What is Noise: A Flaw in Human Judgment about?

Bias gets most of the attention in discussions of judgment error. Kahneman, Sibony, and Sunstein's central claim in Noise is that a different and equally important source of error has been largely overlooked: noise, which is the variability in judgments that should be identical. Two doctors seeing the same X-ray give different diagnoses. Two judges given identical cases hand down sentences that differ by years. Two underwriters assessing identical insurance applications price them differently. Wherever professional judgment is involved, noise is present — and usually at levels that would be shocking if anyone measured them.

The book distinguishes several types of noise. Level noise is the variation between individuals in their average judgments — one judge is consistently harsher than another. Occasion noise is variation within the same individual across time — the same judge is harsher before lunch than after, harsher on cold days. Pattern noise is variation in how different people respond to the same specific case, independent of their average level. All three contribute to the total noise in a system.

Bias and noise are not the same problem and do not have the same solutions. Bias is a systematic directional error — a tendency to judge too high or too low. Noise is scatter around whatever center a person has. Bias causes errors in the same direction; noise causes errors in random directions. You can cancel noise by averaging over many judges, but averaging does not remove bias. The book argues that interventions designed to address bias — training, feedback, awareness — often do not reduce noise, and vice versa.

The prescriptions center on decision hygiene: structures and processes designed to reduce noise rather than eliminate it entirely. These include using algorithms and structured checklists, sequencing information to reduce anchoring, having judges make initial independent assessments before deliberation, and using scoring rubrics that break judgments into components. The book's most provocative claim is that mechanical rules and algorithms, even imperfect ones, will typically outperform unaided human judgment in most domains because they are noise-free — consistent in a way humans cannot be.