Representativeness Heuristic
Overview
The representativeness heuristic is judging probability by how closely something resembles a prototype — overriding actual base rates. Named by Tversky & Kahneman (1972); produces three systematic errors: base rate neglect, the conjunction fallacy (A-and-B feels more likely than A), and insensitivity to sample size.
Composes with bayesian-reasoning (restores the prior), survivorship-bias (failures are invisible in the prototype), confirmation-bias, and anchoring.
When to Use
- Evaluating candidates, screening investments, or persona-based product decisions
- Any probability judgment where a vivid profile or narrative is present
- When "she/he/it looks like X" drives a decision without a stated base rate
- Auditing for the conjunction fallacy (more specific = seemingly more likely)
- Judging AI startups/valuations/capex by resemblance to a prototype ("the next OpenAI/Stripe," AI capex "obviously pays off," AI adoption curves read as durable revenue)
Not when: judgment is purely quantitative; base rate and profile align and Bayesian updating has been done explicitly.
Coaching Novices (Adaptive Front Door)
- Engine mode: user has a concrete judgment → audit for base rate neglect and conjunction fallacy immediately.
- Coach mode: user is unfamiliar → guide step by step.
In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.
- One-line: we judge probability by similarity to a prototype, which feels like reasoning but ignores base rates.
- Check fit: is a vivid profile driving the judgment? Is a base rate available but being ignored?
- Elicit their real case: what is the judgment, what profile is driving it, what is the actual base rate?
> [WAIT — do not advance until user responds] - Run The Process one step at a time: state base rate → apply Bayesian update → decompose any conjunction.
> [WAIT — do not advance until user responds] - Close by naming the calibrated probability that replaced the prototype-matching estimate.
> [WAIT — do not advance until user responds]
The Process
Step 1 — Identify judgment + profile: What probability estimate is being made? What profile information is driving it?
Step 2 — Reference class + base rate: What is the base rate of this outcome in the relevant reference class? Was it stated before the profile assessment? (If N while a profile is present, representativeness is active.)
Step 3 — Conjunction test: Does the judgment involve multiple attributes (X and Y and Z)? P(A AND B) ≤ P(A). If the conjunction was rated more probable than a component, the conjunction fallacy is active.
Step 4 — Bayesian correction: Prior (base rate) × Likelihood ratio → Posterior. Never start from the profile; always start from the base rate.
Step 5 — Prototype audit: Was the prototype built from survivorship-biased examples? What did failures that matched this prototype look like?
Step 6 — Calibrated output:
Base rate: [X%] | Profile adjustment: [direction + magnitude] | Corrected estimate: [Y%]
Conjunction (if applicable): P(A AND B) ≤ [Z%]
Recommendation: proceed / do not proceed / gather more data
→ Method in Action: Tversky & Kahneman 1972 + 1983 — The Linda Problem · The Hot Hand in Basketball
→ 2026 lens: "The Next OpenAI" — Representativeness in AI-Startup Investing (2023–2026)
Pack: Representativeness Across Domains
| Domain | Prototype trigger | Base rate ignored | Correction |
|---|---|---|---|
| VC screening | Founder matches "winner" archetype | 75–90% of startups fail | State failure base rate before profile |
| Hiring | Candidate looks like top performers | 50–70% of hires underperform | Structured assessment replacing holistic match |
| Customer persona | "Target customer = power user" | Actual distribution is wide | Segment by behavior data, not persona |
| Product launch | "This looks like Slack / Airbnb" | Analogous success rate < 10% | State success rate for the category explicitly |
Applying It Well
- State the base rate before assessing the profile — never after
- Treat profile as evidence that updates a prior, not as the primary input
- Audit every prototype for survivorship bias: what did the failures look like?
- When a conjunction is involved, decompose it — each added attribute reduces joint probability
→ Primary sources: references/sources.md
Common Rationalizations
[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.
| Rationalization | Reality |
|---|---|
| [D] "He just has the DNA of a founder — I know one when I see one." | Representativeness at maximum activation. What is the base rate for "looks like a winner" actually winning? |
| [D] "This is exactly the profile of a successful hire." | Profile built from visible successes. How many with this profile failed? Never checked. |
| [D] "She's a bank teller AND deeply committed to the cause." | Conjunction stated. P(A AND B) ≤ P(A). Narrative fit is not probability. |
| [D] "Every great company looked like this in the early days." | Survivorship bias generating the prototype. How many that looked like this failed? |
| [D] "I've seen this pattern before — it always plays out the same way." | How many times did a similar pattern occur and play out differently? |
| [D] "The more I learn about her, the more I'm convinced." | Added detail increases coherence, not probability. Each condition decreases joint probability. |
| [D] "I trust my gut — I've been doing this for 20 years." | Accumulates calibration error, not correction, without structured feedback loops. |
| → Add [O] entries here after each real use — paste the actual failure pattern | What went wrong and why |
Red Flags
- Probability judgment driven by vivid narrative with no explicit base rate
- Conjunction of positive attributes treated as more likely than any single attribute
- Prototype drawn entirely from visible successes; failure base rates never checked
- "Pattern recognition" cited without stating the pattern's historical hit rate
Verification
- [ ] Reference class defined and base rate stated before profile assessment
- [ ] Conjunction decomposed; no conjunction rated more probable than its least probable component
- [ ] Prototype audited for survivorship bias
- [ ] Bayesian update applied: prior × likelihood → posterior
- [ ] Final estimate anchored to base rate, not profile similarity
Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/representativeness-heuristic · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.
Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/representativeness-heuristic.json
