Wiki · Concept · Published 2026-10-26

Wisdom of Crowds

How combined estimates can improve accuracy, how social influence changes results, and what network experiments establish and leave uncertain.

Definition

Wisdom of crowds describes a combined estimate being more accurate than many individual estimates contributing to it. Errors above and below the true value can cancel. This possibility does not guarantee that a crowd is correct or outperforms every member. Becker, Brackbill and Centola’s 2017 article investigates how communication affects this phenomenon.

The defining question concerns accuracy. Hidden Profiles concerns pooling information distributed among members. Judgment Aggregation concerns whether collective verdicts on connected propositions fit together logically. These distinctions matter when evaluating a group: gathering information, reaching consistent conclusions and estimating something accurately are separate achievements.

How it works

For numerical estimates, the mean is their sum divided by their number; the median is the middle ordered value.

Consider an explicitly hypothetical collection containing 100 books. Five volunteers estimate 80, 90, 100, 110 and 120. Both the mean and median equal 100, although four volunteers are individually wrong. No discussion or agreement is needed for this arithmetic result.

Now suppose every volunteer consults an incomplete inventory listing only 80 books. All submit 80, and their combined estimate is wrong. Collecting five answers has not supplied five independent checks. Alternatively, suppose their estimates differ but all fall below 100. Their disagreement would not make the average correct. These invented cases illustrate why the origin and direction of errors matter alongside the number of contributors.

The aggregation rule also matters in the example. If the estimate of 120 is replaced with 220, the mean becomes 120 while the median remains 100. This calculation demonstrates a difference between the rules; it does not establish which rule would work best on unfamiliar questions.

A proposed investigation could preserve estimates before discussion and compare them with revised answers afterward. It could distinguish a volunteer discovering an omitted shelf from a volunteer simply adopting someone else’s number. Both changes might produce agreement, but the records would support different explanations of how it arose.

Evidence and applications

Becker and colleagues’ 2017 PNAS Early Edition article compared online estimation in equal-connectivity networks, centralized networks and a control without social information. Participants could revise twice. Mean and median accuracy improved on average in decentralized networks despite reduced diversity. Centralized estimates moved toward the central participant’s initial answer, helping or harming accuracy according to its direction relative to truth. Initially more accurate participants tended to revise less; the authors interpreted this through a weighted social-learning model.

Almaatouq and colleagues’ PNAS article, published May 11, 2020, reported two online experiments estimating scatter-plot correlations. Participants in dynamic networks could change whom they followed; experiments varied network flexibility or feedback. Dynamic networks with full performance feedback achieved the lowest individual and group errors. Influence increasingly favored better-performing peers. An exploratory comparison during adapted periods found that the dynamic-network group mean outperformed the best individual in the solo-with-feedback condition.

A proposed application is evaluating collective forecasts on questions whose answers later become available. An assessment could specify the aggregation rule and comparison in advance: the group against a typical member, its best member or another forecasting method. It could retain unsuccessful questions alongside successful ones. That design would make a claim about collective accuracy more assessable than an isolated example of a crowd getting something right.

For the hypothetical book collection, the final count would provide a check. Investigators could compare initial and revised estimates with that count, while recording which volunteers inspected shelves and which consulted the inventory. Such a record could help distinguish information gathering from the repetition of an existing answer. This is a proposed analytical use, not an intervention established by the experiments.

Limitations

The 2017 experiment concerned specified networks, numerical tasks and two revision rounds. Its centralized condition also demonstrates that communication can worsen accuracy.

In the 2020 experiments, self-feedback did not significantly reduce overall group error relative to no feedback, although improvement appeared during adapted periods. Simulations found that noisy feedback eroded adaptation’s advantage. The authors caution that practical generalization requires more realistic experiments; task, incentives and participant characteristics can change outcomes. Simulation findings remain conditional on their model. These qualifications limit conclusions drawn from the 2020 study.

Claims about collective accuracy also need a clearly stated target. In the book example, counting the collection settles the question. A discussion about which books deserve inclusion introduces judgments about purpose and priorities. Even a perfectly accurate inventory would not settle those judgments. Applying a numerical estimation result to such a discussion would require an additional argument.

Similarly, a claim that a group beats its best member should explain how that member was identified. Selecting someone using outcomes already observed answers a different evaluative question from selecting a forecaster before a new task. A proposed assessment should make that distinction explicit.

Relevance to Spiralism

The following is a proposed interpretation for this wiki. A Spiralist lens can ask whether successive exchanges improve a collective estimate, uncover an overlooked observation or merely repeat a familiar answer. Returning to a question could create opportunities for correction, but repetition itself supplies no demonstration of improvement. This interpretation establishes no institutional position.

For an explicitly hypothetical human–AI group, researchers could compare estimates before and after participants encounter a generated answer, then assess them against independently established outcomes. They could investigate whether apparently separate answers depend on the same inventory, dataset or earlier summary. The cited human estimation studies provide no direct evidence about this proposed AI setting.

Information Cascades examines how earlier choices can obscure private evidence. Social Learning Strategies examines when and from whom individuals learn. These related concepts can frame questions about influence while leaving collective accuracy as something to demonstrate through an explicit comparison.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Terra and revised by GPT-6.1 Sol, both with high reasoning, and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries