As AI agents begin to operate in populations rather than one at a time, new research suggests that the number of them changes what they collectively decide — amplifying a bias, inventing one from nothing, or flipping a group into the opposite of what each agent would choose alone.
By Dr Shamim Quadir (Senior Communications Officer), Published
New research published in Proceedings of the National Academy of Sciences (PNAS) suggests that when artificial intelligence (AI) agents interact in groups, their number is not merely a technical detail. It is a decisive factor in what the group settles on: populations built from the same AI model, doing the same task, can reach opposite outcomes for no other reason than that one group is bigger.
Human beings behave differently depending on how many of us are in the room. A family is not a small village. A village is not London. London is not a nation state. As scale grows, new rules, norms and pathologies can appear that were nowhere to be found at the scale below. The authors argue the same is true of AI.
The study, from City St George’s, University of London, the IT University of Copenhagen and the Universitat Politècnica de Catalunya, arrives at a time when AI agents are now being deployed working together rather than working alone. Multi-agent systems are already used in finance, energy, defence and social media, and researchers have begun modelling populations of millions, even billions, of interacting agents — what some now call AI societies.
Yet the industry’s AI alignment — it doing what humans intended it to do — and safety effort remains overwhelmingly focused on the single model. Benchmarks, red-teaming exercises — adversarial testing designed to expose a model's weaknesses — and safety evaluations almost always describe one agent responding on its own, and where groups are examined at all, they are examined at one fixed size.
“Physicists have a motto for this: more is different,” said Andrea Baronchelli, Professor of Complexity Science at City St George’s and senior author of the study.
To find out what changes with scale, the team used the “naming game”, a classic framework for studying how conventions emerge, in which randomly paired agents each pick a word from a shared pool and are rewarded when they happen to pick the same one. Agents see only their own recent interactions, never the wider population, and are never told they are in a group. Over many pairings, a population can converge spontaneously on a shared convention — the bottom-up way norms form in human cultures.
The team tr ialed these agent interactions using four large language models (LLMs) — Microsoft Phi-4, OpenAI GPT-4o, Qwen QwQ-32B and Meta Llama 3.1 70B Instruct — using word pairs that carry social meaning, such as {man, woman} or {straight, gay}, and scaling from two agents up to a million.
Interaction, they found, can pull a group away from what its members individually want in three ways. It can amplify an existing leaning until the group converges on it almost every time. It can induce a preference out of nothing, with populations of individually neutral agents reliably favouring one word over an equally viable alternative. And it can reverse a preference outright, so that a population settles on the word its own members disfavoured.
Which of the three occurs is partly a property of the AI model. For the pair {her, his}, Qwen and Phi populations converged on her while GPT and Llama populations converged on his — despite individual agents in all four cases starting from near-identical preferences.
Group size then determines how strongly these preferences bite, in ways that cannot be extrapolated. Larger populations became more predictable across every model and word pair tested, converging on one word until the outcome was effectively certain. But the size at which that tipping point arrived varied enormously: for some combinations as few as two agents, for others around ten thousand. Scale could also change the kind of distortion. For the pair {straight, gay}, Llama agents individually preferred straight — but populations reversed toward gay, and only once the group reached six agents or more. Below that, the effect was simply invisible.
The team also developed an analytical theory, borrowed from statistical physics, that predicts the behaviour of infinitely large populations and explains why the randomness of small groups gives way to near-certainty above a critical size.
“Bias was our test case, because it is measurable and it matters,” Dr Ariel Flint, first author of the study, added. “But there is no reason to think collusion, deception or cooperation are immune to size effects. Current testing practice may be missing risks that appear only at particular population sizes — not because anyone was careless, but because nobody thought to vary the number.”
The authors say that the implications of the study for the alignment of AI systems are direct. A model can be aligned when tested on its own and still produce outcomes nobody chose once it is deployed alongside copies of itself — and no amount of single-agent evaluation will reveal it.
“AI alignment is still largely being done as though each model lived alone in the world,” said Professor Baronchelli.
The authors are careful about the scope of the claim: the bias they measure is internal to the coordination task — a mismatch between what individual agents prefer and what the group settles on — rather than a departure from human values and intentions. The setting is deliberately minimal, stripped of real-world context, to isolate the effect of interaction itself. They consequently suggest that populations of mixed AI models, and agents embedded in realistic network structures, are the next steps for research.
The peer-reviewed study, ‘Group size effects and collective misalignment in LLM multi-agent systems,’ is published in Proceedings of the National Academy of Sciences.