Group decisions can outperform individual judgments, but the result depends on how people interact, how large the group is, and how individual opinions are combined. Research from 2010 to 2026 reports measurable gains in collective intelligence, dyadic accuracy, and diagnostic performance, alongside evidence that deliberation can move teams away from the average member choice.
Table of contents
- Collective intelligence across group tasks
- Conversational equality and group composition
- What group size changes
- Deliberation and choices under risk
- Dyads, accuracy, and response bias
- Aggregation rules in diagnostic decisions
Collective intelligence across group tasks
Two studies published in 2010 tested collective intelligence in 699 people working in groups of two to five. Study 1 included 40 three-person groups that completed multiple group tasks. The average correlation among group scores across tasks was r = 0.28, indicating a shared pattern across different activities. A first factor explained more than 43% of group-task variance, while a second factor explained 18%.
The collective-intelligence score was also associated with performance on a criterion task. It predicted criterion-task performance at r = 0.52. In a joint regression, collective intelligence remained a significant predictor, with beta = 0.51 and p = 0.001. These results describe a relationship between a group-level capability and performance; they do not establish that one causes the other.
The group score was not simply a reflection of the smartest participant. Average member intelligence was not significantly correlated with collective intelligence in Study 1 (r = 0.19), and maximum member intelligence was also not significantly correlated (r = 0.27). Study 2 included 152 groups ranging from two to five members. Its first factor explained 44% of group-task variance and its second explained 20%.
Across both studies, average member intelligence correlated with collective intelligence at r = 0.15, p = 0.04, while the highest-scoring member’s intelligence correlated at r = 0.19, p = 0.008. For 63 people working alone on the architectural-design criterion task, individual intelligence predicted performance at r = 0.33, p = 0.009. In groups performing the same task, average member intelligence was not a significant performance predictor (r = 0.18).
Source: Evidence for a Collective Intelligence Factor in the Performance of Human Groups (2010).
Conversational equality and group composition
The same 2010 research links collective intelligence with the quality and balance of interaction. Average social sensitivity correlated with collective intelligence at r = 0.26, p = 0.002. Meanwhile, variance in members’ speaking turns was negatively correlated with collective intelligence at r = -0.41, p = 0.01. In practical terms, the measured groups with more evenly distributed speaking participation tended to have higher collective-intelligence scores.
Group composition was also associated with the score. The proportion of women in a group was positively correlated with collective intelligence at r = 0.23, p = 0.007. Women in the study scored higher than men on social sensitivity, with t(441) = 3.42, p = 0.001. The reported social-sensitivity mediation effect for the relationship involving women’s share had Sobel z = 1.93, p = 0.03.
The relationships should not be treated as interchangeable explanations. In a regression including social sensitivity, speaking-turn variance, and women’s share, only social sensitivity reached statistical significance: beta = 0.33, p = 0.05. The result identifies social sensitivity as the statistically significant variable in that model, not a universal rule that group composition guarantees better decisions.
What group size changes
Evidence from studies summarized in The primate workplace: Cooperative decision-making in human and non-human primates shows that more people do not always produce proportionally better decisions. In a 1971 US Air Force cadet study with groups of 3 to 15, consensus decisions were more accurate than the average of members’ individual decisions. Accuracy for groups of six was not significantly different from accuracy for groups of 12 or 15.
In a 2006 letters-to-numbers problem study, groups of two to five were compared with individuals. Groups larger than two outperformed the best individual efforts. Groups of two performed at the level of the best individuals, while groups of three, four, and five did not differ from one another.
Another problem-solving and consensus study from 1975 examined groups of two to 18 and reported 11 as the most effective group size. Consensus fell sharply as members were added at first, then the decline became less pronounced at larger sizes. A 1951 participation study covering groups of two to 12 found that average member participation declined as group size increased. In that study, groups of six were most conducive to leadership emergence.
| Study finding | Group sizes or result |
|---|---|
| US Air Force cadet consensus study, 1971 | Groups of 3–15; six was not significantly different from 12 or 15 in accuracy |
| Letters-to-numbers study, 2006 | Groups of 2–5; groups above two beat the best individual efforts |
| Problem-solving and consensus study, 1975 | Groups of 2–18; size 11 was most effective |
| Participation study, 1951 | Groups of 2–12; participation declined as size increased |
These findings measure different outcomes, including accuracy, consensus, participation, and leadership emergence. They therefore do not identify one universally optimal group size.
