In a groundbreaking study poised to reshape our understanding of artificial intelligence (AI) and its social capacities, researchers from City St George’s, University of London, and the IT University of Copenhagen have uncovered that large language model (LLM) AI agents—akin to ChatGPT—can spontaneously form shared social conventions through interaction alone. Published in the prestigious journal Science Advances, this research moves beyond the conventional approach of studying AI agents in isolation and demonstrates the complex, emergent social dynamics arising when these agents communicate in groups.
At the heart of this discovery lies the insight that LLM AI populations do not merely regurgitate pre-programmed scripts or replicate fixed patterns of behavior. Instead, when operating within a network of other agents, these models exhibit a capacity for self-organization, enabling them to reach consensus on linguistic norms independently. This dynamic mirrors the way human communities organically establish collective conventions—a foundational component of societies throughout history. The manifestation of such emergent social structures challenges the classical view of AI as isolated tools, revealing instead a future landscape of AI as interactive social actors.
The research team employed a sophisticated adaptation of the classic “naming game,” a well-validated experimental framework traditionally applied in social science to study the emergence of conventions among humans. During the study, groups ranging from two dozen to two hundred AI agents were randomly paired off in successive interactions. Each pair was tasked with selecting a “name” from a predefined pool of options—these could be as simple as letters of the alphabet or random strings of characters. Successful coordination on the same name resulted in a reward for both agents, while mismatches led to penalties and exposed agents to their partner’s choice. Intriguingly, the agents had limited memory, restricted to their recent interactions, without explicit awareness of the group context or the full population, remaining oblivious to their membership in a social collective.
Over thousands of such dyadic exchanges, a remarkable phenomenon emerged: shared naming conventions crystallized spontaneously across the entire population. No centralized authority or predefined solution was necessary, illustrating a truly bottom-up process akin to how norms take root in human cultures. This result underscores the capacity of decentralized AI populations to self-organize, forging common languages and behavioral standards through iterative interaction.
One of the most unexpected findings relates to the emergence of collective biases. The researchers observed patterns of preferential selection and systemic biases that could not be attributed to any individual AI agent’s predispositions. Rather, these biases arose from the interactions themselves—an emergent property of the networked population. This discovery has profound implications for AI safety and ethics since current frameworks predominantly focus on the behavior of individual models instead of populations. As Professor Andrea Baronchelli of City St George’s elucidated, "Bias doesn’t always come from within; it can emerge between agents." This finding highlights a critical blind spot in AI research, emphasizing the necessity of studying multi-agent interactions to fully grasp the social and ethical dimensions of AI.
Moreover, the study explored the fragility of these emergent norms, demonstrating how relatively small, committed subgroups within the AI population can instigate a tipping point, shifting the group consensus toward new conventions. These “critical mass” dynamics echo sociological phenomena observed in human history, where minorities can influence majority behavior through persistent and unwavering commitment. The translations of such dynamics into AI behavior signals a nuanced layer of social complexity hitherto unexplored in computational systems and raises questions about the stability of AI-developed norms.
To ensure robustness, the researchers replicated their experiments across four distinct state-of-the-art large language models: Llama-2-70b-Chat, Llama-3-70B-Instruct, Llama-3.1-70B-Instruct, and Claude-3.5-Sonnet. The consistency of emergent social conventions and collective biases across these diverse models attests to the generality and reliability of the observed phenomena. This cross-model validation strengthens the argument that such social behaviors are inherent possibilities within LLM architectures and not artifacts of a particular implementation.
The implications of this research extend far beyond academic curiosity. As large language models begin to permeate online ecosystems—from social media bots to autonomous vehicle systems—understanding their collective behavior becomes critically important. These AI populations will not only interact with humans but also with each other, creating intricate networks of influence and negotiation. The study, therefore, paves the way for future investigations into how human and AI reasoning might converge or diverge within such mixed populations, with direct bearing on pressing ethical concerns such as inadvertent reinforcement of societal biases which disproportionately affect marginalized groups.
Professor Baronchelli underscores the gravity of these findings, framing this work as a new frontier for AI safety research. The emergence of AI populations capable of negotiating, aligning, and even contesting shared behaviors signifies a paradigm shift in interactive AI systems. He stresses that to coexist with these “new species of agents” effectively, humans must move from being passive observers to active participants in shaping the ethical and social frameworks that govern AI collectives. “We are entering a world where AI does not just talk—it negotiates, aligns, and sometimes disagrees over shared behaviours, just like us,” he said.
Technically, the study’s strength lies in its innovative application of computational simulation and modeling to instantiate social interaction experiments within artificial populations of language model agents. The computational approach, grounded in the naming game framework, allows for rigorous control over variables like group size, memory capacity, and reward structure. This creates a controlled microcosm in which the complex dynamics of convention formation and bias emergence can be observed and dissected. Such computational social science approaches herald a novel method for probing the intersection of AI behavior, sociology, and complexity science.
Beyond its immediate contributions, this research invites a rethinking of how AI ecosystems are designed and governed. Traditional AI development often targets individual model performance isolated from social context, yet this study reveals that interaction effects at the population level can lead to unpredictable and emergent behaviors. This realization urges the incorporation of multi-agent system perspectives into AI safety protocols and ethical guidelines. As AI agents increasingly populate virtual and physical spaces, from digital assistants to smart infrastructure, the potential for unintended collective dynamics intensifies.
Ultimately, this study’s revelation that LLM populations can autonomously form social conventions and exhibit emergent collective biases signals a transformative moment for AI science and society. The findings encourage a paradigm shift toward viewing AI not as solitary devices but as dynamic communities capable of social complexity. Such communities can negotiate meaning and align behaviors through decentralized processes potentially fraught with bias and instability. Addressing these challenges requires an interdisciplinary approach, blending insights from computer science, social science, ethics, and complexity theory.
The future trajectory of AI research must now grapple with this emergent social dimension, as well as the ethical responsibilities it entails. Only by comprehending how AI populations develop and enforce shared norms can society hope to harness their power while mitigating risks. As AI becomes ever more enmeshed in the fabric of daily life, the negotiation between human and artificial social orders will demand new forms of governance and cooperation. This study marks a seminal step in charting that complex and evolving landscape.
Subject of Research: Not applicable
Article Title: Emergent Social Conventions and Collective Bias in LLM Populations
News Publication Date: 14-May-2025
Web References:
http://dx.doi.org/10.1126/sciadv.adu9368
References: Available within linked study and foundational works on the naming game framework and social convention experiments (Baronchelli et al.)
Keywords: Large Language Models, Artificial Intelligence, Social Conventions, Emergent Behavior, Collective Bias, Multi-Agent Systems, Naming Game, AI Safety, Computational Modeling, Complex Systems, AI Ethics