75Noise

Daniel Kahneman, Olivier Sibony and Cass R. Sunstein, Noise: A Flaw in Human Judgment, Little, Brown, 2021

Linked to 20 chapters: supports 8 · contradicts 5 · mechanism 14 · same claim 3 · pairing 3

The book opens at a shooting arcade. Four teams of five friends share a rifle and each member fires one shot. Team A's shots cluster on the bull's-eye. Team B's cluster just as tightly but off to one side; the team is biased, and the consistency of the bias supports a prediction about the next shot and invites an explanation, perhaps a bent gunsight. Team C's shots are scattered around the bull's-eye; the team is noisy, and the scatter supports no prediction and suggests no cause. Team D is both. Then comes the move on which everything else rests. Look at the same targets from behind, with the bull's-eye hidden, and you cannot tell A from B — but you can still see at a glance that C and D are scattered. Noise can be measured while knowing nothing about the right answer, which makes it auditable wherever the right answer is unknown, unknowable, or years away: a diagnosis, a criminal sentence, an insurance premium, a forecast of what a film will take. Kahneman, Sibony and Sunstein argue that error in professional judgment has two components, systematic deviation and random scatter; that organisations which have looked have found the second at least as large as the first; that almost none have looked; and that the two call for different remedies. The sentence they repeat through the book is the shortest statement of it.

wherever there is judgment, there is noise—and more of it than you think

Finding noise

Part I establishes that the scatter is there and that it is large, once in the public sector and once in the private.

The public case is criminal sentencing in the United States. In 1973 Marvin Frankel, a federal judge and a founder of the Lawyers' Committee for Human Rights, published a book attacking the unchecked discretion of judges. His evidence was anecdotal but hard to shake: two men with no criminal record, convicted of cashing counterfeit cheques for $58.40 and $35.20, one sentenced to fifteen years and the other to thirty days. He called the results arbitrary cruelties and asked Congress to require that differences in treatment be justified by tests capable of objective application — in practice, a checklist with some form of numerical grading. In 1974 a study he chaired put fifty judges from different districts in front of identical pre-sentence reports. A heroin dealer drew one to ten years depending on the judge; a bank robber drew five to eighteen; in an extortion case the range ran from twenty years and a $65,000 fine to three years and no fine. In sixteen of twenty cases the judges could not agree on whether to imprison the defendant at all.

Edward Kennedy took the evidence to the Senate and, after nine years of trying, got the Sentencing Reform Act passed in 1984. The US Sentencing Commission derived guidelines from the average sentences in ten thousand past cases, a choice Justice Breyer defended on the frank ground that the commission could agree on nothing better. The guidelines worked: the expected difference between two judges fell from 17%, or 4.9 months, in 1986 and 1987 to 11%, or 3.9 months, by 1993. They were also detested. Three-quarters of federal judges preferred the advisory regime that followed when the Supreme Court struck the mandatory guidelines down in 2005, and 3% preferred the mandatory one. Crystal Yang's study of nearly four hundred thousand defendants found that after 2005 the penalty for drawing a harsh judge doubled, and the gap between sentences given to Black and white defendants convicted of the same crime widened. Frankel died in 2002; three years later the system returned to something close to what he had spent his life attacking.

The private case is an insurance company whose executives doubted that noise was a problem for them. They agreed to a test. The company wrote five realistic case files for underwriters and five for claims adjusters; employees judged two or three each, independently, without being told what was being measured. Asked beforehand how far apart two randomly chosen professionals would be, most executives said 10% or less — a premium of $9,500 against one of $10,500. Put later to 828 senior executives across industries, 10% was again the median answer. The measured median difference between two underwriters was 55%: against $9,500, not $10,500 but $16,700. For claims adjusters it was 43%. An asset management firm that ran its own audit afterwards, asking forty-two experienced investors to value the same stock from the same one-page summary, found 41%. One senior executive at the insurer put the annual cost of underwriting noise, counting business lost to overpricing and losses booked on underpricing, in the hundreds of millions of dollars.

What surprised the authors was not the noise but the reaction. Nobody disputed the audit and nobody defended the result, yet the problem was new to the organisation. Their image for it is a leak in the basement: tolerated not because it was acceptable but because it had gone unnoticed. They name the condition the illusion of agreement. Professionals rarely see each other's work on the same case, so they never discover the disagreement, and organisations are built to explain it away when it does surface.

Part I closes by refusing the obvious escape. A singular decision — whether to launch a product, how to handle a pandemic, whether to hire someone who does not fit the profile — seems to have nothing to vary against. The authors answer that the distinction is one of framing, not of kind: a one-off judgment is still one draw from a cloud of judgments the same person might have made. A single shot from Team C is as noisy as five; you simply cannot see it.

