What is collective intelligence? Collective intelligence is the ability of groups of individuals — human, machine, or both — to act collectively in ways that seem intelligent, producing judgments, decisions, and solutions that can outperform what any member could achieve alone.
Collective intelligence, as defined by Thomas Malone and Michael Bernstein in the Handbook of Collective Intelligence (MIT Press, 2015), is 'groups of individuals acting collectively in ways that seem intelligent.' The field spans from the Condorcet Jury Theorem (1785) and Francis Galton's 1906 ox-weighing experiment to modern laboratory science: Anita Williams Woolley and colleagues (Science, 2010) found evidence for a general collective intelligence factor — the c factor — that explained roughly 43% of the variance in group performance across tasks and correlated more with social sensitivity, equality of conversational turn-taking, and the proportion of women in the group than with member IQ. The c factor has been debated (Credé and Howardson, 2017) and supported by a meta-analysis of 22 studies and 1,356 groups (Riedl et al., PNAS, 2021). The Collective Intelligence Genome (Malone, Laubacher and Dellarocas, 2010) maps building blocks of smart groups along Who, Why, What, and How. Current MIT Center for Collective Intelligence research (2026) focuses on generative AI and collective intelligence, Supermind Design, and designing human-AI teams; a 2024 meta-analysis (Vaccaro, Almaatouq and Malone, Nature Human Behaviour) found that human-AI combinations do not automatically outperform the best of either alone — genuine synergy requires deliberate design of task allocation and interaction. Current CCI systems include Supermind Ideator, which wraps a large language model in structured creative-problem-solving scaffolding, and a 2026 PNAS competition that studied more than 180,000 autonomous AI-to-AI negotiations. Within the decision-making lifecycle — Attend, Frame, Generate, Investigate, Aggregate, Deliberate, Allocate, Choose, Implement, Monitor — collective intelligence research primarily concerns the Aggregate and Deliberate stages: how groups pool dispersed knowledge and reason together. Argumentree operationalizes these findings through independent asynchronous argument submission, equal-participation structure, multi-dimensional rating aggregated into consensus scores, and a full audit trail of the group's reasoning.
What Is Collective Intelligence?
Collective intelligence is the capacity of groups — human, machine, or both — to act in ways that seem intelligent. It's not a metaphor: research has measured it, debated it, and shown how to design for it. This guide covers the science from Condorcet (1785) to MIT's human-AI team research (2026).
Last updated: 2026-07-18
TL;DR
Groups can be measurably smart. Woolley et al. (Science, 2010) found a collective intelligence factor (c) that predicts group performance across tasks — driven less by member IQ than by social sensitivity, equal turn-taking, and process. The finding is debated (Credé & Howardson, 2017) but supported by a 1,356-group meta-analysis (Riedl et al., PNAS, 2021). The modern frontier is human-AI collective intelligence: MIT's 2024 meta-analysis shows that adding a human to the loop does not automatically improve AI systems — synergy depends on task allocation and process design. Argumentree is built on that lesson: structure the reasoning, aggregate the judgments, keep the humans deciding.
Defining Collective Intelligence
The standard academic definition comes from Thomas Malone and Michael Bernstein's Handbook of Collective Intelligence (MIT Press, 2015): collective intelligence is "groups of individuals acting collectively in ways that seem intelligent." The definition is deliberately broad. It covers a market setting prices, a Wikipedia community writing an encyclopedia, a jury weighing evidence, an ant colony finding the shortest path to food — and, increasingly, hybrid teams of people and AI systems.
Collective intelligence is broader than the wisdom of crowds, which concerns the statistical aggregation of many independent judgments. Collective intelligence also includes genuine collaboration and deliberation — groups that interact, argue, and build on each other's reasoning. It differs from swarm intelligence, where coordination emerges from simple local rules without deliberation. And it connects to mechanisms studied in neighboring fields: voting systems aggregate preferences, prediction markets aggregate probabilistic beliefs, and structured deliberation aggregates reasons — three different answers to the same question of how to turn many minds into one decision.
