This post is also available in:
עברית (Hebrew)
In a sci-fi-esque discovery, a new study has revealed that large language models (LLMs) can spontaneously form social conventions when placed in group interactions, without any explicit programming to do so. The research, published in Science Advances under the title “Emergent Social Conventions and Collective Bias in LLM Populations”, suggests that these AI agents are capable of mimicking the way human communities organically establish norms.
Conducted by scientists from City St George’s, University of London, and the IT University of Copenhagen, the study explored what happens when multiple LLMs engage with each other in a series of structured, yet unsupervised, interactions. Unlike previous research that focused on single-agent behavior, this work examined collective dynamics, revealing that LLMs are capable of social self-organization.
Using a modified version of the “naming game”—a framework from human social science—the researchers tested groups of AI agents ranging from 24 to 200 individuals. In each round, a random pair of agents was asked to choose a name for an object from a common list. If both selected the same name, they were rewarded. If not, they were penalized and shown each other’s choice, according to TechXplore.
Importantly, the AI agents only had access to a short history of their own interactions and were unaware they were part of a larger population. Nevertheless, a group-wide naming convention often emerged over time—completely spontaneously and without any centralized coordination.
This behavior closely mirrors how social norms and linguistic conventions evolve in human societies. More intriguingly, researchers also discovered that collective biases could form within the group—biases not present in individual agents, but arising purely from group dynamics.
Further experiments revealed that these norms could be surprisingly unstable. A small cluster of agents introducing a new naming pattern could sway the entire system, demonstrating effects well known in sociology.
The study used several LLM architectures—Llama-2-70B-Chat, Llama-3-70B-Instruct, Llama-3.1-70B-Instruct, and Claude-3.5-Sonnet—highlighting the fact that this phenomenon occurs across different models.
These findings underscore a new frontier in AI research: understanding how populations of autonomous AI systems may evolve shared behaviors, independent of human oversight. As LLMs become increasingly embedded in platforms across the board, the ability to predict and manage these emergent behaviors will be vital for long-term safety and ethical governance.
