What it is
Large language models now power autonomous agents that can form social networks, communicate and develop shared and diverging opinions on political issues, but their collective behaviour and its mechanisms are poorly understood. The authors simulated networked systems of thousands of LLM agents across different backbone models (GPT-3.5, GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3). In most networked simulations, including the intervention tests, 1,000 agents decided whether to talk to a connected agent or a random new one, exchanged LLM-generated messages and reported their opinions on a five-level left-to-right scale for partisan alignment, gun control and abortion ban. The opinion distributions moved away from their near-Gaussian start in fewer than 10 rounds of interaction and went on to polarize: the neutral share, initially the largest at 40%, fell to 22.5%, 0.4% and 5.1% for the three issues. The agents spontaneously formed networks with properties of human social networks, such as homophilic clustering, and their collective opinions evolved in ways consistent with phenomena discussed in studies of human behaviour, including the echo chamber effect and confirmation bias.
Why it matters
LLM agents left to talk among themselves polarized on their own and clustered with the like-minded, patterns consistent with studies of human opinion dynamics. The authors use the system as a synthetic testbed for interventions applied once it had polarized: every strategy reduced polarization significantly, but changing the network structure had limited effect, while individual-level strategies, instructing agents to be open to persuasion by diverse opinions or giving them neutral messages from newly introduced top influencers, gave the largest reductions, 11.8% and 8.8%.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underOpinion Dynamics and Social Influence, Social Media and Politics, Sentiment Analysis and Opinion Mining
