The Representativeness Heuristic — Judging by Resemblance
When similarity hijacks probability — and how to correct the substitution
What is the representativeness heuristic, and its traps?
The representativeness heuristic, identified by Daniel Kahneman and Amos Tversky, is the mental shortcut of judging how probable something is by how closely it resembles a prototype or stereotype — substituting the easy question "how similar is this?" for the hard question "how often does this actually happen?" It is fast and often useful, because resemblance and frequency really do correlate much of the time. But it misfires in recognisable ways: it overrides base rates (guessing a quiet, bookish stranger is a librarian rather than a salesperson, when salespeople vastly outnumber librarians); it produces the conjunction fallacy, where a detailed, story-like description is judged more likely than the broader category that contains it; it treats small samples as though they carried the reliability of large ones; and it makes random sequences look non-random, which is the engine of the gambler’s fallacy. The correction is not to distrust intuition wholesale but to notice the substitution as it happens and deliberately ask for the base rate.
Kahneman and Tversky identified the representativeness heuristic as one of the most consequential shortcuts in human judgment: we assess probability by asking "how much does this look like a member of that category?" rather than "how often does this actually occur?" The heuristic is not always wrong — resemblance often correlates with membership — but it fails systematically when base rates are ignored, when sample sizes are treated as irrelevant, or when a coherent narrative obscures the underlying frequencies. What makes it hard to catch is that the substitution is invisible from the inside: you do not experience yourself skipping the probability question, you simply experience an answer arriving with confidence attached. Detail makes this worse rather than better — the more specific and coherent a story is, the more representative it feels and the more probable it seems, even though every added detail can only narrow the odds. That is why the correction has to be procedural rather than attitudinal: resolving to "be more rational" does nothing, whereas the habit of asking "out of all the cases like this, how many actually turn out that way?" reliably surfaces the number the shortcut skipped. The practices below address the specific failure modes, each with the cognitive mechanism responsible.
Practices
- Look up base rates before forming a resemblance judgment
Before deciding "this looks like X," ask how common X actually is in the relevant population.
- Identify which features are actually diagnostic
Separate features that genuinely differentiate categories from ones that just complete the picture.
- Recognize that random sequences don’t "owe" balance
Random processes have no memory — a run of heads doesn’t make tails more likely.
- Treat small samples with explicit skepticism
A short sequence can look representative without being statistically reliable — adjust confidence for sample size.
- Audit person-judgments for prototype substitution
Check whether a judgment about a person is based on their actual behavior or on resemblance to a type.
- Expect extreme results to regress toward average
Unusually good or bad performance predicts more average performance next time — account for this before giving praise or blame.
- Name the prototype you’re comparing against before deciding
Make the prototype you’re using as a reference explicit — hidden templates bias decisions without scrutiny.
How these practices connect
All seven practices correct the same substitution: representativeness answers 'how much does this resemble a category or template?' in place of 'how probable is this, given the actual base rate and sample size?' Because the resemblance judgment is computed automatically and feels like evidence on its own, the substitution is invisible unless something forces the harder statistical question to the front. Each practice is a different point of intervention on that same swap — before the judgment (base rates, diagnostic features), during a sequence (gambler's fallacy, small samples), on people (stereotyping), on outcomes over time (regression to the mean), or on the hidden comparison itself (naming the prototype).
Are these all just the same bias with different names?
No, but they share one root cause. Every practice here corrects the same underlying substitution — judging probability by resemblance instead of by actual statistics — but each targets a different place that substitution shows up: a single judgment about a person or thing, a sequence of random events, or a trend over time. Check-base-rates-first and prototype-trap-in-decisions both intervene before a judgment is made, while gambler-fallacy-correction and small-sample-skepticism intervene on how a sequence of outcomes is read. They're expressions of one mechanism, not seven unrelated tips.
Why doesn't a coin 'owe' you a tails after a run of heads, but a slumping player is expected to bounce back?
Because the two situations differ in whether the underlying process has memory, and each mistake is representativeness pulling in the opposite direction. A coin flip is genuinely random with no memory, so a long run of heads doesn't make tails more likely — expecting a correction is the gambler's fallacy, because the corrected sequence merely looks more representative of 'randomness.' A slumping player's next performance, by contrast, partly reflects real skill and partly reflects chance, so extreme results tend to regress toward that player's true average — which can look like a rebound but is actually regression to the mean, not a correction being paid off. Telling them apart requires asking whether the process is actually random or whether it mixes skill and luck.
