Skip to main content


Variability in high-stakes decision-making is more prevalent than we might like to believe

Noise: A Flaw in Human Judgment

Daniel Kahneman, Olivier Sibony and Cass R. Sunstein
Little, Brown Spark
464 pp.
Purchase this item now

It is not uncommon to read about criminals receiving drastically different sentences for seemingly similar crimes. We may rationalize these differences, supposing we do not know all the relevant facts or the nuances of the laws applied. We generally trust the expertise of the judges tasked with sentencing the accused. Yet, shortly after Judge Marvin Frankel first donned his robe and was required to render a verdict in a federal case, it occurred to him that he “had no experience or special knowledge whatever about the subjects of penology, criminology, sentencing philosophies, or any other pertinent learning” (1).

Unsettled by this reflection, Frankel began to observe sentencing patterns among his colleagues in the 1970s and found enormous discrepancies in judgments of very similar crimes. In one early study he oversaw, 50 judges from different districts were asked to evaluate a set of identical hypothetical cases, and the variability of their opinions was alarming. For example, the very same extortion case was judged to merit anywhere from 3 years imprisonment, in the most lenient case, to 20 years, in the harshest, with certain judges additionally imposing considerable fines. Subsequent legal analysis found variation of this sort to be the rule rather than the exception. How is it that such serious judgments could essentially amount to a high-stakes lottery?

Daniel Kahneman, Olivier Sibony, and Cass Sunstein’s new book, Noise, offers a sustained series of responses to this question while cataloging similar phenomena across a number of other domains. Their primary contention is that variability in decision-making contexts where one expects consistency is both more prevalent and more consequential than one would imagine. In their words, “wherever there is judgment, there is noise, and more of it than you think.”
To statisticians, the concept of noise is familiar, often the product of measurement limitations. If you are observing a series of measurements and you find that repeated readings (absent any environmental changes) are inconsistent, then the measurements are noisy.

Unlike bias, which can be thought of as a systematic deviation from a given outcome, noise occurs in decision contexts without normative frameworks or known outcomes. One does not need to know the true value of a proposition—and, in fact, there may be no “true” value—to observe noisy results.

Consider the pronouncements of wine tasters or movie critics, for example. We might reasonably be suspicious of there being a “correct answer” in matters of taste and might observe that individuals tend to diverge in their evaluations. In these situations, divergences are inconsequential and expected, but in others, they can be both surprising and devastating.

The authors of Noise are especially adept at gathering examples of the latter type of noise, and they produce evidence of it in the most disparate of fields, ranging from hiring practices and performance evaluations to cancer diagnoses, financial forecasts, and even forensic science. They find that experts not only regularly disagree with each other in situations laypeople consider matters of fact but also sometimes disagree with their own earlier pronouncements.

The authors dedicate a great deal of attention to characterizing and accounting for sources and variations of noise, but their work aims to be constructive as well as critical. They prescribe what they call “decision hygiene,” urging decision-makers to regularly assume external perspectives, making use of relevant statistics instead of mere impressions; establish common anchors for shared decision-making frameworks to minimize scale subjectivity; and segment and strategically sequence the presentation of different dimensions of complex decisions to avoid cascades of influence, in which an early impression colors all subsequent assessments.

Noise is, by the authors’ own admission, more anecdotal than their previous works. Whereas both Kahneman’s Thinking, Fast and Slow and Sunstein’s Nudge (with Richard Thaler) were the culmination of well-vetted research programs, Noise is quite tentative at times, relying on ad hoc analyses and unpublished research. Furthermore, although impressive in its scope and ambition, statisticians are apt to be disappointed with the occasionally clumsy handling of concepts critical to their field. With this compromise, however, comes the assurance that the text will likely appeal to a broader audience than it might otherwise have and that such readers will discover a great deal about the sources, types, and means of minimizing decision-making noise.

References and Notes:

  1. M. E. Frankel, Current Contents 25, 14 (1986).

About the author

The reviewer is a lecturer at Sciences Po, Paris, France, and director at Ensemble Insight.