Understanding Bayesian Updating

Bayesian updating means combining what you believed before with new evidence, then changing your confidence by only as much as that evidence warrants. The catch is that priors can also distort how you interpret the very evidence meant to correct them.

Two candidates take the same aptitude test. The high scorer looks strongest—until you learn they attended a poorly regarded local college, while the low scorer graduated near the top of IIT-B. Suddenly the good score suggests cheating and the bad score suggests an off day. The evidence stayed fixed; the story you told about it reversed.

R1

Your starting belief sets the weight of the evidence

A judgment combines a prior—your expectation before seeing the current result—with the new evidence. That prior is not automatically irrational: it may encode accumulated experience. New observations should move the belief, and the revised belief then becomes the starting point for the next observation. But updating requires more than collecting data. You must notice how the prior changes the meaning you assign to it. If prestige makes a poor result look accidental and a strong result look suspicious, the prior is no longer merely being updated by evidence; it is screening the evidence first. This is where Believing Is Seeing can masquerade as Bayesian reasoning.

E1 R1

Where it shows up

The developer you did not expect

Seeing a well-dressed woman wearing makeup at a software company triggered the assumption that she could not be a developer. Discovering otherwise supplies exactly the kind of observation that should weaken that stereotype for future judgments.

E1

The aptitude test changes meaning

College reputation did not add evidence about what happened during the test; it altered the explanation attached to each score. The example exposes the crucial difference between using a prior to weight evidence and using it to explain inconvenient evidence away.

E1 R1

Updating fails when the prior edits the evidence

A prior can legitimately affect confidence, but it cannot rescue every preferred belief. When every contrary result becomes cheating, bad luck, or an exception, no possible observation can produce an update. At that point the belief is insulated, not Bayesian.

E1 R1

Record the verdict before the biography

The next time you assess a candidate, proposal, or performance, write down what the observable result suggests before revealing reputation or background. Then reveal that context and note exactly what changed, why it changed, and whether the new explanation is supported or merely convenient.

R1

Episodes that teach this