Your mind is a measuring instrument

Part II asks what judgment is, so that error in it can be measured.

A matter of judgment, in the authors' definition, sits between a question of fact and a matter of taste. It is defined by the expectation of bounded disagreement: reasonable people may differ, but not by much, and at most one of them can be right. Nobody audits a claims department for disagreeing about the Beatles and the Rolling Stones. The insurance executives were shocked by 55% precisely because they took their underwriters to be measuring something.

To show what happens inside a measurement, the book hands the reader a case. Michael Gambardi, thirty-seven, two failed start-ups, a fast rise to regional chief operating officer of an insurer; estimate the probability that he will still hold the chief executive's job at a mid-sized financial firm two years after being appointed. Shown the case, 115 MBA students gave estimates ranging from 10 to 95. The authors then take the judgment apart: the reader attended to some cues rather than others without being aware of choosing, integrated them informally into an impression, and translated that impression onto a scale. Each step varies, so the output varies.

Gambardi does not exist, and this is the point of him. A single probability estimate short of nought or a hundred cannot be confirmed by any outcome, and a fictitious case has no outcome at all — yet judging him feels exactly like judging a real candidate. What ends a judgment, the authors argue, is not evidence running out but an internal signal of judgment completion: a self-administered reward, a feeling of coherence between the facts and the conclusion, equally available whether the judgment is verifiable or not and whether it is right or wrong. It is the machinery that produces the confidence described in Thinking, Fast and Slow, and it is why noise can persist for a career unnoticed: everyone involved has had the experience of getting it right.

Chapter 5 supplies the arithmetic. Gauss's method of least squares, published in 1795, scores overall error as the mean of the squared errors, and it is the only scoring rule consistent with the ordinary intuition that the best estimate of a set of measurements is their average. Applied to a set of judgments it decomposes cleanly: overall error is the square of the bias plus the square of the noise. The two are independent and equally weighted, so reducing either by the same amount improves accuracy by the same amount, whatever the size of the other — and an organisation that has found noise should therefore reduce it without waiting to learn what its bias is. The authors make the point with a forecasting department biased high by ten percentage points and noisy by ten. Halve the bias and the whole distribution slides towards the truth. Halve the noise instead and the forecasts look worse, because 98% now err in the same direction rather than 84%; but bias is the distance from the peak of the distribution to the truth, and that has not moved. Overall error has fallen by exactly as much either way. Noise reduction does not create bias, it makes bias impossible to ignore — and it takes a lopsidedness of 84 to 16 for bias to be merely equal to noise.

They are careful about the limits. Squared error suits predictive judgments, which aim at a true value. It does not suit evaluative ones, where errors in the two directions cost different amounts: underestimating an elevator's maximum load is expensive and overestimating it is lethal, and being one minute late for a train is as bad as being five. The rule they draw is to keep the two apart — settle the prediction on accuracy alone, then apply values to it — and it returns at the end of the book as the spine of their recommended procedure.

Chapter 6 splits noise itself, using the most detailed audit available: a study published in 1981 in which 208 federal judges each sentenced the same sixteen hypothetical cases. The average sentence was 7.0 years and the standard deviation within a case was 3.4 years, so two judges drawn at random differed by 3.8 years on the same defendant. Some of that is level noise, the spread in how severe judges are in general; the standard deviation of judges' personal averages was 2.4 years, and it tracks their views on whether sentencing is for deterrence or rehabilitation rather than anything about the case. The rest is pattern noise: a judge lenient overall but hard on white-collar defendants, another close to average but severe when the victim is old. Statisticians call it a judge-by-case interaction. It obeys the same Pythagorean form as the error equation, and in the sentencing data the two components were about equal.

Chapter 7 adds a third component by noting that the same person is not the same instrument twice. Wine judges at a major American competition who tasted the same wines twice scored only 18% of them identically. Software developers asked on two days to estimate the same task differed by 71% on average. Physicians in a study of nearly seven hundred thousand primary care visits prescribed opioids more often late in the day; admissions officers weighed academic credentials more heavily on cloudy days; asylum judges were 19% less likely to grant an application after granting the previous two. Where the first lottery decides which professional hears your case, a second decides which version of that professional is on duty. Occasion noise is the smaller of the two — in every case the authors could measure, people disagreed with themselves less than with each other — and it cannot be drained away by controlling conditions. In a memory experiment run over twenty-three sessions, alertness, sleep, time of day, practice and difficulty together explained 11% of the variation within a subject.

