What is it about?
When AI agents interact, the group can settle on outcomes its members would individually disfavour. Interaction can amplify an individual bias, create one from neutrality, or even reverse individual preferences. We call this collective misalignment. Crucially, group size shapes both the strength and the form of these effects, and larger populations can become increasingly deterministic. As multi-agent AI systems are deployed in areas including finance, energy and social media, understanding how collective behaviour changes with scale becomes a practical necessity, not just a theoretical question.
Featured Image
Photo by Nick Fewings on Unsplash
Why is it important?
A model that looks well behaved on its own can still produce misaligned collective outcomes once it interacts with other agents. Our results show that group size is not a neutral parameter: testing one agent is not enough, and testing one group size is not enough either. A system that behaves one way with ten agents may behave differently with a hundred. Reliable evaluation and governance of multi-agent AI therefore need to test how behaviour changes with scale.
Perspectives
More is different: when many agents interact, the collective can behave differently from any of its parts. This work comes from exactly that complex-systems perspective. What I find most interesting is that we can already see collective misalignment and strong size effects in a deliberately simple coordination setting, where the mechanism is transparent enough to analyse mathematically. The next question is how broadly these effects extend to richer forms of interaction such as cooperation, competition, collusion or deception.
Andrea Baronchelli
City University
Read the Original
This page is a summary of: Group size effects and collective misalignment in LLM multi-agent systems, Proceedings of the National Academy of Sciences, August 2026, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2531697123.
You can read the full text:
Resources
Contributors
The following have contributed to this page
