A roomful of people guesses the weight of an animal. No single guess is especially impressive, yet the middle of the estimates comes close. The familiar story invites a simple conclusion: ask enough people and the group will know. That conclusion is too broad. A crowd can be wise when its members contribute partly independent information or errors that offset one another. It can be confidently wrong when everyone relies on the same misleading source, copies a visible estimate, or answers a question for which the average has no useful meaning.
Wisdom of crowds names a conditional improvement in collective judgment, not a magical property of numbers. The mechanism depends on the task, the distribution of individual judgments, the information people hold, and the rule used to combine contributions. In some settings, interaction improves a group by moving useful information through a network. In others, interaction destroys the independence that made aggregation valuable. This guide explains those mechanisms and limits. Estimating the size of the effect in a particular study is a separate measurement problem.
What the classic example shows
Francis Galton’s 1907 report on estimates of an ox’s weight is a historical starting point. The median of many fairgoers’ guesses was close to the measured value. A median is the middle judgment after guesses are ordered, so unusually high or low guesses have limited influence. The observation illustrates how dispersed errors can cancel or be resisted by a robust summary. It does not show that the crowd understood the animal better than every expert, or that a median is always the best aggregation rule.
The object had a numerical answer. That feature matters. Weight can be measured independently of the guesses, and individual errors can be described relative to the same quantity. Many collective decisions lack such a clean answer. A team choosing a care policy, a town balancing values, or a group interpreting an uncertain social event cannot simply average positions and call the result true. Crowd wisdom travels most easily to tasks with a shared target and a defensible way to score accuracy.
The classic example is also a snapshot. It does not tell us exactly what each participant knew, whether they had seen other guesses, or what would have happened with a different animal or a less familiar task. A historical demonstration can motivate a mechanism without validating it everywhere. That is why later experiments that manipulate information flow and group structure are crucial.
How aggregation can improve an estimate
Suppose each person tries to estimate the same value. Some overestimate; others underestimate. If their errors are sufficiently dispersed and not all driven in the same direction, a mean or median can be closer to the target than many individual estimates. Adding more judgments can reduce the impact of idiosyncratic noise. The exact gain depends on how errors covary. Ten guesses based on ten independent observations may contain more distinct information than a thousand guesses copied from one viral claim.
Independence is therefore relative, not absolute. People share education, culture, news, perceptual biases, and available evidence. A crowd need not be perfectly independent to gain from aggregation. It does need enough distinct information that the combination is not merely a repeated version of the same error. The important question is not only how many people answered, but how many genuinely different evidence paths or reasoning strategies shaped their answers.
Aggregation rules are part of the mechanism. An arithmetic mean is sensitive to extreme values; a median is more robust to a small number of outliers. A weighted average can use information about expertise or calibration, but a poorly chosen weight can amplify a confident mistake. Majority vote is suited to some discrete choices, yet it can fail when answer options are poorly framed or the majority shares a bias. There is no rule that is best for every task, and a successful result should identify which rule made it possible.
It is also possible to aggregate within a person. Someone who makes independent estimates at different times may gain by combining them, though repeated judgments can share memory and anchoring effects. The broader lesson is that independent error, not head count alone, is the resource. A crowd of one repeated without fresh evidence may have little diversity; a small group with genuinely different approaches may have more.
Diversity: useful differences, not a slogan
Groups can differ in experience, expertise, information, perspectives, and problem-solving heuristics. These dimensions are not interchangeable. Hong and Page’s theoretical work shows conditions under which a group of diverse problem solvers can outperform a group selected for individual high ability. Their result concerns diversity of approaches within a specified mathematical search setting. It is not a universal claim that any demographically diverse group automatically beats experts on any question.
Empirical work helps sharpen the boundary. De Oliveira and Nisbett found that demographic diversity, as operationalised in their crowd-judgment studies, typically added little to the wisdom gained from small crowds. That finding should not be turned into the opposite universal slogan either. Demographic attributes may correlate with relevant experiences in some settings and not others. The mechanism needs task-relevant variation in information or errors, and demographic categories are at best imperfect proxies for that variation.
Expertise matters when errors are not symmetric. If only a few participants understand a technical domain, treating every answer equally can wash out crucial knowledge. On the other hand, relying only on a narrow expert group can preserve shared assumptions and blind spots. A sound design can combine expert knowledge with varied independent judgments, then test performance against an answer key where one exists. The aim is not to romanticise lay crowds or to dismiss them, but to match contributors and aggregation to the task.
Why social influence can hurt
Seeing another person’s answer changes what an estimate means. A participant may revise toward a visible group value because it seems informative, because disagreement feels costly, or because it becomes an anchor. If many people do this, their answers become correlated. An aggregate then looks more consensual without necessarily becoming more accurate. Lorenz and colleagues showed experimentally that even mild social influence can undermine a crowd-wisdom effect in numerical estimation, including by narrowing the diversity of estimates and changing confidence patterns.
This result does not imply that people should never learn from one another. A participant who has poor information can improve by hearing a knowledgeable peer. The danger is undifferentiated influence: participants may copy an answer without learning why it is better, or a high-status member may steer everyone toward the same wrong assumption. In such cases, apparent agreement is evidence of a social process, not evidence that the target was correctly inferred.
The timing of exposure matters. Collecting initial judgments independently preserves a baseline of distributed information. Showing answers only afterward allows researchers or decision-makers to observe what discussion changed. If the group sees a running average before answering, it becomes difficult to know how much independent information remains. Anonymity, turn-taking, and explicit requests for reasons can affect these dynamics, but their benefits depend on the setting and should be tested rather than assumed.