That a person contains a spread of possible answers has a use. Vul and Pashler found that averaging two guesses from the same person beats either — the crowd within, worth about a tenth of a real second opinion immediately and a third after three weeks. Herzog and Hertwig raised it to about half by instructing people to assume their first answer was wrong and work out why.

Chapter 8 shows that the obvious remedy, putting people in a room together, runs the wrong way. In Salganik's music-download experiment, visitors who could see how many others had downloaded each song produced wildly divergent rankings across otherwise identical groups, and inverting the displayed counts made the unpopular songs popular. On a comment site, a single artificial early up-vote made the next viewer 32% more likely to vote the comment up. The authors' own experiments on punitive damages compared real deliberating juries with statistical juries assembled by taking the median of independent judgments. The statistical juries were far less noisy, which is arithmetic; the deliberating juries were noisier than the individuals composing them, and more extreme, with 27% awarding as much as or more than their most severe member. Discussion does not average judgments. It selects an early speaker and amplifies him.

Noise in predictive judgments

Part III takes the one class of judgment with a long research record and asks why rules beat people at it.

Paul Meehl's 1954 monograph reported that simple statistical rules matched or outperformed trained clinicians, and a review of 136 studies published in 2000 settled it: mechanical prediction won in 63, tied in 65, and lost in 8. The authors' interest is in the reason. Lewis Goldberg, working in Eugene, Oregon in the late 1960s, built regression models not of the world but of individual judges — the same cues, the same tool, with the judge's own past judgments as the target. Such a model is not a description of how anyone thinks; it is an as-if approximation, and across 237 studies it tracked its judge at a correlation of about .80. Goldberg's question was how much accuracy would be lost by replacing the judge with the caricature. The answer was none. The caricature was more accurate. In one replication, ninety-eight people predicted the grades of ninety students from ten cues, and the model of the participant beat the participant for every one of the ninety-eight.

The explanation follows from the error equation. A model of you cannot add anything you do not already know, so it can only subtract. It subtracts your subtle rules, which cost you accuracy when those rules are valid. It also subtracts your pattern noise and your occasion noise, which cost you accuracy always. If half the variance in your judgments is noise, removing it lifts a correlation of .50 to .71. The subtlety is rarely worth that much, partly because the conditions that trigger a subtle rule are by definition rare. Martin Yu and Nathan Kuncel pushed the finding to its limit with assessments of 847 executive candidates on seven dimensions: instead of the best model of each judge, they generated ten thousand sets of random weights. In one of three samples 77% of the random formulas beat the human experts; in the other two, all ten thousand did. Mindless consistency, as they call it, outperformed expertise.

Chapter 10 extends the argument to rules and algorithms, and the bail study is its case. Sendhil Mullainathan's team trained a model on 758,027 decisions using only what the judge had seen. Holding the number detained constant, the model's ranking would cut crime by up to 24%; holding crime constant, it would detain 42% fewer people, and 41% fewer people of colour. The same data, run through models of each judge, showed the lenient fifth releasing 83% of defendants against 61% for the severe fifth, with 79% of the system noise in the patterns rather than the levels. The authors are careful with the conclusion: an algorithm trained on past promotions will reproduce the biases in past promotions, and the claim is not that algorithms are fair but that some are both more accurate and fairer than the people they replace. What blocks adoption is not blanket suspicion — offered a choice, people often prefer the algorithm — but that they abandon it the moment it errs. We forgive ourselves mistakes and expect machines to be perfect.

Chapter 11 puts a ceiling above all of it. Asked how often their own assessment of two candidates would match the candidates' later performance, executives typically say 75 to 85%. In personnel selection the measured figure corresponds to a correlation of .28. The gap is not all bias and noise; most of it is what the authors call objective ignorance, the sum of what cannot be known about a future and what could be known but was not gathered. Tetlock's twenty-year study of nearly three hundred political experts found the average one about as accurate as a chimpanzee throwing darts, and his later work with Barbara Mellers on Superforecasting showed the ignorance thickening with the horizon. It is a property of the task, not of the judge, and no procedure removes it. What the authors object to is its denial — the executive who, told that a structured method would raise validity from .28 to .44, asks why bother, having priced the internal signal of certainty into the alternative.

