A recent study from City St George's, University of London, and the IT University of Copenhagen reveals that AI agents can develop human-like social conventions when communicating in groups. The research challenges conventional thinking, which usually views large language models (LLMs) as solitary entities. Collaboratively, the models demonstrated a tendency to create shared naming conventions through interaction, emulating human social behavior. The findings suggest that AI can form collective behavior that transcends individual actions, marking a significant shift in understanding AI's social capabilities.
When we posed the question, 'can these models coordinate their behaviour by forming conventions, the building blocks of a society?', the answer was yes, and their collaborative behavior cannot be reduced to isolated actions.
This study goes against the mainstream approach in AI research, which typically views LLMs as solitary agents rather than as part of a collective that can exhibit social interaction and shared norms.
Collection
[
|
...
]