md) - [Epistemic Humility](https://www.criticalthinkingexercise.org/pages/glossary/epistemic-humility.md) - [Bayesian Reasoning](https://www.criticalthinkingexercise.org/pages/exercises/bayesian-reasoning.md) - [Basic Probability Intuition](https://www.criticalthinkingexercise.org/pages/exercises/probability-intuition.md) - [Lesson: core-skills/scientific-thinking](https://www.criticalthinkingexercise.org/pages/learn/core-skills/scientific-thinking.md) ## Sitemap See the full [sitemap](https://www.criticalthinkingexercise.org/sitemap.md) for all pages.