Within the decision-making lifecycle — Attend, Frame, Generate, Investigate, Aggregate, Deliberate, Allocate, Choose, Implement, Monitor — collective intelligence research concentrates on the aggregation and deliberation stages: how dispersed knowledge gets pooled, and how groups reason together without the pathologies of information cascades and groupthink.
A Brief History of Collective Intelligence Research
The field has roots in Enlightenment mathematics, statistics, economics, biology, and computer science. The through-line: under the right conditions, groups outperform their members — and the conditions can be engineered.
The Marquis de Condorcet proves that if each voter is even slightly better than chance and votes independently, the probability that the majority is correct approaches certainty as the group grows — the founding mathematical result of collective intelligence.
Francis Galton analyzes 787 guesses of an ox's dressed weight at a 1906 country fair. The median guess (1,207 lb) lands within 1% of the true weight (1,198 lb) — better than the cattle experts. Published in Nature.
Friedrich Hayek's "The Use of Knowledge in Society" (American Economic Review) argues that markets aggregate dispersed, local knowledge through prices — the theoretical foundation for prediction markets.
James Surowiecki codifies the four conditions for crowd wisdom (diversity, independence, decentralization, aggregation). The same year, Hong and Page publish their contested "diversity trumps ability" model in PNAS.
Woolley, Chabris, Pentland, Hashmi, and Malone (Science) find evidence for a general collective intelligence factor in 699 people working in small groups — a group-level analogue of individual IQ.
Malone, Laubacher, and Dellarocas (MIT Sloan Management Review) catalog the building blocks — "genes" — of collective intelligence systems along four questions: Who, Why, What, and How.
Malone and Bernstein's MIT Press handbook consolidates the field and its standard definition: groups of individuals acting collectively in ways that seem intelligent.
Thomas Malone's Superminds describes five enduring types of collective intelligence — hierarchies, democracies, markets, communities, and ecosystems — and asks how computers can make each smarter.
Riedl, Kim, Gupta, Malone, and Woolley (PNAS) analyze 22 studies with 1,356 groups and find evidence for a collective intelligence factor across student, military, gamer, and online-worker populations.
Vaccaro, Almaatouq, and Malone's meta-analysis of 106 experiments (Nature Human Behaviour) finds human-AI combinations on average performed worse than the best of human or AI alone — synergy must be designed, not assumed.
Burton and 27 co-authors (Nature Human Behaviour) map how large language models can enhance collective intelligence — and how they risk homogenizing viewpoints and manufacturing false consensus.
MIT's Center for Collective Intelligence pursues three umbrellas: Generative AI and Collective Intelligence, Supermind Design, and Designing Human-AI Teams — including a PNAS study of 180,000+ AI-to-AI negotiations.
The c Factor: Can a Group Have an IQ?
In 2010, Anita Williams Woolley and colleagues asked a simple question with a century of individual-psychology behind it: just as one general factor (g) predicts an individual's performance across many cognitive tasks, is there a general factor for groups? They tested 699 people working in groups of two to five on a wide variety of tasks (Science, 330, 686–688). The answer was yes: a single statistical factor — c — explained about 43% of the variance in group performance across tasks.
What predicts a group's c — and what doesn't
Member IQ barely matters
Both the average and the maximum individual intelligence of group members were only weakly correlated with c. Hiring the smartest people does not, by itself, make a smart group.
Social sensitivity matters
The group's average score on the Reading the Mind in the Eyes test — the ability to infer mental states from facial cues — was a significant predictor of c (r = 0.26).
Equal turn-taking matters
Groups whose members contributed more equally to conversation had higher c; variance in speaking turns correlated negatively with c (r = −0.41). Groups dominated by one or two voices were collectively less intelligent.
Composition effects
In the 2010 study, groups with a higher proportion of women scored higher on c (r = 0.23) — an effect the authors link statistically to women's higher average social sensitivity, i.e. to process, not demographics per se.
Is the c factor real? The replication debate
Like most influential findings in psychology, c has been stress-tested — and readers deserve the honest state of the evidence:
The critique (2017)
Credé and Howardson (Journal of Applied Psychology) reanalyzed six samples and argued the empirical support for a single general group factor is weak — that group performance may be better described by task-specific abilities than one c.