Source: The primate workplace: Cooperative decision-making in human and non-human primates. The source summarizes the legacy studies; the dates above are the study periods reported in that review, not newly conducted measurements.
Deliberation and choices under risk
A 2023 field experiment on team allocations recorded 488 team allocations. Teams averaged 1.83 deliberation rounds, with a maximum of seven rounds before agreement. Among 266 classified team decisions, 46.24% fell strictly between the two members’ individual choices. A further 25.56% equaled the less risk-averse member’s choice, and 16.17% were greater than that member’s choice.
At the more risk-averse end, 6.77% of classified decisions equaled the more risk-averse member’s choice, 1.13% were lower than that choice, and 4.13% equaled both members’ choices. Relative to the individual mean, 66.92% of team decisions were more risk-taking, 22.93% were more risk-averse, and only 10.15% equaled the mean of the two members’ choices. The study therefore reported that about 90% of team decisions differed from the average of the members’ individual choices.
The three team-choice categories differed significantly, with chi-square(2) = 30.337, p < 0.001. Deliberation was not always extensive: about 46% of teams reached a decision from their initial allocations without further discussion. The remaining 54% disagreed with their initial allocations and needed subsequent deliberation. Among teams requiring more discussion, 29.7% agreed in the second round and 20.3% after a third round.
Source: From individual decisions to team decisions under risk: evidence from a field experiment (2023). The percentages for classified decisions are based on the 266 decisions identified in that analysis.
Dyads, accuracy, and response bias
A 2026 dyadic change-detection experiment included 44 participants organized into 22 dyads. Mean hit rate was 0.532 for individual judgments and 0.703 for dyad judgments. Mean false-alarm rate was 0.009 for individual judgments and 0.011 for dyad judgments.
The dyads’ mean sensitivity, measured by A-prime, was 0.918, compared with 0.889 for the better member in each dyad. The difference was statistically significant: t(21) = 13.34, p < 0.001, with Cohen’s d = 2.843. The dyads also had a mean decision criterion B-double-prime of 0.845, compared with 0.897 for the better member. That difference was t(21) = 3.120, p < 0.01, Cohen’s d = 0.665.
The results distinguish sensitivity from response tendency. Dyads had higher measured sensitivity than their better member, while the decision-criterion comparison also shifted. These are experiment-specific measures from a change-detection task, so they should not be generalized automatically to every group decision.
A follow-up individual experiment analyzed 26 participants’ first and second judgments. Mean hit rate was 0.563 on the first judgment and 0.551 on the second. Mean false-alarm rate was 0.065 and 0.050, respectively. Follow-up A-prime was 0.853 on the first judgment and 0.860 on the second, with t(25) = -1.443, p = 0.162. Follow-up B-double-prime was 0.659 and 0.720, with t(25) = -1.926, p = 0.065.
Source: Dyadic interaction enhances sensitivity and shifts decision criterion in a change detection task (2026).
Aggregation rules in diagnostic decisions
A 2024 general-practice aggregation study used 260 GPs diagnosing nine vignette patient cases. A second dataset used 30 GPs diagnosing 12 patient actors in face-to-face cases. The simulated diagnostic groups ranged from 3 to 9 GPs.
In the vignette dataset, plurality aggregation improved accuracy as group size increased in 8 of 9 cases and reduced accuracy in 1 of 9 cases. The actor-patient control condition produced a different pattern: plurality aggregation reduced performance in 5 of 12 cases. In the actor-patient early decision-support condition, plurality aggregation reduced performance in 2 of 12 cases.
These results show why the aggregation rule and case format matter. A larger simulated diagnostic group was not uniformly beneficial across all cases or conditions. The findings concern plurality aggregation in the reported general-practice datasets and do not establish that voting is always the best way to combine clinical judgments.
Source: Collective Intelligence Increases Diagnostic Accuracy in a General Practice Setting (2024).