Is Collective Intelligence Real

Why We Act · 23 October 2026 · Reviewed by Faisal Zaidi

Galton's crowd guessed an ox to within 0.8 per cent, and the result has been read as proof that groups are smarter than people ever since. The later record is narrower: averaging works while guesses stay independent, groups pull ahead mainly on complex tasks, and the general factor called collective intelligence is real enough to survive a large pooling and contested enough to have failed a direct replication in part.

Answer to the question

Are groups smarter than individuals?

Sometimes. Averaging many independent estimates can beat most individuals, as Galton's 1907 ox contest showed, but discussion can erode that independence. Groups outperform individuals mainly on complex, interdependent tasks and gain no advantage on simple ones. Whether a general group intelligence exists is still contested.

Key takeaways

Galton's 1907 crowd of 787 put the middle guess for an ox at 1,207 pounds against a true 1,198, an error of 0.8 per cent, and the average only works while the errors stay independent.
In Lorenz and colleagues' 2011 experiment with 144 students, showing people one another's estimates narrowed the spread of guesses without improving the crowd's accuracy.
Almaatouq and colleagues found in 2021 that groups matched the fastest individual and beat the most efficient individual on complex tasks, and held no advantage on simple ones.
The 2010 collective intelligence factor survives a 2021 pooling of 22 studies and 5,279 people, yet Bates and Gupta found individual intelligence explained around 80 per cent of group differences.
Wageman's Xerox research attributes 37 per cent of the variance in team performance to design and under 1 per cent to the quality of coaching.

Two different claims travel under the name collective intelligence, and the evidence for them is not the same size. One is statistical: average enough independent estimates and the errors cancel. The other is psychological: a group, like a person, may have a general level of performance that carries from one task to the next. The statistical claim is more than a century old and the psychological one is not yet twenty, yet they are often cited as though the first confirmed the second.

The first holds within narrow limits. The second has been challenged by an independent team, and rests in part on a test whose validity is disputed. Colin Fisher's The Collective Edge draws on both.

Whether groups are smarter than individuals depends on what the group was asked to do, and on whether its members were allowed to talk first.

Independent guesses cancel out each other's errors

In 1907 Francis Galton reported in Nature, under the title Vox Populi, a contest at a livestock fair in Plymouth in which visitors guessed the weight of a dressed ox. He counted 787 usable cards. The middle guess was 1,207 pounds and the ox weighed 1,198, an error of 0.8 per cent. It has served as the standard exhibit for collective intelligence ever since.

The data arrived as separate cards, one estimate per person, and the middle value landed close because individual mistakes fell on both sides of the truth. The mechanism needs errors that do not line up. Kenneth Wallis's 2014 re-examination in Statistical Science adjusts the figures slightly, putting the outcome at 1,197 pounds and the true median at 1,208, and shows that the mean of the entries was right on the weight.

What happens when guessers can see one another was tested in 2011 by Jan Lorenz, Heiko Rauhut, Frank Schweitzer and Dirk Helbing, using 144 students at ETH Zurich in twelve sessions of twelve, answering six factual questions about Swiss geography and crime and giving five successive estimates to each. Where people were shown the others' estimates, the spread of guesses narrowed. The crowd's accuracy did not improve. As the estimates converged, the true value tended to sit nearer the edge of the group's range, and participants grew more confident while being no more correct.

A crowd that talks before it answers is a different instrument from the one Galton read. Whatever the crowd had to offer has been traded for agreement. Fisher's own endnote to his wisdom-of-crowds passage sets out boundary conditions that the sentence in the text leaves out, and the endnote is the safer guide.

A crowd that talks before it answers is a different instrument from the one Galton read.

Why does talking leave what members know unshared?

