Representativeness heuristic
People judge probability by resemblance to a stereotype, not by frequency.
When asked how likely it is that something belongs to a category, people assess how much it resembles their mental prototype of that category and use that as the probability. Because resemblance ignores prior odds and sample size: How many observations a test needs before its result means anything., the judgement can violate basic probability, including ranking a specific conjunction as more likely than one of its own components.
How it shows up in software
Personas are representativeness in a slide deck. A team builds a vivid archetype, then treats every decision as a question of whether a feature fits that archetype rather than what share of users it touches. It shows up in classification UI too: a fraud reviewer looks at a transaction that matches the picture of fraud, and a hiring tool ranks candidates by resemblance to past hires.
Using it well
- Attach a size to every persona. If a persona has no share of users next to it, it is a story rather than a segment.
- In any scoring or ranking UI, show the prior alongside the match, so a reviewer sees a similarity signal and a base rate at once.
- Report small-sample findings with the sample size in the same sentence, and say plainly when five interviews cannot settle a question.
- Where you ask users to judge likelihood, use natural frequencies (3 out of 100) rather than percentages or odds.
Where it turns manipulative
- Shipping a model that ranks by resemblance to historical outcomes and calling it merit. This is how hiring and lending tools reproduce the bias in their training data.
- Marketing a product with a customer archetype so specific that people outside it assume the product is not for them, when it is.
- Using a representative-sounding case study to imply a typical result without stating the distribution.
Where you have seen it
LinkedIn Recruiter
Candidate search ranks profiles by similarity to a described role and to previously contacted profiles.
Gmail
Spam classification surfaces a reason banner on filtered mail rather than only the verdict.
Spotify
Recommendation rows are labelled by the similarity basis, such as artists related to something already played.
What the research says
- Tversky and Kahneman, 1983Well evidenced
Given a description of Linda as a philosophy graduate concerned with social justice, most participants rated 'bank teller and active in the feminist movement' as more probable than 'bank teller', which cannot be true.
- Tversky and Kahneman, 1971Well evidenced
Researchers systematically expected small samples to show the same properties as large ones, a belief the authors called the law of small numbers.
- Hertwig and Gigerenzer, 1999Contested
Rewording the Linda problem to ask for frequencies out of 100 rather than probability sharply reduced conjunction errors, supporting a reading where part of the effect is conversational inference about what 'probable' means.
The phenomenon reproduces easily. What it proves is disputed. Read it as evidence that wording and format drive probabilistic reasoning, which is a more useful lesson for product work than 'people are bad at maths'.
Grades are a judgement about the evidence, not about how useful the idea is. Plenty of contested effects are still worth knowing, as long as you do not cite them as settled.