belief & evidence · metaphor 6 of 100
Two reasonable people watch the same decade of news and end further apart than they began. Neither is stupid; neither is lying. Bayes' rule explains how honest minds diverge on shared evidence — and when they are doomed to.
We like to imagine evidence as a referee: pile up enough facts and everyone converges, and whoever doesn't is arguing in bad faith. Bayes says something subtler. A fact has no meaning by itself. It gets multiplied into what you already believe, through your model of how the world generates appearances — how likely this headline would be if the accusation were true, and how likely if it weren't.
That single sentence hides two different tragedies. Same fact, different starting beliefs — different conclusions, for a while. Same fact, different models of what evidence means — permanently different worlds. Below, two watchers follow one news stream about one public figure. Give them minds, and run the decade.
The hidden world is nothing but a fixed frequency table over three headlines. Every curve below is Bayes' rule applied exactly, one event at a time — nothing staged. Changing a prior, lens, or world restarts the stream.
The rule
The folk model of persuasion is a pile: facts accumulate, and at some height the pile wins. Bayes replaces the pile with a product. Each event carries a likelihood ratio — how probable this appearance is under one story versus the other — and your odds get multiplied by it. Nothing is deposited; everything is scaled. Which means every fact needs something to land on: a prior, the compressed residue of everything you believed before this morning. The same ×2 that barely dents a skeptic launches a believer toward certainty. The fact was identical. The minds were not, and the fact was never going to be experienced apart from a mind.
And the ratio itself comes from each watcher's generative model — their private theory of what the world would show them if each story were true. Watch the flagship headline: he denies everything. The trusting watcher reads it at ×0.60 — innocent men deny. The cynical watcher reads it at ×2.00 — that is exactly what a guilty man does. One event, honestly processed by both, is evidence for opposite conclusions. Neither has made an arithmetic error.
What to try
The fault line
This is the deep result the instrument keeps demonstrating: disagreement about where you start is temporary, because under a shared model of evidence, the data multiplies both of you by the same ratios, and the stream eventually swamps any finite head start. Prior disagreement is an argument about initial conditions, and evidence is a machine for making initial conditions irrelevant. But disagreement about what evidence would look like compounds. Different likelihoods mean the same stream is multiplied into the two minds with different — sometimes oppositely-signed — weights, forever. The gap doesn't survive the data; the gap is manufactured by the data.
So the fault line between two minds is what they think the world would show them under each hypothesis — whether a denial is what innocence sounds like or what guilt sounds like, whether a clearing panel is exoneration or capture. This is why "just look at the facts" fails between people with different generative models: they are looking at the facts. Looking at the facts is the mechanism of their divergence. The productive argument is one level up — not what happened? but what would we expect to see if each story were true?
The practice
Cromwell's rule first: never probability 0, never probability 1 — Lindley's name for it, after Cromwell's plea, "think it possible you may be mistaken." Certainty is the decision to stop multiplying. The zealot's flat line is arithmetic: no ratio, however extreme, moves a zero. Keeping your log-odds finite is what keeps you reachable.
Second: the question "what would change my mind?" is stating your likelihoods aloud — naming, in advance, the events you'd count against yourself and by how much. Someone who cannot answer it has likelihoods that read every possible event as ×1.0 or better for their side, and such a model doesn't process evidence; it digests it. "Strong opinions, weakly held" is, on this reading, a slogan with exact content: let the prior be as bold as you like, provided |log-odds| stays finite and the ratios stay honest — bold about where you stand, negotiable about what would move you.
The mapping
| Mathematics | Life |
|---|---|
| prior P(H) | Where you start — the residue of your whole history, compressed into a number the next fact will land on. |
| likelihood P(E | H) | Your model of what the world would show you if each story were true — what innocence sounds like, what guilt sounds like. |
| posterior ∝ prior × ratio | The update: one honest multiplication per event. Nothing accumulates; everything is scaled. |
| shared likelihoods → convergence | Disagreements that facts can settle — a chasm of priors, healed by a few dozen honest observations. |
| different likelihoods → divergence | Disagreements that facts deepen — the same news stream, multiplied into two minds with opposite signs. |
| P = 0 or P = 1 | The mind no evidence can reach: zero times anything is zero, forever. |
Where the metaphor tears
Real people are not Bayesian. We update on vividness rather than diagnosticity, on the order in which we heard things, on what our friends concluded first; we double-count repeated tellings of one fact and forget the base rates entirely. The rule is an aspiration, not a description — a plumb line for noticing which way your actual updating leans, not a portrait of it.
Updating happens inside a hypothesis space, and choosing that space is prior to all of it. The instrument admits exactly two stories, guilty or innocent — but the truth might be a third thing neither watcher ever formulated, and the option you never considered gets probability zero by omission. No amount of scrupulous multiplication corrects it. This is the deepest zealotry, and it is invisible from inside the model.
It would be comfortable to stop at "different likelihoods, different worlds, who's to say." But likelihood models make frequency predictions, and frequency predictions can be scored: a watcher whose lens keeps assigning high probability to events that don't come, or low probability to events that do, is measurably miscalibrated. The two watchers are symmetric in the algebra, not before the record. That is where the shrug honestly breaks.