Groups meet in order to pool what each member knows. The founding test of whether they do is Garold Stasser and William Titus's 1985 study, which gave four-person groups of students a political caucus to run. Each member received only part of the information about the candidates, arranged so that pooling it all would reveal the best one. Groups chose the candidate that a plurality of members had favoured before they began. Recall taken before and after discussion showed that talking had kept the distorted pictures in place rather than correcting them. The shopping version of this problem is set out in [Group Decisions at the Checkout](https://thewhyweproject.com/articles/group-decisions-at-the-checkout).

The usual accounting is a subtraction. In Ivan Steiner's 1972 formula, actual productivity is potential productivity minus the losses from faulty process, so a group can only fall short of what its members could do. Richard Hackman later added the other term, process gains, the things a group can do that no member could do alone. With both terms on the page the question stops being whether groups are good and becomes which tasks let the gains outrun the losses.

Do groups pull ahead on hard tasks?

That question went to an experiment, published in PNAS in 2021 by Abdullah Almaatouq, Mohammed Alsobay, Ming Yin and Duncan Watts. Individuals and groups worked on tasks of varying complexity, built from a room-assignment problem in which students had to be placed under constraints. Complexity was set by the number of components and the interdependencies between them.

On the complex tasks, groups were as fast as the fastest individual and more efficient than the most efficient individual. On simple tasks the advantage did not appear. The authors measured both synergistic gains and process losses and found that the balance between them moves with task complexity. They suggest this may reconcile earlier studies that disagreed, because a simple laboratory task will make a group look worse than a real piece of knowledge work does.

That changes how the Zurich result reads. Estimating a number is a simple task, so a crowd gains little beyond what averaging already gives it and can lose a good deal by talking. A complex task with many linked parts is the setting where dividing the work can pay for the cost of coordinating it. The same group can show collective intelligence on one and none on the other.

One general factor, three studies and no settlement

The psychological claim has one origin, a 2010 paper in Science by Anita Woolley, Christopher Chabris, Alex Pentland, Nada Hashmi and Thomas Malone. Across a range of group tasks they found a general factor, which they called collective intelligence by analogy with general intelligence in a person. The factor was not strongly correlated with the average or the maximum intelligence of the members. It was correlated with the members' average social sensitivity, with how evenly they took turns to speak, and with the proportion of women in the group.

A direct challenge came from Timothy Bates and Shivani Gupta, whose three studies in Intelligence in 2017 gave a different picture. Individual intelligence accounted for around eighty per cent of the differences between groups, and the turn-taking and gender hypotheses were not supported. In 2021 Christoph Riedl and colleagues pooled twenty-two studies and 5,279 people in PNAS. The factor held up, predicting performance on new tasks, and was best predicted by the group's collaboration process, then individual skill, then composition.

A second dispute runs underneath. Social sensitivity was measured with the Reading the Mind in the Eyes test, whose construct validity is itself in question: a 2024 scoping review in Clinical Psychology Review of 1,461 studies that used it found that 63 per cent reported validity evidence in none of six categories, and concluded that the test's validity is largely unsubstantiated. The Collective Edge states the 2010 result at pages 57 and 58 without either dispute.

Read together, the three records leave collective intelligence in an unusual position. As a statistical effect it is well supported and easy to break. As a trait of groups it has a published original, a three-study challenge from an independent team, and a large pooling in its favour from authors that include two of the original five, Woolley and Malone. The summary the record supports is that a general factor probably exists, that what predicts it is unsettled, and that the best-known predictor was measured with an instrument whose validity is disputed.

Does design outweigh coaching in teams?

If a general factor exists, what builds it? The best-supported answer in Fisher's book is design. His chapter 12 leans on Ruth Wageman's research with teams of Xerox service technicians, in which features of team design accounted for 37 per cent of the variance in team performance, against less than one per cent for the quality of leaders' coaching. Design also accounted for 42 per cent of the variance in how far teams managed themselves, against under ten per cent for coaching.

Wageman's 2001 paper in Organization Science reports the same pattern in its abstract. Leaders' design choices affected how well teams performed their task, while coaching did not, and the two interacted: well-designed teams gained more from good coaching and lost less from bad coaching than poorly designed ones. It is a field study of one company's technicians, so it speaks most clearly to teams like them.

