A thousand LLM agents talked freely and split on their own: polarization grew out of social interaction, not model bias
Synopsis
Thousands of large-language-model agents freely formed relationships, tried to persuade each other and updated their opinions on political issues, and ended up spontaneously polarized, organizing a social network with homophilic clustering. Earlier work on LLM political leanings mostly treated the model as an isolated counterpart to a human and rarely examined collective interaction among agents. The system gives agents only minimal initialization and lets the LLM generate everything else; on partisan alignment, gun control and abortion ban, the neutral share fell from an initial 40% to 22.5%, 0.4% and 5.1%.
Interpretation
Thousands of agents driven only by large language models freely formed social ties, sent each other persuasive messages and updated political opinions without any pre-assumed rules, and opinion polarization emerged on its own. Earlier work on LLM political influence mainly treated the model as an isolated conversational counterpart to humans, studying how its text or dialogue affects people's views, rather than the collective behavior of agents interacting with each other. On partisan alignment, gun control and abortion ban, the neutral share, initially the largest at 40%, fell to 22.5%, 0.4% and 5.1%, and the opinion distribution departed from its initial near-Gaussian shape in fewer than 10 rounds of social interaction.
Each agent is given only three basic social capabilities: expressing reasons for its opinion, deciding whom to talk to and generating a message to persuade that partner, and updating its opinion from received messages. All three stages are carried out by the LLM through prompts, so the system relies on no traditional hard-coded parameters; prompts act as soft parameters shaping how agents perceive, reason and interact. The main experiments use ChatGPT (GPT-3.5) through the public OpenAI API at temperature 1 in a zero-shot setting; experiments with GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3 give comparable results.
Perspective
The demonstrated scope is a synthetic LLM-agent system: main results use ChatGPT and three political issues, with 1000 agents per networked experiment and 100 agents in individual-mechanism experiments; the authors also obtain consistent results on immigration restriction and flat-Earth topics and after varying the initial network structure and initial opinion distribution.
How quantitatively these patterns correspond to real human behavior still needs further validation; prompt-based design and simplified opinion representations reduce interpretability; the intervention results should be read as hypothesis-generating, and the authors call for complementary human experiments and more systematic calibration against fine-grained real-world data.
