Noise

Human judgment is damaged not only by bias, which pushes errors in a shared direction, but by noise: unwanted variability among judgments that should be similar. Noise argues that many organizations worry about prejudice, incentives, and predictable distortions while missing a quieter defect: two authorized professionals can see the same kind of case and produce different answers for no defensible reason. The bad outcome is not just an occasional mistake. It is a system that behaves like a lottery.

The book’s central move is to make inconsistency visible. A single sentence, diagnosis, insurance premium, inspection result, job interview score, or performance rating may sound reasonable when considered alone. The flaw appears only when comparable judgments are placed side by side and the spread is larger than the institution can justify. That is why the practical question is not only “What bias caused this decision?” but “How variable is this decision system, and is that variability acceptable?”

Bias has direction; noise has scatter

Bias is average error. A group may be too optimistic in forecasts, too severe in punishment, or too generous in ratings. Noise is variability of error. Decision makers who are supposed to be interchangeable disagree in ways the system did not intend. In settings where a correct answer exists, the authors use the error logic that total error contains both bias squared and noise squared. The managerial implication is simple: a process can reduce bias and still leave a large amount of error if judgments continue to scatter.

This distinction matters because bias is easier to narrate. It has a culprit, a motive, a stereotype, or a known cognitive tendency. Noise is statistical. It becomes visible across cases, not inside one case. If an insurer prices one risk, a court issues one sentence, or a doctor gives one diagnosis, the decision can be defended with reasons. But if equally authorized underwriters, judges, or doctors reach markedly different conclusions from similar or identical evidence, the reasons no longer settle the matter. The pattern itself is evidence of poor quality.

The authors are careful that not all variation is bad. Noise is unwanted variability. A society does not need film reviewers, artists, voters, or creative teams to respond identically. Diversity of taste and perspective can be valuable. Noise becomes a flaw when the institution itself promises like treatment for like cases. Courts, hospitals, insurers, food inspection systems, hiring processes, and employee evaluations usually cannot defend large arbitrary differences as personal style.

The book’s public examples make the point concrete. Judges in the same system may give different sentences for similar crimes. Doctors in the same city may give different diagnoses to identical patients. The point is not merely that professionals are fallible. It is that the person being judged may receive a different outcome depending on which professional happens to be assigned.

A system can be noisy even when its people are competent

The book’s main target is system noise: unwanted variability inside a system that is supposed to speak with one voice. System noise does not require incompetence or bad faith. A noisy system may contain trained, sincere professionals who can each explain their decisions. The defect is that the institution has not created enough structure, calibration, independence, or feedback to make their judgments reliably comparable.

The authors divide system noise into components that help locate the problem. Level noise is variation in decision makers’ average levels. One judge is generally harsher, one interviewer more forgiving, one performance rater more generous. Sibony uses performance reviews to illustrate this: some raters are higher or lower on average, so an employee’s rating can partly reflect the rater’s habitual leniency or severity rather than the employee’s work.

Pattern noise is more subtle. Two decision makers may be equally severe on average while disagreeing about which specific cases deserve severity. One may react strongly to one feature of a case, another to a different feature. Their average level looks similar, but their case-by-case rankings diverge. This is why calibration around general strictness may not be enough. A system can align people on the average and still leave them disagreeing about what matters in particular cases.

Occasion noise adds instability within the same person. The same professional may judge differently depending on timing, sequence, fatigue, mood, or other irrelevant conditions. The book’s public examples include the difference between morning and afternoon, or Monday and Wednesday. The authors treat occasion noise as a contributor to system noise, though differences between decision makers are more central. The practical lesson remains: even a conscientious expert is not a fixed measuring instrument.

This framing prevents a common but unhelpful response. Telling professionals to “be consistent” does not specify what should change. A common form or policy also does not guarantee a common interpretation. If the system allows each person to bring a private scale, private weighting of evidence, and private timing effects, it can look orderly while producing arbitrary spread.

A noise audit measures the spread before arguing about causes

Because noise belongs to a pattern of judgments, it usually cannot be diagnosed from one decision. The book’s practical tool is the noise audit: give multiple interchangeable decision makers the same realistic cases and measure how much their judgments vary. The audit does not always require knowing the correct answer. If the organization expects alignment, the spread among judgments is itself the object of concern.