The reply (2018)
Woolley, Kim, and Malone responded that the critique rested on scoring procedures and simulation assumptions that did not match the tasks groups actually performed.
The meta-analysis (2021)
Riedl, Kim, Gupta, Malone, and Woolley (PNAS) aggregated 22 studies covering 1,356 groups — students, military personnel, online workers, gamers — and found evidence for a c factor across all of them, with collaboration process and member social perceptiveness among its strongest correlates.
Bottom line: a large meta-analysis supports the existence of a measurable collective intelligence factor, while its exact structure and generality remain actively debated. What is not debated is the practical lesson: group process — who gets heard, how judgments are combined — predicts group performance better than member brilliance.
Statistical vs. Interactive Crowds
There are really two kinds of collective intelligence, and they run on different fuel:
Statistical crowds
Many independent estimates, aggregated mechanically — averaging, voting, market prices. This is Galton's fair and every prediction market since. Success depends on independence and diversity; the exercise is often one-shot.
Interactive crowds
People influence each other to build shared understanding — discussion, deliberation, collaboration. This is Wikipedia, a science community, a design team. Success depends on communication quality and the safety to speak; the work is ongoing.
The two modes pull in opposite directions: independence helps statistical crowds but starves interactive ones. A team that never talks can aggregate well but builds nothing new; a team that always talks can build shared understanding while quietly destroying the independence its judgments need.
Structured deliberation is an attempt to get both. In Argumentree, contributions and ratings are submitted independently — the statistical mode — while the argument tree they land in is a shared, evolving object the group deliberates over — the interactive mode. The structure decides when each mode applies, instead of letting the loudest mode win.
The Collective Intelligence Genome
If collective intelligence can be measured, can it be designed? Malone, Laubacher, and Dellarocas (MIT Sloan Management Review, 2010) cataloged the recurring building blocks — "genes" — of collective intelligence systems, organized around four questions. Any system, from Wikipedia to a corporate strategy process, is a combination of these genes:
Who is performing the task?
A defined hierarchy assigns the task, or an open crowd self-selects. Crowds bring scale and diversity; hierarchies bring accountability and coordination.
Why are they doing it?
The motivation genes: money (payment, prizes), love (intrinsic enjoyment, community), and glory (recognition, reputation). Sustainable systems align incentives with contribution.
What is being done?
Create (generate something new — designs, ideas, arguments) or decide (evaluate and select among alternatives). Most real processes chain the two.
How is it being done?
Independently (contributions aggregated without interaction — contests, voting, averaging) or dependently (collaboration, where contributions build on each other). The choice determines which failure modes you inherit.
Five Superminds
In Superminds (2018), Malone describes five enduring forms of collective intelligence that human societies keep reinventing — and argues the real question about AI is not "will it replace us?" but "how does it make each supermind smarter?":
Hierarchies
Decisions delegated up and down a chain of authority. Fast and accountable; vulnerable to information bottlenecks at the top.
Democracies
Decisions by voting. Legitimate and inclusive; how votes aggregate preferences is its own deep science (social choice theory).
Markets
Decisions through prices and trades. Superb aggregators of dispersed beliefs — the mechanism behind prediction markets.
Communities
Decisions by norms, reputation, and informal consensus. Powers open source and Wikipedia; slow but resilient.
Ecosystems
No shared decision mechanism at all — outcomes emerge from power and survival. The default when no other supermind is designed.
Supermind Design: Six Moves
MIT CCI's Supermind Design methodology turns the genome into a creative-problem-solving practice. Six "moves" systematically reframe a problem to surface new designs for groups of people and computers:
Zoom Out
Step back to the broader system or purpose the problem sits inside.
Zoom In
Decompose the problem into parts and processes that can each be redesigned.
Analogize
Ask how other domains — nature, other industries, history — solve a structurally similar problem.
Groupify
Ask how a different supermind — a market, a community, a democracy — would tackle the task now handled by one person or a hierarchy.
Cognify
Ask which cognitive processes (sensing, remembering, deciding, learning) the solution needs, and who or what should perform each.
