Why most people never learn Bayes' theorem

Most people don't fail to learn Bayes' theorem because the math is hard. The math is one division. They fail to learn it because their intuition, every single time, supplies the wrong answer with overwhelming confidence, and the theorem — the actual formula — feels like it must be doing something exotic to land on the right answer. So they memorize the formula, fail the next intuition test it gives them, and conclude that they don't really understand it.

This isn't a math problem. It's a self-defense problem against one specific cognitive bug your brain ships with: the base-rate fallacy. Bayes' theorem is the formal patch.

The intimidating formula, in plain English

The formula you've seen:

P(H | E) = P(E | H) × P(H) / P(E)

Translated:

The probability that a hypothesis is true, given some new evidence, equals: how often the evidence shows up when the hypothesis is true, times how likely the hypothesis was to begin with, divided by how often the evidence shows up at all.

The denominator — "how often the evidence shows up at all" — is where everyone slips. It includes the cases where the evidence shows up even though the hypothesis is false. Those cases — the false alarms — are the entire reason Bayes matters. Your brain ignores them. The formula doesn't.

The medical test problem

This is the standard worked example because it weaponizes the failure mode directly.

A disease afflicts 1 in 1,000 people. A test for the disease is 99% accurate — meaning 99% of people who have it test positive, and 99% of people who don't have it test negative. You test positive. What's the probability you have the disease?

Most people, including most doctors in the original 1970s studies, answer something like "99%." The actual answer is about 9%.

Here's why. Imagine 10,000 people get tested:

• 10 of them actually have the disease (1 in 1,000). The test correctly flags ~10 of them as positive.
• 9,990 of them don't have the disease. The test falsely flags 1% of them — about 100 people — as positive.

So 110 people test positive. Of those, 10 actually have the disease. 10 / 110 ≈ 9%.

The "99% accurate" test, applied to a rare disease, produces 10 true positives for every 100 false positives. The base rate — the 1-in-1,000 — dominates the math. The accuracy of the test, however impressive in isolation, doesn't matter as much as we feel it does.

This is the entire point. Bayes' theorem is the formal patch for the fact that your intuition ignores base rates.

Why your brain ships with this bug

This is not a personal failing. It's a species-level pattern, and it's been measured for fifty years.

In Kahneman and Tversky's original "cab problem," subjects were told a witness identifies a cab as Blue, the witness is 80% accurate, and the city's cabs are 85% Green and 15% Blue. People consistently say there's roughly an 80% chance the cab was Blue. The correct answer — using Bayes — is about 41%. The witness's testimony shifts the probability, but the base rate (Green cabs vastly outnumber Blue) keeps the answer below 50%.

People literally do not update their estimate in the direction of the base rate. They use the witness's accuracy as if it were the answer.

Why? A reasonable evolutionary story: in ancestral environments, when something signals "tiger in the grass," the cost of acting on a false alarm is low (you ran for nothing) and the cost of ignoring a true signal is catastrophic (you got eaten). A brain that responds to strong signals without weighting them by base rate is a brain that survives. The price you pay, ten thousand years later, is that you suck at interpreting medical test results.

The bug isn't going away. The mitigation is the model.

The mental model: think in counts, not percentages

The single most useful technique for Bayesian reasoning is convert every percentage to a count, then count.

Reread the medical test problem above. Notice that I never used the formula — I converted "1 in 1,000" and "99% accurate" into "10 sick people, 100 false positives, 110 total positives." Counts collapse the algebra into arithmetic. They make the base rate visible: there are 9,990 healthy people and only 10 sick ones, so even a small false-positive rate creates a flood of false positives.

This is the move that distinguishes people who can use Bayes from people who can only state Bayes. Counts beat percentages every time.

Try it on this one before reading the answer:

A particular AI-detector tool flags 90% of AI-written essays as AI-written, and incorrectly flags 5% of human-written essays as AI-written. In a class of 100 students, 10 used AI. If a randomly picked essay is flagged, what's the probability it's actually AI-written?

Counts:

• 10 AI essays. 90% flagged → 9 true positives.
• 90 human essays. 5% flagged → 4.5 false positives.
• Total flagged: 13.5. True positives: 9.
• P(AI | flagged) = 9 / 13.5 ≈ 67%.

A "90% accurate" detector is right two-thirds of the time on this dataset. If the AI-use rate were 1% instead of 10%, the same detector would be right about 15% of the time. The base rate eats the accuracy.

The base-rate fallacy in the wild

Once you've internalized this, you'll see it everywhere. Some of the most consequential examples:

Airport security. A "99.9% accurate" terrorist detection system applied to a million travelers, of whom maybe one is a terrorist, produces ~1,000 false positives per real positive. The mathematics of rare events is brutal.

Pre-employment drug screening. A drug test with 95% specificity (5% false-positive rate) applied to a workforce with 2% actual drug use generates more false positives than true positives. Companies that fire on the first positive without a confirmatory test will, mathematically, fire mostly innocent people.

Cancer screenings in low-risk populations. This is why screening guidelines distinguish "average risk" from "high risk" populations. The same test, in a population with a lower base rate, produces a higher false-positive ratio. The test didn't change; the math did.

Tech interview signals. If you reject any candidate who fails a particular take-home, and 5% of the people who would succeed on the job happen to bomb take-homes, you're trading away qualified candidates at a rate the rest of your funnel can't see — because rejected candidates don't show up in your data.

In every case the pattern is identical: a signal that feels informative because it's "accurate" turns out to be much weaker once you compute its conditional probability against the base rate. Bayes is the formula that surfaces the weakness.

The thing the formula is doing

Now you can read the formula without flinching:

P(H | E) = P(E | H) × P(H) / P(E)

Translate it directly into the counting move:

• P(H) — the base rate. The number of people who actually have the disease before anyone tested anything.
• P(E | H) — the true-positive rate. Of the people who have it, what fraction test positive.
• P(E) — the total fraction of positives, summed across both groups. This is the denominator that does the work, because it accounts for false positives.

Counting style, the same formula reads: (actual positives among the sick) / (positives among everyone) = (true positives) / (true positives + false positives). That's the form that drops out when you sketch the four-cell table.

When you see the formula in a textbook, do the table. When you see the formula in a paper, do the table. When you see a startup pitching a "99% accurate" anything, especially do the table.

The most important habit to build

Adopt one habit and you will have learned Bayes' theorem in the way that matters: always ask for the base rate before reacting to a signal.

• "The model has 95% accuracy" → on what dataset, with what class balance?
• "The test came back positive" → what fraction of people who get tested actually have it?
• "I had three bad experiences in a row" → out of how many total?
• "The competitor's feature got a lot of complaints" → out of how many users?

Without the base rate, you're answering a different question than the one in front of you. Once you ask for it, you can do the four-cell table in your head — true positives, false positives, true negatives, false negatives — and the real informativeness of the signal drops out. Sometimes that's a relief (the test is more decisive than it sounded). More often, it's deflation (the signal is much weaker than the headline). Either way, you'll be operating on the actual evidence instead of the felt-sense of the evidence.

That instinct is the entire learnable skill. The formula is its proof of correctness. Confuse the proof for the skill, and you'll spend years memorizing without ever using it. Lead with the counts, and the formula becomes obvious — just a way of saying out loud what your table already told you.

If you want to keep going, the natural neighbors are probability vs statistics (Bayes lives at the edge of both), p-values (a frequentist tool with its own base-rate traps), expected value (the other primary tool for reasoning under uncertainty), and the central limit theorem (which tells you when noisy individual signals start to behave predictably in aggregate).