SupercivilizationThe Wrong Assumption
We have spent a century optimizing for individual intelligence. Hiring screens filter for it. Schools rank by it. Compensation rewards it. The implicit model is additive: put smart people together and you get a smart group. Put the smartest people together and you get the smartest group.
This model is wrong. Not slightly wrong — structurally wrong.
In 2010, a research team set out to answer a deceptively simple question: do groups have a measurable general intelligence, the way individuals do? Individual IQ is one of psychology's most replicated findings. A person who scores well on verbal reasoning also tends to score well on spatial reasoning, pattern recognition, and working memory. This cross-task consistency is called the g-factor — general intelligence.
The question was whether groups exhibit something similar. Does a group that performs well on one task tend to perform well on different tasks? And if so, what predicts it?
The Discovery
The researchers assembled nearly 700 people into groups of two to five. Each group worked through a battery of tasks: brainstorming, puzzle-solving, moral reasoning, planning, negotiation. The tasks were deliberately diverse — if a group did well across all of them, it would not be because of task-specific expertise.
The results were unambiguous. Groups do have a measurable general intelligence — a c-factor — that predicts performance across tasks as reliably as the g-factor predicts individual performance. This was not a weak statistical signal. It was strong, replicable, and consistent across different group compositions.
Then came the finding that changed everything.
What Does Not Predict Collective Intelligence
The researchers tested every obvious candidate for what drives the c-factor.
Average intelligence of group members. Weak predictor. A group of moderately intelligent people with good interaction patterns outperforms a group of brilliant people with poor ones.
Maximum intelligence — the smartest person in the room. Also a weak predictor. The "genius in the group" model does not hold. Having a standout member does not reliably lift the group. Sometimes it harms it, especially when the most intelligent member dominates discussion and suppresses contributions from others.
Group motivation and satisfaction. Not predictive. Groups that reported high motivation and enjoyment did not systematically outperform groups that did not. Feeling good about collaboration is not the same as being good at it.
Group cohesion. Also not reliably predictive. Tight-knit groups sometimes performed well, sometimes performed poorly. Cohesion can produce echo chambers as easily as it produces coordination.
What Actually Predicts It
Three factors explained most of the variance in collective intelligence.
1. Social Sensitivity
The single strongest predictor was the average social sensitivity of group members — measured by a standardized test that assesses the ability to read emotional states from facial expressions. This is not empathy as a personality trait. It is a perceptual skill: the ability to detect what others are thinking and feeling from nonverbal cues.
Groups with high average social sensitivity performed dramatically better across every category of task. They detected when someone had a relevant insight but was hesitant to share it. They noticed when a team member was confused and adjusted their explanation. They picked up on subtle disagreement before it became conflict.
This finding inverts the standard hiring model. Social sensitivity is rarely tested, rarely rewarded, and rarely even discussed in most organizational contexts. It is treated as a "soft skill" — pleasant but peripheral. The research suggests it is the hardest skill of all when it comes to group performance.
2. Conversational Turn-Taking
The second predictor was the equality of conversational turn-taking within the group. When all members contributed roughly equally to discussion, collective intelligence was high. When one or two members dominated — even if those members were the most knowledgeable — collective intelligence dropped.
This correlation between equal turn-taking and the c-factor (r = −0.41 for variance in speaking turns) is one of three statistically significant predictors Woolley identified, alongside social sensitivity and proportion of women. What makes a group smart or stupid is meaningfully bound up in whether everyone gets to talk.
The mechanism is straightforward. Information held by a single person is only useful to the group if that person shares it. Unshared information is the most valuable kind — it is what the group has that no individual has. But people do not share information in environments where they expect to be interrupted, overruled, or ignored. Turn-taking equality is not politeness. It is an information extraction mechanism.
3. Proportion of Women
The third predictor was the proportion of women in the group. Groups with more women performed better. But this effect was fully mediated by social sensitivity — women in the sample scored higher on social sensitivity on average, and when social sensitivity was controlled for, the gender effect disappeared.
The takeaway is not about gender essentialism. It is about what skills we screen for when composing groups. If we select for social sensitivity directly, we get collectively intelligent groups regardless of demographic composition.
Why This Matters for Social Architecture
The c-factor research has implications far beyond team composition.
Organizations Are Groups of Groups
Every organization is a nested hierarchy of groups. If collective intelligence at the team level depends on interaction quality rather than individual talent, then organizational intelligence depends on how groups interact with each other — not on how many talented individuals the organization contains.
This means the design of interaction patterns — who talks to whom, how decisions flow, how conflict is handled — is more important than the design of hiring filters. Most organizations invest heavily in selecting the right people and barely at all in designing the right interactions.
