Annie Duke · 2018

Thinking in Bets

Making Smarter Decisions When You Don't Have All the Facts

Rough AI truth score

75/100

Duke's central distinction is solid: uncertain outcomes do not reveal decision quality by themselves, and outcome and hindsight bias can corrupt learning. Probabilities, records, and premortems help make reasoning auditable. The social prescription is more conditional—truth-seeking groups can improve judgment, but conformity, status, selection, correlated information, and incentives can make groups worse.

Based on three central claims · high confidence

The three claims

01Supported

A decision's quality should be judged from the information and process available at the time, not the realized outcome alone.

Outcome-bias experiments directly show that identical decision processes receive better evaluations when random outcomes are favorable. Normative decision theory also evaluates choices against beliefs, information, probabilities, and payoffs available ex ante rather than information revealed afterward.

Outcome Bias in Decision Evaluation The Hindsight Bias: A Meta-Analysis

02Mostly supported

Explicit probabilities and decision records reduce hindsight distortion and improve learning from uncertain results.

Recording forecasts and assumptions preserves the ex-ante information set, while proper scores and calibration curves separate probability quality from one outcome. Documentation cannot remove hidden assumptions, strategic reporting, ambiguous resolution, or feedback delays, and direct transfer evidence varies by setting.

Good Judgment Project Calibration Analysis A Meta-Analysis of Research on Hindsight Bias

03Mixed

Truth-seeking groups with norms of dissent and accuracy reliably improve individual decisions.

Independent estimates, constructive disagreement, aggregation, and accuracy incentives can expose errors and improve forecasts. Groups can also amplify shared bias, polarization, hierarchy, reputational pressure, and common information; benefits depend on composition, independence, facilitation, scoring, and task structure.

Cross Validation for Structured Expert Judgment Good Judgment Project Calibration Analysis

Other claims worth checking

Premortems can reveal failure modes before commitment.
Separating belief confidence from identity can make updating less threatening.

What this number means. It is an AI-generated first-pass judgment of three central factual or causal claims—not a rating, exhaustive fact-check, or human peer review. Claim credits are 100% for supported, 75% for mostly supported, 50% for mixed, and 25% for weak, then averaged and rounded. Lower confidence means the score should move more as better evidence arrives.

Method three-central-claims/0.1.0 · checked 2026-09-01 · 3/3 selected claims assessed · method and source audit