The insurance-underwriting example is the cleanest demonstration. Kahneman describes executives who expected about 10 percent variation among underwriters pricing the same risks. The observed variation was roughly 55 percent. The surprise matters because leaders did not discover a subtle philosophical disagreement about judgment. They discovered that a system they trusted was far noisier than they thought.

A noise audit changes the politics of improvement. Starting from one controversial case invites argument about details: perhaps the case was unusual, perhaps the decision maker had a reason, perhaps the critic is hindsight-biased. Starting from a distribution of independent judgments on common cases asks a harder question: can the institution defend this much variability? Sometimes it may choose to tolerate some spread because reducing noise has costs or because discretion is genuinely valuable. But without measurement, it is not choosing discretion. It is inheriting randomness.

The audit also shifts attention from blame to design. It may identify extreme judges, but its larger purpose is to reveal a system property. Once the spread is visible, the organization can consider calibration, clearer scales, better subjudgments, aggregation, rules, checklists, or selective automation. The aim is not to humiliate professionals. It is to learn whether the process produces a pattern compatible with its promise.

Decision hygiene delays intuition until evidence is disciplined

The book’s remedy is decision hygiene: preventive practices that reduce noise before anyone knows exactly which error would have occurred. The hygiene metaphor is important. These practices can feel thankless because they prevent invisible mistakes. Their value appears statistically, not dramatically.

The authors do not argue that intuition should be eliminated. Kahneman’s position is that intuition should be delayed. The danger is premature coherence. Once a person or team forms an overall impression, later evidence is interpreted through that story. Early discussion can make this worse by creating anchors, authority effects, social pressure, and shared errors. A group may feel more confident after talking while actually becoming less independent.

The main hygiene practices follow from that mechanism. Break complex judgments into separate assessments. Define the dimensions before the overall story forms. Have people make fact-based assessments independently before discussion. Use clearer scales, and where appropriate prefer relative comparisons over vague absolute labels. Aggregate or compare judgments before allowing the final holistic evaluation.

The Mediating Assessments Protocol is the structured version of this logic for complex decisions:

1. Define the key assessments the decision requires before making the final call.
2. Gather fact-based assessments for each dimension separately.
3. Keep assessors’ judgments independent long enough to avoid early social influence.
4. Review the complete assessment profile before forming an overall evaluation.
5. Let intuition enter at the end, after the evidence has been organized.

The point is not procedure for its own sake. It is to stop teams from deciding early and then retrofitting reasons to a favored conclusion. In hiring, strategy, investment, or performance evaluation, a compelling first story can colonize the evidence. Decision hygiene makes the story wait.

Consistency helps only when the target is worth standardizing

Rules, checklists, structured interviews, formulas, and algorithms can reduce noise because they give the same input the same treatment. When a human process contains large arbitrary variation, a consistent rule may remove one major source of error. This is a direct consequence of the distinction between scatter and direction: if the problem is excessive spread among people, standardization can improve quality.

But consistency is not the same as fairness or truth. A rule can be consistent and consistently wrong. An algorithm can reduce human variability while introducing systematic bias if it is trained on biased data, optimized for the wrong target, or deployed without scrutiny. The book’s argument therefore supports rules and algorithms only with the adjacent qualification that they must be audited for bias and fit.

The same qualification applies to human decision hygiene. More structure, more independent ratings, more calibration, and more auditing all cost time and attention. The authors later emphasized that noise reduction has costs, so the benefits may not always justify the intervention. A low-stakes decision may not deserve an elaborate protocol. A repeated, high-stakes judgment affecting liberty, health, employment, money, or rights is a stronger candidate.

The durable lesson of Noise is that decision quality must be judged by patterns, not only by reasons. A process is not reliable merely because each decision maker can explain a decision, or because the organization has a policy and trained professionals. The relevant test is how much variation reflects real differences in cases and how much reflects the luck of the draw.

If variation expresses legitimate pluralism or useful expertise, protect it. If it means similar people receive different sentences, diagnoses, prices, inspections, interviews, or ratings because of who happened to judge them, treat it as a design flaw. The book gives that flaw a name, a way to measure it, and a discipline for reducing it without pretending that human judgment can become error-free.