Governance Is a Collective Intelligence Problem
Democratic governance is, at bottom, a problem of collective intelligence. A society must make decisions under uncertainty, integrate information from many sources, and adapt to changing conditions. The c-factor research suggests that the quality of democratic outcomes depends less on the intelligence of individual citizens and more on the structure of their interactions.
When a few voices dominate public discourse — through wealth, media access, or institutional position — collective intelligence drops, regardless of how informed those voices are. When participation is distributed and turn-taking is roughly equal, collective intelligence rises, even if individual participants are less informed.
This is not an argument for populism. It is an argument for structural design. The question is not whether experts should inform decisions — they should — but whether the structure of deliberation allows distributed information to surface.
Scale Is the Enemy (Without Design)
The c-factor research was conducted in small groups. As groups grow, conversational turn-taking becomes mechanically impossible — 150 people cannot all speak in a meeting. Social sensitivity becomes less effective when you cannot see faces.
This creates a design challenge: how do we preserve the conditions for collective intelligence as organizations scale? The answer is not to avoid scale. It is to maintain small-group dynamics within larger structures through careful nesting — an idea we will explore in detail.
The Human-AI Complication
By 2024, the most important question in collective intelligence research stopped being "what makes human groups smart?" and started being "what happens when one of the teammates is an AI?"
The 2010 study assembled humans. By 2024, the practical reality is that nearly every knowledge-work team includes an AI collaborator — sometimes formally (Claude, ChatGPT, Copilot in the workflow), sometimes informally (the team member silently feeding work through an LLM). The empirical evidence is now clear enough to take seriously.
Vaccaro, Almaatouq, and Malone (MIT Center for Collective Intelligence, 2024) published a meta-analysis covering 106 studies and 370 effect sizes on combinations of humans and AI. The headline finding: on decision-making tasks, human-AI teams currently underperform the better of human-alone or AI-alone. They outperform only on creative or generative tasks — where the AI extends what the human can produce — and only when the workflow is structured to use the AI as a tool, not a teammate.
This is not a small finding. It directly inverts the optimistic c-factor narrative when applied to hybrid teams. The c-factor was derived from human groups where social sensitivity and turn-taking equality could function — properties an LLM does not have. An AI does not feel safety, does not reciprocate, does not have stake in the outcome, and does not need to be brought into the conversation by other members. The interaction mechanics that make human groups collectively intelligent break down because half the "group" is not actually playing the same game.
Heyman, Malone, and colleagues at MIT CCI demonstrated in late 2024 that scaffolded human-AI ideation — their Supermind Ideator tool — measurably raises creative output. The mechanism: the AI is treated as a generator of variants and provocations, not as a peer. When the workflow respects the asymmetry, the human-AI combination wins. When the workflow assumes the AI is just another collaborator, the combination loses.
The implication for the Supersociety thesis is direct: building collective intelligence at scale in 2026 requires designing for the AI seam, not assuming it can be ignored. Two productive directions are visible:
- AI-mediated deliberation. DeepMind's Habermas Machine paper in Science (October 2024) tested whether an AI mediator could help around 5,700 UK participants find common ground on contested questions. Participants preferred AI-mediated common-ground statements to those produced by human facilitators. The AI was not a peer; it was a structure that made human collective intelligence work better.
- Public-input model alignment. Anthropic's Collective Constitutional AI project used pol.is to ingest roughly 38,000 votes from 1,000 participants into a Claude variant — using human collective intelligence to constrain a model rather than using a model to substitute for it. The pattern generalizes: humans deliberate; AI captures and structures the deliberation.
The Vaccaro/Malone finding is not a refutation of the c-factor. It is a boundary condition: the c-factor is a property of human groups, and extending it to hybrid groups requires explicit design. Without that design, adding AI to a team makes the team measurably worse at decisions and only better at generation. This is going to be the load-bearing coordination-science finding of the next decade, and the implications run through every other piece on this site — from psychological safety to network states to the design of DAO governance.
The Structural Implication
If collective intelligence depends on interaction quality, then the structures that shape interaction are the most important infrastructure a society builds. Not roads. Not networks. Not institutions in the abstract. The specific patterns of who speaks, who listens, who decides, and how disagreement flows.
Most social failures are not failures of individual intelligence. They are failures of collective intelligence — groups that contain the information and capability to solve their problems but whose interaction patterns prevent that information from surfacing.
The c-factor is not a discovery about teams. It is a discovery about the fundamental mechanism by which groups of humans become more — or less — than the sum of their parts.
We already know how to build collectively intelligent small groups. The open question is whether we can design institutions that preserve collective intelligence at scale. The evidence says we can — but only if we treat interaction design as infrastructure, not afterthought.