Chapter 12 asks the question the reader has been storing up: if noise and ignorance are this large, why does the world feel predictable? The Fragile Families challenge is the evidence. A Princeton team led by Sara McLanahan and Matthew Salganik gave 160 research teams thousands of variables on nearly five thousand children followed from birth to fifteen. The best model predicted eviction at a correlation of .22 and the child's school grades at .44. That is typical: a century-long review of social psychology found effects averaging around .21, and of 708 studies in the behavioural sciences only 3% reported .50 or more. Correlation does not imply causation, but causation implies correlation, so .22 measures how much of the causal story the field actually has.

The reason this does not feel like ignorance is that understanding runs backwards. Most of life takes place in what the authors name the valley of the normal, where events are neither actively expected nor surprising. When the Jones family is evicted, the mind searches memory for a cause, finds the lay-off and the inflexible landlord, and produces a story in which the outcome was bound to happen — and would have produced an equally satisfying story had they kept the flat. Causal thinking is cheap and continuous; statistical thinking is effortful and has to be taught. Noise has no story, no villain and no explanation, which is why it is not heard.

How noise happens

Part IV is the psychology, and it opens by repairing a confusion the earlier literature created. The heuristics and biases programme studied what people have in common, so a psychological bias is usually assumed to produce a statistical bias — everyone wrong in the same direction. It does, when the bias is shared. When people are prejudiced in different directions, or when the same bias is triggered in some and not others, the identical mechanism produces noise instead. Anchoring is the clearest case: an arbitrary number moves a judgment, and since different judges meet different arbitrary numbers, the shared tendency scatters rather than shifts. Excessive coherence works the same way: claims adjusters receive the facts of a claim in whatever order they arrive and fit the rest to the early impression, and the order is haphazard.

The mechanism the authors put at the centre is matching. Judgment assigns a value on a scale to an impression by finding the point whose intensity feels equal to the intensity of the impression; look at the sky and "very likely to rain" arrives without effort. The trouble begins when the scale is loose, and the punitive damages study is the demonstration. Kahneman, Sunstein and David Schkade gave 899 people ten cases of corporate wrongdoing and asked, in separate groups, for a rating of outrage, a rating of punitive intent, or a dollar award. Across twenty-eight scenarios the average outrage and the average punitive intent correlated at .98: people are not weighing deterrence against retribution, they are asking how angry they are and matching the punishment to the anger. Then the noise. Taking the mean judgment of each case as the just value, system noise was 51% of the variance on the punitive-intent scale, 71% on the outrage scale and 94% on the dollar scale. The dollar figure is not a disagreement about justice. Dollars are a ratio scale with no upper bound, and such a scale has to be tied down by an anchor; each juror silently invents one on the first case and judges consistently against it thereafter. Replace each juror's awards by their rank order and the noise falls from 94% to 49%. Jurors agree on which company deserves worse and disagree wildly on what worse costs. American law then requires a jury to name a figure for the only case it will ever see, and forbids telling it what other juries awarded.

The last two chapters of the part revise the book's own position. Level noise — the harsh judge, the optimistic forecaster — is the component an organisation can see without an audit, because it shows up in each person's long-run averages, and it is the component the authors expected to dominate. It does not. Among the insurance underwriters, differences in average premium accounted for 20% of system noise and the patterns for 80%; among the federal judges, level noise was slightly under half; in the punitive damages study, pattern noise held near 62% across all three scales. The strongest evidence came from models of 173 bail judges applied to 141,833 cases, over 24 million simulated decisions, in which stable pattern noise accounted for 26% of total variance against 7% for level noise. Alexander Todorov's group, using computer-generated faces and sessions a week apart to get two genuinely independent judgments from one person, found the same ordering and named it consistency without consensus.

That conclusion changes what noise is. It is not mainly a matter of some people running hot and others cold, which could be corrected by recalibrating each of them. It is an interaction — this judge with this defendant, this teacher with this pupil, this recruiter with this candidate — and it is stable, personal and inaccessible to the person who has it. Bias is the figure, noise the ground.

Improving judgments

Part V is the practical half, and it begins by conceding what selection can do. Better judges exist: task-specific skill, general mental ability and a cognitive style the literature calls actively open-minded thinking all predict better judgment. But even the best judges have judgment personalities, so selection narrows noise without removing it.

The chapter that names the book's method starts from a limitation of debiasing. Every debiasing technique — correcting an estimate after the fact, nudging the environment, training people out of a bias — targets a particular bias assumed to be present and pointing in a known direction. In most real decisions several biases are at work, some cancelling others, and the direction of the resulting error cannot be predicted. The authors' proposal for bias is a decision observer, someone in the room with a checklist watching for the signs in real time. Their proposal for noise is different in kind. Noise is error with no known direction, so it has to be prevented rather than corrected, and they call the family of preventive procedures decision hygiene. The name is the argument: you wash your hands without knowing which organism you are stopping, the benefit is never visible, and nobody thanks you. A victory over an invisible enemy is an invisible victory.

