A patient walks out of a clinic shaken: they tested positive for a rare disease, and the test is “99% accurate.” In their head the verdict is already in — I have it. Down the hall, a doctor who actually does the arithmetic looks at the same result and thinks: probably a false alarm — let’s run a second test before anyone panics. Same test, same positive result, two completely different beliefs. Only one of them is calibrated, and it’s not the obvious one. With a disease that strikes 1 in 1,000 people, a 99%-accurate test that comes back positive means there’s only about a 9% chance you’re actually sick.
That gap — between reading a positive test as a verdict and reading it as a nudge — is the entire subject of this course. The patient flipped their belief all the way to certain on one piece of evidence. The doctor moved it a little, from “1 in 1,000” up to “about 1 in 11,” and stopped exactly where the math said to stop. That disciplined, partial move has a name: Bayesian updating, and it’s the most important thinking tool you’ll meet on the whole probability path.
The move, in one breath
Here’s the whole idea before we spend five lessons earning it. Bayesian updating means you never build a belief from scratch when new evidence arrives. You start from what you already believed — the prior — then you ask how well the new evidence fits, and you move your belief in proportion to how strong that evidence really is. Surprising, hard-to-fake evidence moves you a lot. Weak, ambiguous, easy-to-fake evidence moves you barely at all. You land on an updated belief — the posterior — and that posterior becomes your prior for the next piece of evidence.
It sounds modest. It is also the single discipline that separates people who track reality from people who get yanked around by whatever they heard last. The stubborn never update; the gullible update all the way every time. The Bayesian updates by the right amount — and “the right amount” turns out to be a number you can actually compute.
Before you read — take a guess
A home test for a rare condition (it affects about 1 in 1,000 people) is 99% accurate. You test positive. Before reading on, what's your gut: roughly how likely is it that you actually have the condition?
Why “accurate” is a trap
Look at what fooled the gut in that question. The test is “99% accurate,” and the mind reads that as “99% chance the result is right.” But accuracy is a statement about the test — how it behaves when you already know who’s sick and who’s healthy. The thing you actually want is the reverse: given a positive result, how likely is the disease? Those two numbers feel like the same thing and are wildly different, and the gap between them is exactly the size of the base rate you forgot.
That’s the trap Bayesian updating exists to defuse. A positive test is evidence, not a verdict. Evidence shifts a belief; it doesn’t replace it. To know how far it shifts you, you need three things, and only one of them is the test’s accuracy:
- The prior — how likely the disease was before the test (here, 1 in 1,000). This is the base rate from the last course, now doing a new job.
- The true-positive rate — how often the test fires when you do have the disease.
- The false-positive rate — how often it fires when you don’t.
Leave out the prior and you get the patient’s panic. Keep it, and you get the doctor’s calm. The rest of this course is just learning to hold all three at once.
The one-sentence version
Bayesian updating = start from your prior (usually the base rate), weigh how well the new evidence fits, and move your belief in proportion to how strong that evidence is — never all the way, never not at all. Evidence is a nudge whose size you can calculate, not a verdict that overwrites what you knew.
Which sentence shows someone updating like a Bayesian, rather than flipping a switch?
Why this is the model, not just “be open-minded”
“Keep an open mind” is advice everyone nods at and nobody can act on, because it doesn’t say how open or to what. Bayesian updating is the operational version. It gives you a specific move — combine your prior with the evidence’s likelihood — a specific quantity to estimate — how much more likely is this evidence if my hypothesis is true than if it’s false? — and a specific failure to expect — people drop the prior and treat the evidence as the whole answer (the base-rate neglect you met under base rates, now with a name for the cure).
And like every good mental model, it’s portable. The same engine that turns a positive medical test into “about 9%, run another test” turns a glowing job interview into “good signal, but most confident-sounding candidates still don’t work out,” turns a scary headline into “alarming, but how often does coverage like this precede the actual disaster?”, and turns a scientific result into “extraordinary claim — does the evidence clear the extraordinary bar?” It isn’t tied to medicine or to math class. It’s a move you can make about any belief, the moment new evidence lands.
A friend insists their new business will 'definitely' succeed because three friends said they'd buy. What's the Bayesian critique?
The map of the course
Five short teaching lessons, then one exam you can’t undo. The route climbs from intuition to the formula and back to judgment:
- Prior, Likelihood, Posterior — the three ingredients in plain language: your starting belief, how well the evidence fits each hypothesis, and the updated belief you get by combining them. The vocabulary the rest of the course runs on.
- The Medical Test — the canonical worked example, done in full with real numbers and a natural-frequency tree that makes the base rate impossible to forget. You’ll drive an interactive posterior calculator and watch a positive test move you only a little.
- Bayes’ Theorem — the actual equation,
P(H|E) = P(E|H)·P(H) / P(E), plus its much friendlier odds form — prior odds × likelihood ratio = posterior odds — which lets you update in your head. - Strength of Evidence — the likelihood ratio as a single number for how diagnostic a clue is; weak vs. strong evidence; why extraordinary claims require extraordinary evidence; and updating again and again as evidence arrives.
- Where Updating Goes Wrong — the pitfalls: dropping the prior, confusing P(evidence | hypothesis) with P(hypothesis | evidence), double-counting correlated evidence, and refusing to update at all. Ends with a whole-course recap.
Then a Final Exam — graded, one question at a time, one-way: once you answer, it locks. No back button, no retries.
How to use this course
One rule does most of the work: say your prior out loud before you look at the evidence. The instant you commit to a starting number — “I’d have said 1 in 1,000” — the update has something to push against, and you can feel whether the evidence is moving you a sensible amount. Skip that step and every new fact feels like the whole story. The exercises are built to make you state a prior and then catch your own over- or under-reaction.
Next up: lesson 1, where we name the three ingredients — prior, likelihood, posterior — and snap them together into the single sentence the entire model rests on.