A team has fragments of an answer. One person has read the technical report, another knows the people affected, and a third can see how the current plan failed last year. The team can improve its decision if those fragments are brought together in a way that exposes contradiction and preserves useful independence. It can also become more certain and less accurate if the first confident speaker sets the frame for everyone else. Information aggregation is the process that makes one group outcome from distributed information; its quality depends on what was distributed and how the group combined it.
Information aggregation is broader than taking the average of many estimates. It includes combining judgments, sharing evidence, dividing search across members, revising beliefs after discussion and using institutions or technology to collect signals. In some tasks there is a measurable correct answer; in others the group must decide among legitimate but competing aims. The mechanisms guide here explains when aggregation can help, when it can amplify error, and why a good process is not the same as mere consensus. Its companion Methods & Measures entry examines how to evaluate aggregation in studies.
What exactly is being combined?
Sometimes people each estimate one quantity, such as a future price or the number of items in a container. If their errors are partly independent, averaging can reduce some noise. That is the narrow wisdom-of-crowds case. In many real problems, people do not hold equivalent estimates. They hold different kinds of evidence: a direct observation, a technical model, a local history, a warning about implementation or a value judgment. A simple arithmetic mean of these statements would not be meaningful.
The first design question is therefore the unit of information. Are members supplying probabilities, ranked options, observations, reasons, candidate solutions or preferences? A probability about a verifiable event can be scored later. A concern about whose autonomy is affected cannot be judged solely by whether it aligns with the eventual majority. An organization that collapses these inputs into one score may appear decisive while losing the very heterogeneity it needed.
Kameda, Toyokawa and Tindale survey information-aggregation processes that extend beyond independent averaging. Social learning, majority rules, networks, expertise and collective search can all contribute under different conditions. Their review does not offer a universally best rule. It frames aggregation as a family of mechanisms whose success depends on the information ecology and the task being solved.
Input quality is an obvious limit. Combining many confident mistakes does not necessarily produce a good answer. But a group can sometimes extract a useful signal from members who are each imperfect, provided their errors are not perfectly aligned and the combination rule uses their information appropriately. The words “sometimes” and “provided” carry most of the mechanism. A crowd’s size alone does not make it wise.
Independence, dependence and shared evidence
Averaging works differently when people make independent judgments and when they copy the same source. If all members rely on one flawed report, collecting twenty votes mostly repeats the report’s error. If they observed different parts of a problem, their agreement can be more informative. Independence is not an absolute virtue, however: people can also learn from another member’s evidence and correct a mistake. The question is whether sharing adds information or simply propagates a prior answer.
Lorenz and colleagues showed that social influence in a controlled estimation task could narrow the diversity of answers without reliably improving the group estimate. Their result illustrates one failure mode: participants converge because they see other guesses, not because they have learned why an answer is better. It does not prove that discussion is always harmful. In a task where one person knows a crucial fact, withholding that fact to preserve statistical independence would be counterproductive.
One practical approach is to collect initial judgments privately and then allow evidence-sharing. This preserves a record of the group’s starting diversity while creating room for revision. The final judgment can be compared with the independent baseline. Such a sequence does not itself guarantee quality, but it makes the influence pathway visible. A facilitator can ask which new evidence changed a view rather than merely asking who switched sides.
Shared sources should be documented where they matter. Two participants may appear independent because they sit in different departments while both depend on the same vendor analysis. A group can deliberately recruit people with access to different evidence and ask them to disclose major dependencies. This is a practical design principle, not a numerical rule that every source must be unique. Some shared evidence is reliable and should be shared; the concern is hidden correlation in errors.
Social influence can help or hurt
The effect of influence depends on what people can observe, who influences whom and whether accuracy can be learned. In a fully centralized network, one early answer may reach everyone. If it is wrong, the group’s later agreement can be misleading. In a more distributed network, multiple perspectives may persist long enough to be compared. Becker, Brackbill and Centola’s networked estimation experiments showed that network structure altered whether social influence improved collective accuracy. That is evidence against a simple rule that independence must always be maximized or discussion always added.
Networks are not merely communication pipes. They determine which evidence is heard, whose revision is visible and whether disagreement survives. A person who often speaks first can anchor others even without formal authority. A technical specialist may be ignored if the interface shows popularity but not expertise. Conversely, a visible record of reasons can help a group identify why an unusual answer deserves attention. The best structure is task-dependent and has to be evaluated against meaningful alternatives.
Influence also has ethical and social dimensions. People may avoid dissent because of hierarchy or fear of embarrassment. An agreement reached under such pressure is not equivalent to an agreement reached after open examination of reasons. A mechanism account of aggregation should therefore include who can contribute and whether contributions are actually considered. More messages do not automatically mean more information.
Social learning strategies operate at the level of individuals. One person copies the majority, another copies the most successful member, and another samples others only when uncertain. Those rules shape the aggregate, but they are not the same as the aggregate outcome. The companion entry on measuring social learning strategies addresses how to infer those rules in experiments. Here the important point is causal: a group’s answer depends on how individual updating rules interact across a network.
Diversity, expertise and division of search
Functional diversity can expand the set of approaches a group considers. One member notices a modeling flaw; another sees a practical constraint; another has an alternative method. Hong and Page’s formal work shows conditions under which a functionally diverse set of problem solvers can outperform a set of individually stronger solvers in a specified search landscape. It is a theorem under its assumptions, not a claim that demographic diversity automatically solves every problem or that expertise is unimportant.
Expertise matters most when it is relevant to the task and can be recognized without making the group captive to status. A strong method specialist may improve an estimate; a local practitioner may notice a boundary condition the specialist’s model omitted. A good aggregation process can preserve both contributions. It may assign different weights to different claims, but the weighting rule should be explicit and open to challenge.