Technify
Ask where technology — including AI — can perform, connect, or augment those processes.
Inside MIT's Center for Collective Intelligence (2026)
MIT's Center for Collective Intelligence (CCI), founded by Thomas Malone, is the field's flagship lab. Its earlier projects — the Climate CoLab, the CI Design Lab, Measuring Collective Intelligence, Collective Prediction — are now concluded legacy work. As of 2026, CCI's research is organized under three umbrellas:
Generative AI and Collective Intelligence
Using AI to augment human creativity — including tools like Supermind Ideator and DesignAID that scaffold human-AI ideation.
Supermind Design
Designing innovative combinations of people and computers, building on the genome and the six design moves.
Designing Human-AI Teams
Designing and allocating tasks in human-machine teams — the successor question to "human vs. machine." With the Singapore-MIT M3S program, CCI is building a systematic science of task allocation: which subtasks go to people versus AI, who supervises and verifies, and whether the AI acts as tool, assistant, peer, or manager.
Key findings from the 2024–2026 research wave
Scaffolding beats raw LLMs for ideation
Supermind Ideator wraps a large language model in creative-problem-solving scaffolding (including the six moves). In an experimental study, users generated significantly more innovative ideas than people using ChatGPT alone or working unaided; beta testers have created more than 10,000 ideas with the system (ACM Collective Intelligence Conference, 2024).
"Human in the loop" is not automatically better
Vaccaro, Almaatouq, and Malone's meta-analysis of 106 experiments and 370 effect sizes (Nature Human Behaviour, 2024) found human-AI combinations on average performed worse than the best of human or AI alone — with losses concentrated in decision-making tasks and gains in content-creation tasks. CCI now distinguishes strong synergy (the combination beats both humans alone and AI alone) from weak augmentation (it beats humans alone but not the AI) — and finds strong synergy requires deliberately designed task allocation, not just adding a human.
180,000 AI-to-AI negotiations
A 2026 PNAS study ran an international competition in which prompted AI agents conducted more than 180,000 negotiations with each other. Classic human negotiation science held up: warmth — positivity, gratitude, question-asking — was consistently associated with reaching deals and creating value, while dominance tactics predicted impasses.
Mapping where AI can be used at all
A 2026 MIT working paper (arXiv 2603.20619) reorganized roughly 20,000 work activities from the US O*NET database into a deep ontology and classified 13,275 AI applications against it — finding the top 1.6% of activities account for over 60% of AI market value. A tool for reasoning systematically about human-AI task allocation.
Rating AI and tasks on one scale
AGI-Elo (NeurIPS 2025, arXiv 2505.12844), from CCI-affiliated researchers including Malone, rates AI models and individual test cases against each other like chess players — putting a model's competency and a task's difficulty on one comparable scale, a prerequisite for principled task allocation in human-AI teams.
Collective Intelligence in the Age of AI
Large language models are becoming participants in collective intelligence, not just tools for it. Burton and 27 co-authors across disciplines (Nature Human Behaviour, 2024) mapped the double-edged consequences:
Opportunities
Lowering barriers to participation
LLMs can translate, summarize, and draft — letting more people contribute to collective processes regardless of language or writing skill.
Aggregating dispersed knowledge
LLMs can synthesize large bodies of text — meeting transcripts, open-ended survey responses, public consultations — that previously overwhelmed human aggregators.
Deliberation support
AI can surface counterarguments, cluster similar positions, and structure discussions — augmenting the Deliberate stage rather than replacing it.
Risks
Homogenization of viewpoints
If many people consult the same models, their outputs converge — eroding the diversity of perspectives that collective intelligence depends on.
False consensus
Fluent AI-generated text can manufacture the appearance of agreement or majority opinion that does not exist among actual humans.
Shrinking information ecosystems
An information environment increasingly filtered through a few models risks the correlated errors that make crowds collectively dumb rather than smart.
The design implication mirrors the Vaccaro finding: AI improves collective intelligence when it is assigned the sub-tasks it is demonstrably good at — transcription, extraction, translation, summarization — while judgment, evaluation, and the decision itself stay with accountable humans. This is the human-AI-systems frontier (automation bias, trust calibration, meaningful human control) that decision science is now mapping.
