Two decades of popular behavioural science have trained us to look for bias — the systematic tilt that sends judgements consistently off target. Kahneman, working here with strategy scholar Olivier Sibony and legal theorist Cass Sunstein, argues that this focus has left a second error largely unexamined and often larger. Where bias is a shot that misses in the same direction every time, noise is scatter: professionals who should agree reaching wildly different conclusions on identical cases. Judges handing down different sentences for the same crime, underwriters quoting premiums that differ by a factor most insurers would find alarming, doctors reading the same scan and disagreeing.
The core argument
The book’s organising insight is statistical and blunt: noise and bias contribute independently to overall error, and reducing either improves accuracy. Because noise is invisible in any single decision — you cannot tell a judgement is noisy by looking at it, only by comparing it with others made on the same facts — organisations rarely measure it and therefore assume it is small. The authors’ remedy is the noise audit: give a group of professionals identical cases and look at the spread. Wherever this has been done, the spread has been far wider than the practitioners themselves predicted.
They then decompose the problem. Level noise is the stable difference in severity or generosity between individuals: some graders are simply harsher. Pattern noise reflects each judge’s idiosyncratic reactions to particular kinds of case — the largest component in most systems studied. Occasion noise is the variability within one person across time, sensitive to mood, fatigue, hunger and irrelevant context. All three are error, and none is visible from the inside.
Key ideas
- System noise. Unwanted variability in judgements that should be interchangeable; a matter of luck-of-the-draw in which official you happen to get.
- The noise audit. A concrete diagnostic — same cases, many judges, measure the spread — that turns an invisible problem into a number an organisation can act on.
- Simple rules and models beat experts. Even crude linear models, and sometimes random-weight models, outperform the clinicians and forecasters they were built from, chiefly because they are perfectly consistent.
- The illusion of agreement. Professionals believe their colleagues would reach broadly the same conclusion because they almost never see the counter-evidence, which requires deliberate comparison to surface.
- Decision hygiene. Practices that reduce noise without knowing which error they are preventing: structuring judgements into independent components, aggregating multiple independent estimates, using relative rather than absolute scales, and delaying holistic intuition until the evidence is assembled.
- Independence before discussion. Groups amplify noise through cascades and polarisation, so opinions should be collected before people hear each other.
Who it’s for
This is a book for people who run systems rather than people looking to improve their own thinking: HR leads, clinicians, regulators, anyone designing a process where different humans must reach comparable decisions. The recommendations are unusually actionable for the genre, and the reframing — that consistency is itself a form of accuracy — genuinely changes how you look at institutional decision-making.
The fair criticism is length and repetition. The central argument is grasped within the first hundred pages, and the remaining several hundred elaborate, quantify and restate it; the statistical interludes are careful but slow, and a reader comfortable with variance decomposition will find them laboured. There is also a live tension the authors acknowledge but do not fully resolve: aggressive noise reduction means rules and algorithms, which trade discretion, contextual sensitivity and human accountability for consistency — a bargain that is not obviously right in every setting. Worth reading, but skimmable in its middle third.