If I already thought about the base rate, am I safe from the heuristic?
Not automatically. Base-rate neglect is about weighting, not awareness — you can consciously note the base rate and still let a vivid, individuating description override it, because the resemblance judgment is computed easily and feels like stronger evidence than a population statistic. The corrective only works if the base rate actually changes your final answer, not just if you mentioned it in passing. This is the same failure pattern named in distinguish-diagnostic-from-decorative-features: noticing a factor exists is not the same as it actually constraining the judgment.
Why does a long, detailed description feel more convincing than a short one, even when it shouldn't?
Because people treat the internal coherence of a description as evidence of category membership, when in fact only diagnostic features — ones that are actually more common in the target category than in alternatives — carry any real discriminating information. A longer description usually just adds decorative features that are consistent with a prototype but don't distinguish it from competing explanations, so the perceived strength of the case goes up while its actual evidential value doesn't. This is the same mechanism as the hidden-prototype problem: the mind is matching the whole picture to a stored template rather than weighing which specific details are diagnostic.
Is stereotyping just base-rate reasoning, or is it a different mistake?
It's related but not identical. Group base rates can be real statistical information, so using them isn't automatically an error — the mistake named in stereotyping-awareness is treating a group-level pattern as a certainty about one individual instead of as a probabilistic starting point that should update with actual evidence about that person. The failure mode compounds when individuating information is sought but then dismissed as 'an exception' whenever it doesn't match the template, which is representativeness reasserting itself after having appeared to be corrected. So the discipline isn't to ignore base rates about groups, it's to keep updating on the specific person in front of you rather than defaulting back to the type.
How many observations do I need before I can trust a pattern?
There's no fixed number, but the practical rule is that the smaller the sample, the more an extreme or patterned result can occur purely by chance, and representativeness makes people judge a 4-observation streak the same way they'd judge a 40-observation one. Small-sample-skepticism means widening your confidence and, critically, actually changing the decision you'd make — not just verbally noting uncertainty while still acting on the point estimate. It's also possible to overcorrect into dismissing every early result; a more calibrated approach updates a reasonable prior gradually rather than either trusting the first few data points or discounting all of them.
If I punished someone after a bad performance and they improved, doesn't that prove the punishment worked?
Not by itself, because extreme performances mathematically tend to regress toward the average whenever both skill and chance are involved, independent of any intervention. Harsh feedback after an unusually bad outcome will typically be followed by improvement anyway, and praise after an unusually good outcome will typically be followed by a step back, simply because outcomes that far from average were partly luck to begin with. The corrective isn't to ignore feedback altogether — real skill changes do happen — it's to confirm a trend across multiple observations before crediting a single intervention, and to expect regression more strongly in domains where chance plays a bigger role than skill.
How do you choose between these?
Start by locating where the judgment is happening: if you're evaluating a single case against a category (a person, a product, a business idea), name the prototype you're implicitly comparing to and then check its base rate and which features are actually diagnostic. If you're reading a sequence of outcomes over time (a streak, a run of results, a performance trend), the relevant checks are different — ask whether the underlying process is truly random (gambler's fallacy territory) or a mix of skill and chance (regression to the mean territory), and separately ask whether the sample is even large enough to mean anything. Stereotyping-awareness applies specifically when the case involves judging a person, since that's where prototype substitution has the most consequence. The situation — single case versus sequence, person versus event — determines which practice is doing the actual work, not any trait of the person applying it.
What these approaches do not settle
None of these practices are trialed interventions with effect-size data; the evidence behind them is either observational (documented failure patterns in judgment research) or mechanistic (a mathematical inevitability, in the case of regression to the mean, that people reliably fail to account for). Several of the corrections can be overcorrected — small-sample-skepticism can turn into dismissing every early result, and gambler-fallacy-correction can be wrongly applied to situations with genuine sequential dependence, like a skilled performer's real hot streak. Naming a prototype or a base rate is also only a first step: it makes the comparison visible, but correctly estimating that base rate, or judging whether a streak is random or skill-based, is a separate skill these practices don't by themselves supply.