How Argumentree Applies Collective Intelligence Research
Argumentree is a collaborative decision-making platform built directly on these findings. Each design choice maps to a result in the literature:
Independent, asynchronous contribution
Participants add arguments in their own time, before the group converges — protecting the independence and diversity conditions that aggregation research shows crowds need.
Structural equality of participation
Woolley's research links collective intelligence to equal turn-taking. In an argument tree there is no floor to hold: every participant's arguments enter the same structure and are weighed the same way.
Aggregation by multi-dimensional rating
Arguments are rated on helpfulness, clarity, accuracy, and completeness, and ratings aggregate into consensus scores — an explicit aggregation mechanism, so collective judgment is measured rather than assumed.
Deliberation with structure
The 4-step chain of questions, compromises, and reviews lets participants probe and refine arguments — dependent collaboration in genome terms, but inside a structure that keeps reasoning visible.
Human-AI task allocation done right
AI extraction turns meeting transcripts and documents into draft argument trees — the content-processing work AI is good at — while evaluation and decisions stay with people, consistent with the 2024 human-AI synergy evidence.
A memory for the supermind
The full audit trail (draft, open, closed lifecycle with versioning) preserves the group's reasoning, so the organization can learn from its own collective intelligence instead of losing it when the meeting ends.
Frequently Asked Questions
What is collective intelligence?
Collective intelligence is the ability of groups — of people, of machines, or both — to act collectively in ways that seem intelligent (Malone & Bernstein, Handbook of Collective Intelligence, 2015). It covers statistical crowd wisdom, collaborative production like Wikipedia, deliberating teams, markets aggregating beliefs, and modern human-AI teams.
What is the c factor in collective intelligence?
The c factor is a general collective intelligence factor identified by Woolley et al. (Science, 2010): a single statistical factor that explained about 43% of the variance in a group's performance across many different tasks, analogous to IQ's g factor for individuals. It correlated with the group's social sensitivity and equality of conversational turn-taking far more than with members' individual intelligence.
Has the c factor been replicated?
It is supported but debated. Credé and Howardson (2017) argued the evidence for a single general factor is weak, and Woolley and colleagues published a detailed reply in 2018. A 2021 meta-analysis by Riedl et al. (PNAS) covering 22 studies and 1,356 groups found evidence for a c factor across diverse populations. The practical takeaway — group process predicts performance better than member IQ — is robust across the debate.
What is the difference between collective intelligence and the wisdom of crowds?
The wisdom of crowds is the narrower phenomenon: many independent judgments, statistically aggregated, converge on accurate answers (as in Galton's 1906 ox experiment). Collective intelligence is the broader field: it includes crowd wisdom but also genuine collaboration, deliberation, markets, communities, and human-AI teams — any system in which a group acts intelligently.
What is the difference between collective intelligence and swarm intelligence?
Swarm intelligence is coordination that emerges from many simple agents following local rules — ant colonies, bird flocks, robot swarms — with no deliberation and no explicit reasoning. Human collective intelligence involves exchanging and evaluating reasons. They share deep principles (decentralization, local knowledge, aggregation), which is why decision platforms borrow from both.
What is a supermind?
A supermind, in Thomas Malone's 2018 book of that name, is any group of individuals acting together in a way that seems intelligent. Malone identifies five recurring types — hierarchies, democracies, markets, communities, and ecosystems — and argues that computers make superminds smarter mainly by enabling new combinations of them, not by replacing them.
What is the MIT Center for Collective Intelligence?
The MIT Center for Collective Intelligence (cci.mit.edu), founded at MIT Sloan and directed by Thomas Malone, studies how groups of people — and increasingly people plus AI — can be organized to act more intelligently than any individual. Its earlier crowdsourcing-era projects (the Climate CoLab and CoLabs, Measuring Collective Intelligence) are now legacy work; as of 2026 its research is organized under three umbrellas: Generative AI and Collective Intelligence, Supermind Design, and Designing Human-AI Teams, with tools like Supermind Ideator and DesignAID and, with the Singapore-MIT M3S program, a systematic science of human-AI task allocation.
