The Representativeness Heuristic

Tversky and Kahneman defined representativeness as how much something resembles its category "in essential characteristics" and reflects "the salient features of the process that generated it." Your mind keeps a library of prototypes — the typical engineer, the typical fraud case, the typical winning startup — and when a new instance matches one closely, it gets slotted in as likely, fast.

The shortcut is efficient precisely because it skips the step of asking how common that category actually is. Resemblance becomes a stand-in for probability, and the two quietly get treated as the same thing.

In the classic 1983 experiment, subjects read a description of "Linda" written to sound like an active feminist, then ranked the probability of several statements, including "Linda is a bank teller" and "Linda is a bank teller and active in the feminist movement." Most people ranked the second as more probable than the first — even though a conjunction of two things can never be more likely than either thing alone.

The description was representative of a feminist, so the mind reached for the feminist-shaped answer, probability rules notwithstanding. This is the conjunction fallacy, and it is the heuristic's signature error: a good story beats a true number.

Take a disease that affects 1 in 10,000 people and a test that is 99% accurate. Most people who hear "the test came back positive" assume near-certain infection. In fact, because the disease is so rare, false positives from the healthy 9,999 vastly outnumber true positives from the 1 — the realistic chance of actually having the disease is still low.

The test result is representative of infection, so it crowds out the base rate entirely. This is why doctors, underwriters, and recruiters who skip the base rate and go straight to "does this look like the thing" get burned by rare categories that nonetheless produce vivid, convincing-looking cases.

A second, related error: people expect even a small sample to look representative of the whole population — three good quarters "must" mean a skilled manager, five good at-bats "must" mean a hot streak. Kahneman and Tversky called this the "law of small numbers," a tongue-in-cheek nod to the true law of large numbers, which only holds for genuinely large samples.

Small samples are noisy by nature. Treating them as representative is the same underlying error as the Linda problem, applied to a sequence of outcomes instead of a single description.

Representativeness is not a design flaw to be eliminated — it is a fast classifier that is right far more often than it is wrong, which is exactly why evolution and experience kept it around. An emergency physician who has seen ten thousand chest-pain cases can often pattern-match a genuine cardiac event in seconds, and that calibrated intuition outperforms a slow checklist when minutes matter.

The difference between useful expert intuition and a costly bias, per later work by Kahneman and Gary Klein, comes down to two conditions: a reasonably stable environment, and many repetitions with fast, clear feedback. Where both hold, trust the pattern. Where either is missing — a one-off decision, a noisy or shifting environment, delayed or absent feedback — slow down and check the base rate instead.