Why social influence can also help
Network experiments complicate the simple independence rule. Becker, Brackbill, and Centola found that in decentralised communication networks, information exchange could improve the accuracy of collective estimates. The structure of who saw whose judgment mattered. Some social influence can help people correct errors while preserving enough variation for the group to avoid premature convergence. The same study also illustrates why a centralised network can be vulnerable when many people follow a few influential estimates.
The mechanism is not merely “more communication.” It is a pattern of information flow. If individuals receive local signals from different neighbours and adjust gradually, useful evidence can circulate without all judgments collapsing onto one early answer. If the network directs everyone to a single hub, that hub’s mistake can become a common mistake. The distinction is especially relevant in online environments, where rankings, popularity counts, and recommendation systems can make some answers disproportionately visible.
Interaction can also reveal reasons unavailable in bare estimates. A person may know a local fact, spot a hidden assumption, or explain why a source is unreliable. That is not the same mechanism as averaging noisy numerical guesses. Kameda, Toyokawa, and Tindale’s review places classic wisdom-of-crowds effects within a wider landscape of information aggregation and collective intelligence. The review helps distinguish statistical pooling from group processes that generate, transmit, or transform information. It should not be read as proof that every collaborative process produces a wiser result.
Information aggregation is wider than crowd wisdom
Some collective systems combine estimates about a known quantity. Others combine fragments of evidence, explore possible solutions, or negotiate priorities. The classic crowd-wisdom effect is most precise in the first case: a defined answer, independent or partially independent judgments, and an aggregation rule. Information aggregation can include discussion, debate, sequential learning, and institutional procedures. These may be valuable without resembling the ox-weight median.
A forecast about a future event introduces another challenge. The outcome is unknown at the time of the prediction, and probability judgments should be evaluated over many resolved questions. A one-off correct forecast may be luck. A crowd’s average probability can improve on individual forecasts under some conditions, but the result depends on question selection, calibration, incentives, and the mix of forecasters. Calling any group forecast “crowd wisdom” without these details obscures the mechanism.
Complex problem solving is different again. Combining partial solutions may require coordination, shared representation, or an ability to recognise which contribution is useful. An arithmetic average of proposed plans is rarely a workable plan. Hong and Page’s work addresses diversity in search heuristics, not only averaging opinions. A methods paper that tests group problem solving should therefore not be evaluated as if it were a numeric-estimation experiment.
When the crowd can be systematically wrong
Aggregation does not remove a shared bias. If everyone sees a distorted image, hears a false rumour, or uses the same flawed model, their errors may align. The mean of biased judgments can be precise and wrong. A large sample can make the wrong answer look stable. The problem is especially serious when confidence and popularity are treated as evidence of truth, because both can arise from repetition or coordinated influence.
Selection can also distort the crowd. People who volunteer an estimate may differ from those who stay silent. A platform may amplify enthusiastic or extreme responses. A panel may exclude people with firsthand knowledge. Even before aggregation, the population of contributors determines what information is available. Increasing the number of responses from a narrow pool will not recreate missing perspectives.
Some questions have no settled factual target. A moral or political choice can involve competing values, rights, and trade-offs. Public participation may be justified by legitimacy and representation, not only by a claim that averaging preferences reveals an objective truth. Wisdom-of-crowds language should not erase those democratic reasons or suggest that a technically accurate estimate settles what ought to be done.
What the evidence does not show
Galton’s historical result does not show that every crowd is wiser than its best member or that large groups automatically overcome misinformation. Lorenz and colleagues show that social influence can erode a numerical-estimation advantage; Becker and colleagues show that appropriately structured influence can improve it. Neither result licenses a universal rule that interaction is always harmful or always helpful. Task, network, and timing are essential conditions.
Hong and Page’s theorem does not prove that demographic diversity by itself guarantees better decisions. De Oliveira and Nisbett’s empirical findings do not prove that diversity never matters. The theoretical and empirical work refer to different forms of variation and different tasks. A claim about a particular organisation or public issue needs evidence about the information and error structure in that setting.
The reviewed research also does not establish that averaging answers is enough for complex decisions, that a consensus reflects independent agreement, or that a crowd score reveals the truth of a value judgment. The Nature Reviews Psychology synthesis maps a wider set of collective processes, but synthesis is not a single causal test. “Wisdom” remains a result to demonstrate under defined conditions, not a property to assume from the presence of many people.
A practical reading of a crowd result
When someone says a crowd was wise, first ask what the target was and how its truth was checked. Then ask who contributed, what each person could see before answering, and whether their information sources overlapped. Finally ask how responses were combined and what comparison makes the group look better: the median individual, the best individual, a random small group, an expert, or an earlier group result? Different comparisons answer different questions.
Those questions are not a recipe for one perfect group. They protect the mechanism from exaggeration. A group can be useful because independent errors cancel, because diverse approaches find different parts of a solution, or because a well-designed network helps information travel. It can fail because the same error is repeated, because influence centralises too early, or because the task was never one that an average could solve. Crowd wisdom begins where those conditions are examined, not where the crowd is simply counted.
Research references
- Galton, F. (1907). Vox populi. Nature. View source
- Kameda, T., Toyokawa, W., & Tindale, R. S. (2022). Information aggregation and collective intelligence beyond the wisdom of crowds. Nature Reviews Psychology. View source
- Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). How social influence can undermine the wisdom of crowd effect. Proceedings of the National Academy of Sciences. View source
- Becker, J., Brackbill, D., & Centola, D. (2017). Network dynamics of social influence in the wisdom of crowds. Proceedings of the National Academy of Sciences. View source
- Hong, L., & Page, S. E. (2004). Groups of diverse problem solvers can outperform groups of high-ability problem solvers. Proceedings of the National Academy of Sciences. View source
- de Oliveira, S., & Nisbett, R. E. (2018). Demographically diverse crowds are typically not much wiser than homogeneous crowds. Proceedings of the National Academy of Sciences. View source