How rational agents revise beliefs when they observe new evidence.
Bayesian updating is the process of revising a belief in light of new evidence. Starting from a prior belief about how likely something is, you observe data, and combine the two using Bayes’ rule to produce an updated posterior belief. It is the formal framework for learning under uncertainty—how an ideal rational agent should think.
Prior: What you believe before seeing any evidence. “I think there is a 10% chance it will rain today.”
Likelihood: How probable the evidence would be if the hypothesis were true. “If it were going to rain, dark clouds would be very likely.”
Posterior: Your updated belief after combining prior and evidence. “Given the dark clouds, I now think there is a 70% chance of rain.”
Bayes’ rule: The mathematical formula that performs this update. It weights the prior by how well the evidence fits each possible explanation.
Learning from experience: Each new observation shifts beliefs gradually. After enough consistent evidence, even a skeptic updates to the correct answer.
Prior sensitivity: With little evidence, the prior dominates. With abundant evidence, the data overwhelms the prior. This explains why experts and novices can disagree initially but converge over time.
Base rates matter: Ignoring the prior leads to the base-rate fallacy. A medical test that is 99% accurate still produces mostly false positives if the disease is extremely rare.
Medical diagnosis: Doctors update disease probabilities as test results, symptoms, and patient history accumulate—each piece of evidence shifts the posterior.
Spam filtering: Email filters start with a prior about whether a message is spam, then update as they observe words, sender reputation, and link patterns.
Criminal investigations: Detectives begin with priors about suspects and update as forensic evidence, alibis, and witness testimony emerge.
Machine learning: Bayesian models update parameter beliefs as training data arrives—producing not just predictions but calibrated uncertainty estimates.