A pioneering study led by City St George´s, University of London and the IT University of Copenhagen has demonstrated that groups of artificial intelligence agents, specifically large language models, can independently develop shared social norms through repeated interaction—without human intervention or central coordination. The findings, published in Science Advances, challenge the prevailing view that artificial intelligence agents operate simply as isolated systems and highlight the increasing societal relevance of multi-agent artificial intelligence networks in digital environments.
The researchers employed a version of the ´naming game´, a classic experimental framework used in human sociolinguistics, to observe how groups of language models select and converge on shared linguistic conventions. In simulated experiments, clusters of up to 200 artificial intelligence agents, each powered by models such as Llama-2-70b-Chat, Llama-3-70B-Instruct, Llama-3.1-70BInstruct, and Claude-3.5-Sonnet, were randomly paired to choose names from a shared pool. Agents received rewards for choosing the same name as their partner and penalties otherwise, with only limited memory of previous encounters and no explicit instructions about group membership. Over time, stable social norms emerged spontaneously within the groups, resembling the formation of societal conventions among humans.
Strikingly, the study also discovered ´collective biases´—emergent properties in the group’s behavior that could not be traced back to any individual agent but stemmed from their interactions. Senior author Professor Andrea Baronchelli noted this as a key blind spot in current artificial intelligence safety frameworks, which tend to focus exclusively on single-model behavior. Furthermore, a small but committed subgroup of agents was able to tip the majority toward a new convention, mirroring the critical-mass tipping points observed in human social change.
The study’s implications extend to the design and governance of future artificial intelligence populations as they become more integrated into digital societies, from autonomous vehicles to participatory online platforms. The research team emphasizes the need to understand and monitor the collective dynamics of artificial intelligence agents, as their ability to self-organize and propagate biases could amplify risks to marginalized groups. This new line of inquiry opens avenues for developing more robust frameworks for artificial intelligence safety, ethics, and societal coexistence.
