Collective Intelligence and AI, Part 2: How individual AI tools erode collective intelligence

In my previous post, I ended with a question: what happens to the social conditions for collective intelligence as individual-focused AI tools spread, at scale, across teams and over time?

It turns out there’s a lot, and not in some theoretical way. We can already see erosion starting to happen in specific and observable ways. From my work and research, I’ve found five mechanisms that come up repeatedly, and they connect to a deeper systemic problem that goes beyond the social losses. This isn’t an exhaustive list — I think there’s probably more out there to be found — but it’s enough to see the shape.

1. They reduce the need to engage with other people in the first place

The most basic point is that these tools reduce the engagement that collective intelligence relies on. Why ask a colleague when Claude can answer? Why bring a half-formed draft to a team for feedback when you can keep iterating and refining with an AI? Why find an expert when you can prompt AI to act like one?

It’s the easiest, quickest option and makes sense at the time, but over time it chips away at how much we share thinking, collaborate and build on each other’s contributions.

A designer who used to ask a product manager what they thought now drafts five versions with Claude and only brings the polished one. The PM loses early visibility into the thinking, and the designer loses the cheap feedback that might have redirected them before five rounds of polishing. An engineer doesn’t ask the team if anyone has tried something similar, they just ask AI and miss that a colleague spent six weeks on the same problem last year. A leader iterating on strategy with AI can work for days down a tangent before anyone else even sees what they’re considering, by which point the cost of redirecting is much higher.

The work still happens, but with fewer chances for anyone to notice that it’s going wrong, or to bring something new to it.

2. They privatise the thinking that used to happen between people

The value of collaboration isn’t just the output, it’s the understanding built along the way. A meeting creates a shared memory of why decisions were made. A Slack channel becomes a history of working knowledge. Even a dreaded email thread creates a record that people can see, build on and challenge.

A chat with AI is usually private to one person. The documents, slides or tools that came out of that chat may be shared later, but the thinking behind them is invisible to everyone else. That thinking is often where the most important collective work used to happen: where shared mental models develop and where the reasoning behind decisions becomes part of the team’s memory. This isn’t about just showing outputs, but making sure people understand why the output is the way it is.

This can have impacts on how teams grow and function beyond the work itself. For junior people, they lose sight of how senior people work through problems, and the chance to learn from them. Team members can lose visibility of related work happening alongside their own. For leaders, it becomes harder to understand, reconstruct or challenge decision rationale later.

Even if the output looks more polished than before, the team’s shared knowledge and understanding stops growing.

3. They homogenise the creativity of the people who use them

There’s now a substantial amount of research showing that when people use GenAI for creative tasks, an individual’s outputs may improve, but the collective diversity of outputs across people drops significantly.

Doshi and Hauser (2024) found that GenAI-generated plot ideas made individual short stories more creative, but stories produced with AI assistance were significantly more similar to one another at the group level than stories produced without it. Anderson et al. (2024) found the same in ideation tasks, and Padmakumar and He (2024) found it again in AI-assisted text completion.

Zhou and colleagues (2025) went further, identifying what they called a “creativity illusion.” When users lost access to AI, individual creativity dropped sharply, and the homogeneity of outputs kept rising for months afterwards. AI had augmented creativity in the short term, but hadn’t helped people build their own capacity for it.

This is because models are trained on broadly similar sets of data and tend to create outputs around averages of their training distribution. When these models are then used by lots of people, the outputs get pulled towards a similar centre. In practice, if not used carefully, brands using AI can converge on similar voices. Companies using AI for product strategy land on similar propositions. Designers using AI for early concept work end up with similar layouts.

These are the things that distinguish one company’s experience, brand or proposition from another’s, and AI tools can quietly erode them at scale. Collective intelligence needs diverse perspectives interacting, but individual-focused AI tools do the opposite: they pull individual perspectives toward a shared centre, even if each person is using their own account.

4. They accommodate rather than challenge

These tools can be really accommodating. They are designed to be helpful and agreeable so that you enjoy using them. Sure, they can push back when asked, but the default is going along with whatever direction you’ve already set.

Collective intelligence relies on people not just agreeing all the time. Having someone who knows enough of your area to challenge you, ask awkward questions or push you to give a better answer can lead to better work. The friction helps.

Without being challenged, weak ideas can survive longer than they should, and things that should have been killed at week one end up being the priority at week five.

You can use prompts to be more critical or challenge your thinking, but there’s a difference between asking for simulated challenge and being challenged because someone else has a stake in the outcome. When someone has a stake in the outcome, their feedback carries weight. Without that, it’s just AI finding another way to be helpful.

5. They erode the social skills collective intelligence depends on

This is more speculative, but based on how workplaces evolved after the pandemic, it’s something we might begin to see more of.

