Collective decision-making

Collective decision-making is the process by which a group turns multiple people’s information and judgments into a shared conclusion or action. In The Wisdom of Crowds, it is neither a guarantee of wisdom nor a synonym for discussion: its quality depends on how judgments are formed, what information reaches the group, and whether the group has a genuine means of combining its members’ views. Properly organized, a collective can outperform its most capable individual; badly organized, it can suppress useful knowledge behind hierarchy, premature agreement, or an advisory process that never decides.

Conditions for Collective Judgment

The book identifies four conditions that make a crowd’s judgment more likely to be accurate: diversity of opinion, independence, decentralization, and aggregation. Each person needs some private information or perspective; judgments must not simply mirror those nearby; people need room to specialize and use local knowledge; and a mechanism must turn separate assessments into a collective result. When a large, diverse, independent group estimates a question with an objectively correct answer, individual errors can cancel in the average, leaving more of the information each person holds. Cancellation alone is not enough: the group must possess relevant information in the first place.

The market response to the Challenger disaster illustrates the distinction. Investors and potential investors were effectively estimating how much less four companies were worth after the explosion. A trader might not have known that Thiokol was responsible, but the market’s aggregated judgment reflected information dispersed among many people. The book presents this as a case in which the crowd’s verdict could be close to the truth even if no individual possessed the whole explanation. The point is not that markets have privileged access to facts; it is that, under suitable conditions, their prices can combine fragments of knowledge that no one person can assemble alone.

Other forecasting markets provide checks on the claim. In NFL betting, the public’s point-spread judgment is difficult for even informed gamblers to beat consistently, and the Mirage’s closing line is described as the most reliable available forecast of game outcomes roughly three-quarters of the time. The forecast is not exact, and occasional patterns—such as historical value in home underdogs late in the season—can be found. The result is a qualified argument about comparative reliability, not a claim of infallibility.

The Iowa Electronic Markets provide a second measure. A study of forty-nine elections from 1988 to 2000 found average election-eve errors of 1.37 percent in presidential contests, 3.43 percent in other U.S. elections, and 2.12 percent in foreign elections. Across 596 polls released during presidential elections in that period, the market price on the day of each poll was more accurate three-quarters of the time. The market remained comparatively small, with no more than about eight hundred traders, most of them men and a disproportionate, though declining, share from Iowa. Its accuracy therefore did not depend on resembling the electorate: it predicted voters better than asking voters to predict themselves.

Diversity, Deliberation, and Aggregation

Diversity matters especially in small groups and formal organizations, where a few people can exert disproportionate influence. Scott Page’s computer simulations assembled groups of ten or twenty problem-solving agents with different skill sets. Groups containing both more and less effective agents almost always did better than groups made up only of the strongest agents; randomly selecting a group and letting it work could match or beat the strategy of identifying the best individuals and placing them together. The finding does not make intelligence irrelevant—the successful groups included high performers—but shows why intelligence alone cannot supply the variety of approaches a group needs. Similar experts may bring overlapping skills and leave the collective with a narrower set of tools.

Small-group decision-making can also be efficient. In experiments by Alan S. Blinder and John Morgan, student groups making central-bank-style interest-rate decisions responded as quickly as individuals and made fewer wrong moves. The groups’ performance did not track the smartest member’s performance, suggesting that they were not merely borrowing one exceptional person’s ability. A study modeled on the experiment at the Bank of England reached similar conclusions: groups could decide intelligently, quickly, and better than their smartest members. These results support deliberation when it is organized to make collective judgment operative, rather than treating discussion as inherently wasteful.

The crucial institutional requirement is aggregation. A team established only to advise leaders loses the very advantage that justifies bringing its members together if their opinions are never combined into a decision. The Columbia Mission Management Team (MMT), for example, never voted on anything: members reported on separate aspects of the mission, but their actual judgments were not aggregated. Its failure was therefore not simply that a group reached the wrong conclusion; the team lacked a procedure that could turn its members’ views into a collective judgment.

That failure was reinforced by the meeting culture. NASA’s hierarchy meant that engineers on the Debris Assessment Team (DAT) did not receive a serious hearing for their concerns about the foam strike, including their insistence that images of the Orbiter’s wing were needed for an informed analysis. The discussions were at once too structured and insufficiently structured: they often followed a pattern in which a leader asked a question and another person answered, yet there was no deliberate process for inviting each member to address particular questions. At the January 24 meeting, the DAT’s supporting analysis and any division of opinion were omitted from the summary; no MMT member asked to see the study. The team began with a conclusion and interpreted incoming information to fit it. Hierarchy, narrow discussion, and absent aggregation combined to make the group less capable than the information available to it.

From Local Knowledge to Collective Results

Decentralization gives people room to specialize and use knowledge tied to a particular place, job, or experience—knowledge that cannot easily be summarized for distant decision-makers. It can therefore increase both independence and the amount of information available to a system. But decentralization has a corresponding weakness: information discovered in one part of an organization may never reach the rest. The design challenge is to keep knowledge local enough to retain its detail while making its collective significance available across the system.

The pre-September 11 U.S. intelligence community exemplified the distinction between sharing information and aggregating judgments. Its agencies had specialized roles, which could produce more varied analysis and reveal relationships invisible to one set of analysts. Yet there was no effective mechanism for drawing together not only the information but also the assessments of NSA analysts, CIA officers, and FBI agents. Richard Shelby’s proposed solution—a truly central intelligence agency—would have made it easier for one agency to oversee information, but at the cost of some diversity, local knowledge, and independence. The book’s alternative is not to preserve isolation: it is to connect information while retaining specialized agencies’ autonomy. Linking databases could have helped, and remarkably, two years after September 11 the government still lacked a single unified watch list drawing on the entire intelligence community.