Groups can also divide work. Instead of asking everyone to solve the whole problem, they can search different documents, test different hypotheses or examine failure modes. This expands coverage, but it introduces a new challenge: how to reconcile findings when the parts depend on one another. A project that assigns each member a separate piece and never integrates the pieces has distributed work without aggregating information.
Collective intelligence research examines performance across multiple tasks rather than one estimate. Woolley and colleagues reported evidence of a group-level factor associated with performance across tasks in their studied small groups. Later analyses and critiques debate how robustly such a factor generalizes across task batteries. The finding is useful as a research program, not a license to label one team intrinsically “smart” from a single meeting. Group composition, interaction and task design all matter.
Rules, incentives and accountability
A combination rule can be mechanical, such as a mean or weighted average, or deliberative, such as a decision after arguments and evidence. Mechanical rules can reduce the influence of hierarchy when everyone has comparable, scoreable inputs. Deliberation can incorporate reasons and values that do not fit a common numerical scale. Neither is universally superior. A mean cannot resolve a normative conflict; a meeting can lose the statistical benefit of independent estimates.
Incentives affect what people reveal. A participant rewarded for being individually right may keep private information or avoid a risky but informative minority view. A participant rewarded only for agreement may conform. A group can set a norm of explaining evidence and welcoming corrections, but incentives in promotion, funding and public reputation still shape behavior. An aggregation process that ignores those pressures may misread silence as consent.
Accountability need not mean assigning blame for every wrong prediction. It can mean preserving an audit trail: initial judgments, new evidence, revisions, the final rule and the reason for a decision. This trail allows a team to learn whether it improved by aggregating information or merely became more confident. It also helps distinguish a bad outcome from a bad process when uncertainty was high and no option was guaranteed to succeed.
Technology can help gather dispersed input, but interface choices matter. A platform that displays the current group average before a person answers can create anchoring. A platform that shows vote totals but not reasons may favor popularity over evidence. A system that never lets a minority expert add context can erase a crucial signal. Design should follow the information problem: preserve independent input when useful, display provenance, allow reasoned revision and document how the final output was formed.
When an aggregate is better than its members
The relevant comparison must be chosen before declaring success. An average can beat a randomly chosen member but fail to beat a well-informed specialist. A final group decision can improve on its initial median but still miss a known benchmark. The advantage may appear only for some questions. A group may also be more accurate on average while making rare, severe mistakes that matter more in practice.
Prelec, Seung and McCoy’s “surprisingly popular” method illustrates one alternative to simple majority rule. It elicits not only answers but people’s predictions about how others will answer, then looks for an answer more common than those predictions imply. Under their task assumptions, that can identify a correct minority answer. It is not a magic method for contested values or any question with no verifiable truth. It demonstrates how a rule can use more of the information people hold than a simple count.
Aggregation may improve calibration as well as accuracy, or it may not. A combined probability forecast can be evaluated over many resolved events. A one-time decision about a complex social policy is harder to score. In the latter setting, learning may require process evaluation, stakeholder follow-up and explicit attention to harms and benefits. The principle remains the same: declare the target and comparator so that “better” has a meaning that can be examined.
There is also a temporal dimension. A group may aggregate well today but deteriorate after members leave or incentives change. A process should be resilient enough to retain useful dissent and revisit conclusions as evidence accumulates. This is not an argument for endless delay; it is an argument for decisions that record uncertainty and specify what new evidence would warrant revision.
What the evidence does not show
The cited research does not show that a large crowd is automatically more accurate than an expert, that social influence always harms accuracy, or that social influence always helps. Experimental findings depend on tasks, information, network structures and combination rules. A formal result about functional diversity depends on the model’s assumptions and does not identify a universal hiring rule. A collective-intelligence factor from a task battery is not a personal diagnosis of a group.
Agreement is not itself evidence of truth or ethical legitimacy. A group can converge on a false premise, exclude affected people or aggregate preferences while ignoring rights. Statistical accuracy on a verifiable target and justification of a decision with distributional consequences are different standards. A process can use good forecasts while making an unjust decision, or make an ethically defensible choice under genuine uncertainty.
No single aggregation design works across all settings. Independent first estimates, discussion, expertise weights, networks and prediction markets can each be useful, but each changes incentives and information flow. The appropriate question is what problem is being solved, what information is held by whom and what could be tested against an alternative. The Institute’s design questions here are synthesis from the cited literature, not a validated intervention package.
Return to the team
The technical report, lived impact and memory of an earlier failure do not become a better decision merely by being placed in one room. The team must give each source a route into the choice, test contradictions and avoid confusing the loudest or most repeated message with the strongest evidence. It can collect independent starting views, invite reasons, revise the shared model and retain a record of what changed.
Information aggregation is the mechanism by which distributed knowledge becomes a collective answer or action. It is strongest when its rule fits the information, its influence channels are visible and its outcome is compared with a meaningful alternative. That account is less romantic than saying that groups are wise by nature, but it gives researchers and practitioners something specific to examine and improve.
Research references
- Kameda, T., Toyokawa, W., & Tindale, R. S. (2022). Information aggregation and collective intelligence beyond the wisdom of crowds. Nature Reviews Psychology, 1, 345–357. 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, 108(22), 9020–9025. 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, 114(26), E5070–E5076. 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, 101(46), 16385–16389. View source
- Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., & Malone, T. W. (2010). Evidence for a collective intelligence factor in the performance of human groups. Science, 330(6004), 686–688. View source
- Credé, M., & Howardson, G. N. (2017). The structure of group task performance: A second look at “collective intelligence.” Journal of Applied Psychology, 102(10), 1483–1492. View source
- Prelec, D., Seung, H. S., & McCoy, J. (2017). A solution to the single-question crowd wisdom problem. Nature, 541, 532–535. View source