The c-factor research suggests social skills play a big role in what makes groups smart. Reading what people are feeling, sensing when to speak and reading the room are learnt social skills, and they develop through practice. We have evidence that reliance on AI erodes other cognitive skills, like creativity and recall, so it wouldn’t be surprising if social skills faced the same challenge.

The worry isn’t that any one person becomes less socially capable from using ChatGPT, it’s that a generation that routes around the friction of human interaction may simply build less of the capacity to do it well. We’ve probably all experienced how the growth of text messaging has influenced whole generations to avoid phone calls.

If social skills are what makes collective intelligence possible, we may be setting ourselves up for a long-term loss: short-term productivity benefits in exchange for a gradual drop in the skills that produced the most impactful work humans have ever done.

None of these mechanisms are inevitable, but they are based on the default for tools designed and optimised for one person at a time rather than teams working together.

---

So far, I’ve focussed on what we can lose socially or creatively when we design AI tools to empower individuals, but there’s also a deeper, more systemic argument about how we work with and manage the risks of AI systems.

To manage a complex situation, you need to have at least an equal number of ways to sense and respond to that system as it can produce. The common example of this is a thermostat: if a room can get too cold, a thermostat needs to be able to turn heating on or off to maintain the temperature. It has enough ways to sense and respond to manage it. However, if the weather gets too hot and the thermostat doesn’t have air conditioning to turn on, then it can’t control the temperature anymore. The system it’s trying to control has more variety than the thermostat has ways to influence it. This is Ashby’s law of requisite variety.

This need for variety to control complex systems is why teams are better for hard, complicated problems. It’s not just so that you can produce more stuff, but because more people can sense more things, have different mental models and can respond in different ways. A team of five isn’t like having five identical people: each person has their own perspectives, notices different things, and might make different mistakes. This means they have more capacity for managing complicated systems, and if one of them makes a mistake the others are more likely to spot it.

Now think about AI tools designed for the individual. That’s one human sensor, working with one person’s mental model, with AI that builds on and amplifies that one person’s prompts. All the content AI can produce, and all the variety of decisions that need to be reviewed and checked, goes through that one person. They have limited capacity to check everything, and only one perspective.

What produces collective intelligence isn’t just the number of perspectives, it’s about independent and diverse perspectives interacting. Having four engineers critique a design would give you feedback on similar things; add a designer or a business analyst and they’ll catch things the engineers missed. A reliable system depends on the independence of its parts.

Planes have two pilots, not two autopilots, for exactly this reason. The pilots are different because each catches what the other might miss. They trained on different days, made different errors getting their licences, have different intuitions about the same instruments. Two autopilots running the same software can fail in exactly the same way at exactly the same time, which gets you nothing.

Some people will argue you can solve the solo person plus AI problem by adding more AI: supervisor agents, model-on-model checking, multi-agent systems built around internal disagreement. This is becoming a popular position, but it doesn’t quite work for the same reason the cockpit makes clear. AI regulating AI is closer to two autopilots than to two pilots. The systems share training distributions, architectural assumptions, and blind spots. The supervisor catches failures its own model can recognise and lets through everything else. Stack more layers of AI and you don’t get more independence; you get a taller tower of similar things.

The errors that don’t get caught also aren’t stable: they can compound over time. Each cycle’s output becomes the next cycle’s input, and small failures that the system can’t see get embedded in the next generation of work. A team is better at this not because individuals don’t make errors, but because their different perspectives mean errors get caught in different places at different times. AI-on-AI doesn’t have that property, so the ones that slip through are precisely the ones that keep slipping through.

Eventually any control system needs anchoring in something outside itself, or it becomes a reinforcing loop with no external checks. In a cockpit, that anchoring is physical reality: the plane really does respond to physics, and physics is the ultimate referee. In a team, it’s the independence of the people. In a one-person plus AI system regulated by more AI, there isn’t anything providing that, which is what makes the failures both invisible and serious.

---

Going back to Altman’s “one person and 10,000 GPUs” idea. There’s no second pilot, no one who would catch what the founder missed, because the founder is the only one who could miss it. This can work when things are predictable, but it can fail spectacularly when things get complicated, and most things that actually matter happen in complicated situations.

This is the systemic cost of designing AI for individuals. The social erosions and the structural problem point in the same direction, and they reinforce each other. The five mechanisms in the first half of this post show what happens to the conditions for collective intelligence inside any one team. The cockpit problem shows why even adding more AI on top doesn’t recover them. Stacking correlated systems together doesn’t make them independent.

Which means the answer can’t be “work alone, but better” or “add more AI to compensate.” What it might be instead is the question for the next post in this series.