The global SARS investigation shows decentralization and aggregation working together. After learning of a severe respiratory illness in Guangdong Province and subsequent cases in Hanoi and Hong Kong, the World Health Organization (WHO) called on eleven laboratories in countries around the world to investigate the cause. Beginning on March 17, 2003, they held daily teleconferences, shared images and test results through a WHO website, and exchanged virus samples. Laboratories worked in parallel, eliminated candidate causes, and tested one another’s findings. By March 21, researchers in Hong Kong had isolated a likely candidate, while the Centers for Disease Control and Prevention separately identified a virus that looked like a coronavirus. By April 16, after laboratories had detected and analyzed the virus and infected monkeys had developed SARS, the network was confident that the coronavirus caused the disease.

Cynthia Goldsmith first spotted the coronavirus through an electron microscope, but that observation alone did not establish that it caused SARS. The evidence depended on weeks of work across laboratories, including the elimination of other possible causes and tests showing the virus in a wide range of SARS cases. The WHO coordinated the network, but did not dictate each laboratory’s tasks or control how information was exchanged. The discovery was therefore collective not as a ceremonial attribution, but as a result produced by distributed inquiry: separate contributions became decisive when the laboratories made them mutually available and tested them together.

Organizations, Authority, and Decision Markets

The book contrasts collective decision-making with the corporate pursuit of consensus. Although American companies repeatedly praised teamwork and employee empowerment, collective decision-making was often confused with unanimity or prolonged discussion. Consensus-seeking groups can avoid disagreement and settle for the least controversial answer, particularly when members know one another. Inclusion in a meeting is not the same as sharing power: after World War II, decisions could pass through layer after layer of committees while authority remained concentrated at the top. At General Motors (GM), designing a new headlight required fifteen meetings, with the chief executive attending the final five.

This structure could dilute both information and responsibility. In a GM plant in Van Nuys, California, a supervisor showed two assembly-line workers the costly repair required when a bracket holding the sunshade was missed. One worker had been doing the job for two years without knowing what part she was welding. At Ford, more than fifteen managerial layers separated the chairman from a factory-floor supervisor, compared with five at Toyota. A chart from an unnamed company showed 223 possible bureaucratic paths for a new product idea before approval. Such layers made it harder for executives to know what was happening on the ground and gave workers little incentive to act on knowledge they possessed. Studies cited in the book also found that upwardly mobile young executives communicated problem-related information less accurately, while middle managers’ advancement was associated with withholding news of failures and conflicts.

A wider distribution of authority would let people closest to a problem make more of the decisions it requires, rather than sending every issue upward and filtering every solution back down. Local experience can be more useful than managerial rank, and responsibility can itself improve performance: in one experiment, people given a button that could silence loud, random noise solved five times as many puzzles and made fewer proofreading errors than people without the button, even though no one pressed it. The book does not treat decentralization as a cure-all. Its value depends on preserving channels through which local knowledge can inform decisions beyond the place where it originated.

Internal decision markets offer one such channel for forecasting. In the late 1990s, Hewlett-Packard ran small artificial markets in which employees from different parts of the company traded shares tied to printer sales in the coming month or quarter. With twenty to thirty participants trading for a week at a time, the markets outperformed the company’s forecasts 75 percent of the time over three years. In a separate experiment at InnoCentive, employees traded on realistic profiles and experimental data for six drug candidates, three known to be approved and three rejected by the FDA. The market rapidly sent the likely winners’ prices upward and the losers’ prices downward. Anonymity, incentives to seek out useful information, and a result less dependent on status could help these markets bypass political infighting and sycophancy.

The book nevertheless draws a firm boundary around their authority: decision markets should inform major corporate decisions, not make them. Forecasting and deciding are not identical tasks, and concentrated executive authority remains risky in conditions of complexity and uncertainty. Boards formally have the final say on important corporate decisions, but in practice they may defer to the CEO. The principle applies even when leaders are exceptionally capable: important decisions benefit from a team whose members’ judgments can be combined, rather than from the assumption that one person’s expertise can stand in for the group.

The Limits of the Crowd

Collective judgment is most readily tested when a question has an objectively assessable answer. A bubble experiment at Caltech makes that condition vivid by showing that aggregation can fail when traders follow one another rather than attend to the underlying value. Students traded imaginary company shares over fifteen five-minute periods; each share paid a dividend of 24 cents per period, for a maximum total of $3.60 if held throughout. Although the value declined predictably after each dividend and was never uncertain, prices rose to $3.50 and remained there almost to the end. Students bought at prices far above the remaining dividends because they expected to resell to someone else who would pay just as much. Trading activity here did not produce wisdom: imitation and the hope of finding a greater fool displaced judgment about value.

Democratic decisions pose a different limit. If the common good has no objective standard, political disagreement cannot be resolved by treating votes like estimates of a measurable quantity. Citizens may advocate conflicting policies while each believes the policy serves the public interest; collective intelligence does not by itself settle what that interest should be. Crowds that can perform well on cognition problems with checkable answers should not therefore be assumed to choose objectively correct candidates or policies. Nor does that establish that dictatorship, aristocracy, or rule by elites would do better: those systems carry the additional danger of unchecked and unaccountable power.

The broader account distinguishes cognition problems from coordination and cooperation problems. In the latter, collective solutions are less like a single correct estimate and more like rules or norms that emerge over time. They can help people coordinate or sustain cooperation without being definitively optimal, and they remain vulnerable to exploitation by free riders and tax evaders. Collective decision-making is thus a conditional capacity, not a universal verdict machine. Its promise lies in arranging institutions so that distinct knowledge can remain independent, become visible to others, and contribute to decisions—without confusing numerical participation, consensus, or authority with collective intelligence.