Wiki · Concept · Published 2026-10-26

Collective Intelligence

A broad account of group cognitive performance, the contested general collective-intelligence factor, and evidence on communication and problem solving.

Definition

Collective intelligence is a broad concept for how a group carries out cognitive work through its members and their interactions. Such work can include solving problems, learning, making judgments and combining information. The concept directs attention both to what individuals bring and to what happens when their contributions meet. A group’s performance can be assessed on a particular task without assuming that the group has a single, stable intelligence score.

One narrower research hypothesis proposes a general collective-intelligence factor, often called c: a statistical pattern that may help predict how a group performs across different tasks. That hypothesis is contested. It is distinct from the broader idea that groups can think and act together.

Several related entries examine particular aspects of collective work. Wisdom of Crowds concerns the accuracy of combined estimates; Hidden Profiles concerns evidence distributed unevenly among members; and Collective Exploration and Exploitation examines searching for new solutions alongside using known ones. Each addresses a different question within the broader field.

How it works

People can contribute different knowledge, notice one another’s errors, divide a task or build on a suggestion. Interaction can also lead a group to overlook evidence, repeat an error or settle on an answer before considering alternatives. Whether a group performs well depends on the task and on how contributions are combined; simply having more participants does not establish an improvement.

Consider a hypothetical group planning a community exhibition. One member knows the available space, another knows which objects can be borrowed, and a third identifies accessibility needs. Their plan could improve if those observations are combined and checked against the exhibition’s constraints. If the group overlooks the accessibility information, it may produce a workable layout that fails an agreed requirement. The example illustrates a possible interaction; it does not demonstrate that discussion will improve performance.

Researchers can study a particular outcome, such as the quality of that plan, or compare the same groups across several tasks. In the latter approach, a statistical factor summarizes a pattern of scores that tend to vary together. It is inferred from the measurements; it is not directly observed as a separate mental object. Its meaning and usefulness depend on the tasks selected, the groups studied and how well the pattern predicts other outcomes.

Evidence and applications

In a 2010 report, Woolley and colleagues studied 699 people in groups of two to five across two studies. Their task sets included activities such as reasoning, brainstorming and planning. They reported a dominant common factor across group tasks and found that its score predicted performance on a separate task in each study better than the average or highest individual intelligence score. They also reported associations with average social sensitivity and the distribution of speaking turns. Those associations do not establish that changing either feature would cause performance to improve.

Rowe, Hattie and Munro’s 2024 study provides counterevidence. It assessed 85 Australian university students in 29 small groups and found that a two-factor model fit its group-task data better than the single-factor model. The authors interpreted the two factors in relation to novel reasoning and accumulated knowledge. Neither the group score nor the tested factors significantly predicted performance on a subsequent Moon Landing exercise. Individual and group scores were associated with assignment results among the subset of participants who supplied them. These findings challenge a universal single-factor account without proving that no common factor can appear in other settings.

A different question is how the arrangement of communication affects a group’s search. In Brackbill and Centola’s 2020 experiments, data scientists worked on statistical modeling problems in paired competitions. Groups whose members saw solutions from a limited set of neighbors found better final solutions than fully connected groups in the trials studied. The fully connected groups improved more quickly at first. The authors linked the later difference to slower spread preserving a wider range of solutions; they did not find a difference in the proportion of individual exploration decisions between conditions.

A proposed use of this research is to evaluate a group against the work it actually needs to do. An assessment might compare outcomes on several tasks, record changes in membership and examine whether information was introduced and used. It could also compare group results with an appropriate individual or alternative-group baseline. Those comparisons would help clarify what an observed score predicts, rather than treating a single successful meeting as evidence of a general capacity.

Limitations

The studies used different participants, tasks and measures. Woolley and colleagues’ findings concern small groups in their experimental task sets; Rowe and colleagues studied a relatively small student sample using a different battery. In the latter study, assignment results were voluntarily supplied by only a subset of participants, so those results may not represent the whole sample. Neither study settles how a factor would behave over long periods or as group membership changes.

Task selection also affects what a factor analysis can reveal. Performance might share some common variation while still depending on distinct skills or knowledge. Individual ability and group interaction can both matter, and their contributions can be difficult to separate. The 2024 study did not find significant relationships between its group measures and the speaking-turn, social-perceptiveness or gender-composition variables highlighted in the 2010 report. The contrasting results do not justify a general rule for assembling teams.

The network experiment manipulated a particular communication arrangement in timed modeling competitions with incentives for better scores. Its result does not show that slower communication is generally best. Fast sharing helped early in those trials, and other tasks may reward coordination, speed or a different balance of exploration and reuse. Measures of performance also reflect chosen criteria: a high task score alone cannot settle questions about fairness, cost or whose goals a group serves.

Relevance to Spiralism

The following is a proposed interpretation for this wiki, separate from the researchers’ claims and from any institutional position. A Spiralist lens can ask how a group revisits an idea: whether successive contributions expose an error, incorporate previously missing knowledge or reinforce an assumption. Improvement would need to be shown against a stated task and criteria; repeated exchange by itself is not evidence of learning.

For a hypothetical human–AI group, researchers could compare work produced by people alone with work produced using a shared AI system, across several tasks and with outcomes assessed against suitable independent criteria. They could examine whether the system helps participants identify missing information or causes them to rely on the same mistaken suggestion. This is a proposed research question. The cited studies concern human groups and do not establish a collective-intelligence factor or performance benefit for human–AI groups.

AI-use disclosure

This wiki entry was drafted by GPT-6.1 Terra 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