Our hardware wasn't built for it

The human brain handles many things effortlessly — recognising faces, navigating space, throwing a ball at a moving target. These are problems evolution shaped us for over millions of years.

Probability is not one of those problems. Our ancestors didn't need to compute "what's the chance of a tiger over a 10-year horizon given a base rate of one per village per decade." They needed to react fast to the tiger in front of them. When you ask a modern person to reason about base rates, conditional probabilities, large samples, or randomness over long horizons, the answer often comes back wildly wrong — and the wrongness is predictable.

This is the orientation article. The deeper pieces in this cluster cover the birthday paradox, regression to the mean, Bayes' theorem, and the gambler's fallacy. This one introduces the patterns.

Pattern 1 — Ignoring base rates

A medical test for a disease is 99% accurate. You test positive. What's the chance you actually have the disease?

Almost everyone says "99%" or some number close to it. The actual answer depends on how common the disease is in the first place — the base rate — and is often shockingly lower.

Example: a disease affects 1 in 10,000 people. The test correctly identifies sick people 99% of the time, and correctly clears healthy people 99% of the time. You test positive.

Of every 10,000 people:

  • 1 actually has the disease; the test catches them: 1 true positive.
  • 9,999 don't have the disease; the test (mistakenly, 1% of the time) flags about 100 of them: 100 false positives.

You're 1 of 101 people who tested positive. Your chance of actually having the disease is about 1 in 101 — less than 1%, not 99%.

This is the most consequential probability error in modern life. It drives unnecessary medical anxiety, miscarriages of justice (the "prosecutor's fallacy"), and bad screening policy. The fix: always ask the base rate first.

Pattern 2 — Misreading independence

If you flip a coin and it comes up heads five times in a row, what's the chance the next flip is heads?

Half. The coin doesn't care what it did before. Flips are independent.

But it doesn't feel like half. It feels like tails is "due." This is the gambler's fallacy, and casinos exploit it ruthlessly. Roulette wheels keep electronic scoreboards showing the last several spins specifically because losers convinced that red is "due" after a long black streak will keep betting on red.

The flip side — the hot hand fallacy — is believing that recent winners are more likely to keep winning. Sometimes this is real (in skill-based games where momentum matters), but in pure-chance contexts it's not.

Pattern 3 — Mis-sampling

If you survey people about car safety while standing outside an ER, you'll get different answers than if you survey them outside a tennis club. Obvious in the extreme case. Less obvious in subtle ones:

  • Surveying engagement on a social media platform by polling its active users (those who left aren't there to answer).
  • Estimating WW2 bomber vulnerability from where damage was concentrated on returning planes (you're sampling planes that survived; the truly vulnerable spots are where damaged planes didn't make it back — survivorship bias).
  • Concluding a startup strategy works because the founders who used it are wealthy (you're not seeing the much larger group who tried it and failed).

The samples that come easily to us are almost never random. This is selection bias, and it's everywhere.

Pattern 4 — Linearizing the nonlinear

A common pattern: a process where every step has a small probability of failure, but enough steps that the cumulative probability of any failure is large.

If a piece of medical equipment has 1,000 components and each has a 99.9% reliability rate, the system reliability isn't 99.9%. It's roughly 0.999^1000 ≈ 37%. Engineers know this in their bones; ordinary intuition doesn't.

Similarly: a treatment that's 90% effective sounds great. But if the underlying outcome (without treatment) happens 10% of the time, the absolute improvement from treatment is only 1 percentage point. Relative risk reduction sounds bigger than absolute risk reduction, and the news will pick whichever sounds more dramatic.

Pattern 5 — Confusing correlation and causation

When two variables move together, three things are possible: A causes B, B causes A, or both have a common cause C. Without an experiment, you can't tell which.

This is why observational studies, even big ones, can be misleading. Coffee drinkers might live longer, but that doesn't mean coffee causes longevity — maybe people who can afford coffee are wealthier, and wealth causes longevity. Maybe people who feel good are more likely to drink coffee and also more likely to live long, so an underlying health factor causes both.

Randomized trials force the issue: by randomly assigning the treatment, you ensure there's no hidden common cause. This is why randomized trials are the gold standard for causal claims and observational studies, however large, never quite are.

Pattern 6 — Underestimating coincidences

If you imagine "a coincidence happening to someone, somewhere, in a year," you'll find such coincidences are guaranteed. With billions of people having millions of daily experiences, even one-in-a-million events happen thousands of times daily.

People who win the lottery twice exist, despite the apparent absurdity of "winning the lottery twice." There are millions of lottery players, and over time the probabilities work out: if a state has 50 million tickets sold over decades, getting a few double winners is expected, not surprising.

This is why "what are the odds?" is usually the wrong question. The right question is "out of all the things that could be a coincidence, what fraction actually become one?" And the answer is "more than you'd think."

If you'd like to drill these intuitions with quizzes, NerdSip can generate a personalized 5-minute course on probability illusions.

Pattern 7 — Anchoring

Whatever number comes first in your head shapes the next number you'll produce. A judge primed with a high number gives longer sentences. A salesperson who starts at $50,000 will, on average, get a higher final price than one who starts at $30,000, even if both eventually negotiate down to a reasonable range. We don't reason from scratch; we adjust from whatever anchor we encountered.

This affects price negotiations, salary discussions, plea bargains, jury awards, charity asks. Once you know about it, you can both deploy it and defend against it — but only if you remember.

Pattern 8 — Overconfidence in narrow ranges

Most people, asked to provide 90% confidence intervals for unknown quantities ("how long is the Amazon river"), produce intervals that contain the true answer about 50% of the time. They're 40 percentage points overconfident. Even experienced experts in their own field do this.

The fix: when stating a 90% confidence interval, deliberately stretch it. If your gut says "between 4,000 and 5,000 km," widen it to "between 3,000 and 7,000 km." You'll be closer to actually 90% accurate. This is the same logic behind professional forecasters using ensemble methods.

The takeaway

Probability mistakes are not random. They cluster into predictable patterns — base rates, independence errors, sampling problems, nonlinear thinking, correlation-vs-causation, coincidence underestimation, anchoring, overconfidence. Knowing the patterns is most of the battle. You won't always catch yourself; you will sometimes. The other articles in this cluster zoom in on three of the cleanest examples: the birthday paradox, regression to the mean, and Bayes' theorem. Even a small amount of training on these tools moves you out of the most common error patterns.