Does adding a human in the loop improve AI decisions?
Not automatically. A 2024 meta-analysis of 106 experiments (Vaccaro, Almaatouq & Malone, Nature Human Behaviour) found human-AI combinations on average performed worse than the best of the human or the AI alone, with losses concentrated in decision tasks. Combinations helped most in content-creation tasks. Effective human-AI collective intelligence requires deliberately allocating sub-tasks to whichever party does them best.
How can teams increase their collective intelligence?
The evidence points to process design: equalize participation (structured or asynchronous input rather than open discussion), protect independent judgment before group convergence, recruit for cognitive diversity and social perceptiveness, use an explicit aggregation mechanism such as rating or voting, and keep a record of the reasoning so the group can learn. Tools like Argumentree build these mechanics in by default.
References & Further Reading
Condorcet, M. de (1785). Essai sur l'application de l'analyse à la probabilité des décisions rendues à la pluralité des voix.
The founding mathematical result: majorities of independent, better-than-chance voters converge on truth.
Galton, F. (1907). Vox Populi. Nature, 75, 450–451.
787 fairgoers guess an ox's weight; the median lands within 1% of the truth.
View source →Hayek, F. A. (1945). The Use of Knowledge in Society. American Economic Review, 35(4), 519–530.
Markets as aggregators of dispersed knowledge.
Surowiecki, J. (2004). The Wisdom of Crowds. Doubleday.
The four conditions: diversity, independence, decentralization, aggregation.
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.
The c factor: 699 people, groups of 2–5, one factor explaining ~43% of performance variance.
Malone, T. W., Laubacher, R., & Dellarocas, C. (2010). The Collective Intelligence Genome. MIT Sloan Management Review, 51(3), 21–31.
The Who/Why/What/How building blocks of collective intelligence systems.
View source →Malone, T. W., & Bernstein, M. S. (Eds.) (2015). Handbook of Collective Intelligence. MIT Press.
The field's standard reference and definition.
Credé, M., & Howardson, G. (2017). The structure of group task performance — A second look at "collective intelligence." Journal of Applied Psychology, 102(10).
The principal critique of the c factor's generality.
Malone, T. W. (2018). Superminds. Little, Brown.
Five supermind types and how computers make them smarter.
Riedl, C., Kim, Y. J., Gupta, P., Malone, T. W., & Woolley, A. W. (2021). Quantifying collective intelligence in human groups. PNAS, 118(21).
Meta-analysis: 22 studies, 1,356 groups, evidence for c across populations.
Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303.
106 experiments: human-AI combinations don't automatically beat the best of either alone.
View source →Burton, J. W., et al. (2024). How large language models can reshape collective intelligence. Nature Human Behaviour, 8, 1643–1655.
28 authors on LLM opportunities and risks for collective intelligence.
View source →Advancing AI Negotiations: A Large-Scale Autonomous Negotiations Competition. PNAS (2026).
180,000+ AI-agent negotiations; warmth wins even machine-to-machine.
View source →Cai, A., et al. (2026). Where can AI be used? Insights from a deep ontology of work activities. arXiv 2603.20619.
A deep ontology of ~20,000 work activities for reasoning about human-AI task allocation.
View source →MIT Center for Collective Intelligence. Research (2026).
Current umbrellas: Generative AI and CI, Supermind Design, Designing Human-AI Teams.
View source →MIT Center for Collective Intelligence. Supermind Design methodology.
The six design moves: Zoom Out, Zoom In, Analogize, Groupify, Cognify, Technify.
View source →Sun, S., et al. (2025). AGI-Elo: How Far Are We From Mastering A Task? NeurIPS 2025.
Joint Elo-style rating of AI models and test-case difficulty; CCI-affiliated co-authors including Malone.
View source →Related Concepts
Turn your team into a measurably smarter group
Collective intelligence is a design problem: equalize participation, protect independent judgment, aggregate explicitly, and keep the reasoning. Argumentree builds those mechanics into every decision — with AI doing the busywork and your people doing the judging.
Start Free Trial