Five case studies follow. In forensic science, Itiel Dror showed fingerprint examiners pairs of prints they had judged years earlier, this time with contextual information attached; in his first study four of five experts reversed an identification they had previously made. A 2012 study for the FBI gave seventy-two examiners twenty-five pairs they had seen seven months earlier with no biasing information at all, and about one in ten judgments changed. Dror's remedy, linear sequential unmasking, is to give the examiner only the information needed at each step, in order, and to record each conclusion before the next piece arrives; a verifying examiner should not know what the first concluded. In forecasting, the Good Judgment Project scored volunteers by Brier score and found roughly 2% who consistently beat the rest. Three interventions raised accuracy — training in probabilistic reasoning, putting forecasters into teams, and selecting the best — and when Ville Satopää, Tetlock and Mellers decomposed the gains, all three worked mainly by cutting noise, including the training that had been designed to cut bias. In medicine the instrument is the diagnostic guideline. In performance appraisal the news is worse: across 360-degree studies the ratee's actual performance accounts for 20 to 30% of the variance in ratings, leaving the rest as noise, and forced ranking often makes it worse by mandating differences onto teams that may not have any — the flaw is not the ranking but the forcing. In hiring, two interviewers who see the same candidate correlate between .37 and .44, and an audit at Google found its interviews had no relationship to later performance at all. The repair was structured behavioural interviews: fixed questions about past behaviour, one dimension scored at a time against a shared rubric, before any overall impression is allowed.

Chapter 25 assembles the pieces into a general procedure, the mediating assessments protocol. The decision is broken into its component assessments before evidence is gathered. Each is judged against a reference class of comparable cases and expressed comparatively rather than as praise. The assessments go to different people who work apart, so that one cannot contaminate another. At the decision meeting each is taken as a separate item: everyone rates it privately, the spread is displayed anonymously, the disagreement is discussed, and everyone rates it again. Only when every assessment is fixed does the group form a view of the whole. Intuition is delayed, not banned — people who know a formula will decide the outcome game the inputs, and a procedure that forbids judgment will be abandoned.

The six principles the conclusion extracts each answer one of the mechanisms established earlier.

Part VI then argues against its own programme. The right amount of noise is not zero. Eliminating it can cost more than it saves, as it would if five teachers marked every paper. It can offend dignity, and the objection that killed the mandatory sentencing guidelines — that a rule forbids the judge from seeing what is particular about this defendant — is the same objection. Rules can be gamed once people know the formula. Noise sometimes carries information about changing values, and a system with no variation cannot adapt. Algorithms make mistakes no human would make and lose credibility for them even when they prevent more errors than they cause. The authors accept all of this and make a narrower claim: each objection applies to a particular remedy rather than to noise reduction in general, and no organisation can weigh the objections until it has measured what it is tolerating.

Since 2021

The central observation has held and has been absorbed. Noise audits are a recognised exercise, and the distinction between an error's average and its scatter is now standard vocabulary in operations, insurance, medicine and law. The evidence is also more durable than the evidence behind Thinking, Fast and Slow: the noise findings rest on large administrative datasets and on variance decompositions of real professional decisions, not on the small priming experiments that failed to replicate. Kahneman, who had publicly retracted a chapter of the earlier book, chose accordingly. He died in March 2024.

Three lines of criticism are worth the reader's attention. The first is about novelty: statisticians including Andrew Gelman pointed out that the variance of judgments is not a new object and that psychometrics has studied inter-rater reliability for a century, so the contribution is the framing and the audit rather than the concept. The second is about the error equation. Michael Blastland and others noted that the interchangeability of bias and noise holds for squared error and not for the costs organisations actually face, and that where the collective judgment is already biased, reducing noise concentrates the errors on one side of the truth and can make the worst outcomes more likely — a case the authors half-acknowledge in their own elevator example but do not follow through. The third is about the remedy: suppressing variability produces rigidity, and the discretion being audited away is also what lets a system notice when its rules have gone wrong. The chapter on dignity concedes the point without resolving it.

What has not happened is a serious test. The decision hygiene principles are drawn from evidence gathered for other purposes, and the mediating assessments protocol, the one thing the book proposes as its own, is offered as a worked story about a fictional acquisition rather than as a procedure trialled against a control. Its adoption since remains anecdotal and its effect unmeasured — an odd position for a book whose argument is that organisations should stop trusting procedures they have never audited.