Networks of autonomous language-model agents can develop polarised social structures through repeated local interactions, without being instructed to form opposing groups.
Researchers constructed simulated societies containing 1,000 agents, each assigned an initial opinion and allowed to exchange persuasive messages with other agents. Participants could maintain or change their communication partners and revise their positions in response to received information. Over 40 interaction rounds, initially mixed networks increasingly separated into like-minded communities. The researchers subsequently tested interventions intended to reduce polarisation, including changing communication partners and modifying how agents responded to opposing viewpoints.
https://phys.org/news/2026-10-ai-agents-political-conversation-partners.html