On this reading, collective intelligence is something a group's design permits before it is something its members supply. Design here means the things set before the team starts: who is on it, what task it owns, what authority it holds.

Several of the findings above are decisions of that kind. Someone decides whether people answer before or after they hear each other, and someone hands a group a simple additive job, the setting in which Almaatouq's groups showed no advantage. People who copy one another when they could answer separately are doing something that [can be rational](https://thewhyweproject.com/articles/when-copying-is-the-rational-move) for each of them and costly for the group, and a [belief that everyone else wants something](https://thewhyweproject.com/articles/everyone-thinks-everyone-else-wants-it) can pass for agreement when nobody does.

Independence is a design choice, since someone decides whether people answer before or after they hear each other.

The evidence does not say that groups are wise. It says that some arrangements let them be, and that among the Xerox technicians Wageman studied, the way the team was set up mattered more than the quality of the coaching its leaders gave it.

Common questions

Are groups smarter than individuals?

What is collective intelligence?

Collective intelligence covers two ideas. In the statistical sense it is the accuracy of an average over many independent judgements. In the psychological sense, proposed by Woolley and colleagues in 2010, it is a general factor in how well a group performs across different tasks.

Does a smart group need smart members?

On the original evidence, collective intelligence does not come from simply recruiting the cleverest people. The 2010 paper found the factor was not strongly tied to members' average or maximum intelligence. Bates and Gupta found in 2017 that individual intelligence explained around eighty per cent of the differences between groups. Riedl and colleagues' 2021 pooling found collaboration process predicted the factor most strongly, then individual skill, then composition, so members matter but the way they work together matters more.

Sources

Fisher, C. M. (2025). The Collective Edge. New York: Avery, an imprint of Penguin Random House

Galton, F. (1907). 'Vox Populi.' Nature, 75, 450-451, 7 March 1907. doi:10.1038/075450a0

Higgins, W. C., Kaplan, D. M., Deschrijver, E., & Ross, R. M. (2024). 'Construct validity evidence reporting practices for the Reading the Mind in the Eyes Test: A systematic scoping review.' Clinical Psychology Review, 108, 102378. doi:10.1016/j.cpr.2023.102378

Wallis, K. F. (2014). “Revisiting Francis Galton's Forecasting Competition.” Statistical Science, 29(3), 420-424. doi:10.1214/14-STS468

Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). 'How social influence can undermine the wisdom of crowd effect.' PNAS, 108(22), 9020-9025. doi:10.1073/pnas.1008636108

Stasser, G., & Titus, W. (1985). 'Pooling of Unshared Information in Group Decision Making: Biased Information Sampling During Discussion.' Journal of Personality and Social Psychology, 48(6), 1467-1478. doi:10.1037/0022-3514.48.6.1467

Steiner, I. D. (1972). Group Process and Productivity. New York: Academic Press

Almaatouq, A., Alsobay, M., Yin, M., & Watts, D. J. (2021). 'Task complexity moderates group synergy.' PNAS, 118(36). doi:10.1073/pnas.2101062118

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. doi:10.1126/science.1193147

Bates, T. C., & Gupta, S. (2017). 'Smart groups of smart people: Evidence for IQ as the origin of collective intelligence in the performance of human groups.' Intelligence (published online 2016). doi:10.1016/j.intell.2016.11.004

Riedl, C., Kim, Y. J., Gupta, P., Malone, T. W., & Woolley, A. W. (2021). 'Quantifying collective intelligence in human groups.' PNAS, 118(21), e2005737118

Wageman, R. (2001). 'How Leaders Foster Self-Managing Team Effectiveness: Design Choices Versus Hands-on Coaching.' Organization Science, 12(5), 559-577. doi:10.1287/orsc.12.5.559.10094

Wageman, R. (1995). 'Interdependence and Group Effectiveness.' Administrative Science Quarterly, 40(1), 145-180