Collective Intelligence and AI, Part 1: What actually makes a group smart together
The dominant story of AI is individual augmentation. But collective intelligence, the emergent capability that arises from social interaction between people, requires something different and may be at risk of being designed away.
The world is awash with AI tools and products being launched; ChatGPT, Claude, Lovable, MidJourney, Gamma, Cursor, NotionAI, the list goes on. Pick almost any of them and you’ll find one target goal: enabling one person, alone, to do more. Not a team or a community, just one person “augmented” by AI.
This isn’t an accidental decision, it’s the explicit vision for many founders of these companies. Sam Altman talks openly about the first one-person billion dollar company as something his “tech CEO friends” are betting on. The vision they have is…telling: “the future of startups could just be one person and 10,000 GPUs.”
This has become the dominant story of what AI is for, but I think it’s the wrong one. Not because AI isn’t useful, but because this story is steering us away from what AI could actually help us do together.
These tools genuinely do a lot. ChatGPT’s launch shook things up, and Claude, Mistral and others quickly followed, putting something close to an intelligent partner, with access to most of the internet’s knowledge, instantly available. A person no longer needs to find an expert to explore a topic, learn a language, or sketch out a business plan. They can just ask.
Tools like Lovable, Cursor and Claude Design have collapsed the barrier to launching digital products. Someone with an idea but no engineering team can now reach a working prototype in a day and something shippable in weeks. Even video, which used to need a scriptwriter, a videographer, a sound engineer, a graphics designer and an editor, is increasingly in the hands of one person with a few tools and a few hours.
It’s hard to overstate how much these tools can empower individuals to bring ideas to life and do more on their own than ever before. People who used to be locked out by training, money or social access can now do work that wasn’t possible a few years ago, even if the gates haven’t come down equally for everyone.
But it’s important to notice what’s happening in these examples. The work that used to be done by whole groups, like a founder with an engineering team or a film crew with five specialised roles, is increasingly being done by one person with a set of tools. It’s not just that these products empower people to do more, it’s that the collaborative parts of the work are being absorbed, by design, into the tools themselves.
That’s a particular type of usefulness, and it’s the type that almost every product I mentioned is optimised for. This usefulness is clearly real, but it’s not the only type of usefulness available, and what are we losing by focussing on it?
Throughout human history, we’ve been able to achieve incredible things when we work collectively, much more than any individual has been able to. It’s how we’ve built whole cities, cured diseases and sent people to the moon.
Even in the modern world, with advanced technology and automation, that’s still the case. The microchips in all of our phones and computers are built by EUV lithography machines that require thousands of people and multiple companies to manufacture, and no individual person knows how it all works.
Wikipedia is a common example, where the value isn’t just from people putting their thoughts onto a page, but from the discussion, debate and disputes that happen to make Wikipedia work. This is the opposite of individual work: the collective effort of hundreds or thousands of people working together to create something that is greater than could be achieved by them all working on their own.
There’s an important distinction I want to make early, between collective action and collective intelligence.
Collective action is when a group of people coordinate to achieve something, like people filling in their section of a Wikipedia article, an assembly line producing a product, or a logistics network moving goods. Working together, the group achieves something no individual could on their own, but the intelligence required for the work is largely held in the design of the coordination itself, not generated by the people in the moment.
Collective intelligence is an emergent capability, where the group is able to reason, understand and solve problems that no individual could on their own, and that even a coordinated group of individuals working in parallel couldn’t reach either: a jury finds something neither juror saw alone, a research team produces a paper that no member could have written, or a senior team makes a strategic call that none of them, separately, would have got to. The intelligence isn’t held in the coordination structure, it’s generated by the interaction itself: the disagreement, the building-on, the back-and-forth that surfaces things no one came in with.
This distinction matters because AI tools handle the two kinds of work very differently. For the most part, collective action is about coordinated execution, and workflow automation and AI are proven to be good at this. Collective intelligence is a different problem, because it’s actually created in the interaction with other people. Deliberation and discussion are the value. If you take the other people out and replace them with an AI tool, you haven’t sped up the deliberation. You’ve collapsed it and removed the thing that produced the value.
You might point out that AI tools can also challenge: you can ask Claude to disagree with you or play devil’s advocate. That’s true, but the disagreement is coming from a single model trained on a single distribution of data. It has the look of pushback, but not the variety of sources or perspectives. Genuine challenge needs genuinely different sources of perspective; an AI can simulate that but without the same substance.
It’s also worth flagging that some popular ideas about collective intelligence aren’t really about intelligence. Crowd wisdom, prediction markets, Galton’s famous estimation of an ox’s weight from a guess pool at a county fair are all examples of aggregation: people making independent guesses, averaged. That can produce strikingly accurate results, but it’s a different thing from what we’re talking about here. Real collective intelligence requires people interacting, building on each other, disagreeing and resolving. That’s the part AI tools don’t engage with.
If interaction is what makes collective intelligence work, then interacting with an intelligent machine should be a shortcut, right? Well, what actually makes a group of people smart together isn’t quite what you’d expect.
In 2010, Anita Woolley published research showing that groups have a measurable intelligence of their own, distinct from the intelligence of the individuals in them. They called it the c factor, and it explained roughly 40% of the variance in group performance across a wide range of tasks.
The interesting bit is that the c factor is not strongly correlated with the average or maximum intelligence of individual group members. The smartest people in the room don’t make a smart room. What predicts it is the average social sensitivity of the group, how well people read what others are feeling and thinking, the equality of conversational turn-taking, and how evenly participation is distributed. The finding has been replicated multiple times since.
This is a really important point: the drivers of group intelligence are social. They’re about how the people in the group sense each other, take turns, and create the conditions where everyone’s input shapes the outcome. None of these are things current AI tools do themselves or develop in the people who use them. A chat with Claude isn’t reading the room. ChatGPT doesn’t notice who hasn’t spoken. Lovable doesn’t repair a social rupture between co-founders.
If we want better collective intelligence, we need to design for the social fabric between people, not just the productivity of any one person inside it.
So, if the conditions for collective intelligence are social, but the overriding design model for AI is one person at a time, what happens to those conditions as these tools spread? This isn’t about looking at single conversations, but at scale, across teams, organisations, and people who are learning to work with them.
It’s a growing question that needs to be explored, and the productivity case for these tools doesn’t usually go near it. That’s what the rest of this series picks up: how these tools may be quietly eroding what makes collective intelligence possible, and what designing for it instead could look like.