Emergence of human-like polarization among large language model agents

TL;DR

The paper demonstrates that autonomous LLM agents, interacting in a network, spontaneously develop human-like polarization and self-organize social networks with homophily, echo chambers, and backfire dynamics. A self-regulation mechanism based on social learning reduces internal inconsistencies and balances polarization, while five targeted interventions—especially at the individual level—mitigate polarization more effectively than network restructuring. Across multiple LLMs, temperatures, and initial conditions, the polarization phenomenon remains robust on divisive political issues but not on neutral, fact-based topics. The work provides a scalable, ethically informed pre-experimental platform to study and design strategies for reducing polarization in human-LMM-interactive environments, with implications for safeguarding democratic deliberation and informing policy design.

Abstract

Rapid advances in large language models (LLMs) have not only empowered autonomous agents to generate social networks, communicate, and form shared and diverging opinions on political issues, but have also begun to play a growing role in shaping human political deliberation. Our understanding of their collective behaviours and underlying mechanisms remains incomplete, however, posing unexpected risks to human society. In this paper, we simulate a networked system involving thousands of large language model agents, discovering their social interactions, guided through LLM conversation, result in human-like polarization. We discover that these agents spontaneously develop their own social network with human-like properties, including homophilic clustering, but also shape their collective opinions through mechanisms observed in the real world, including the echo chamber effect. Similarities between humans and LLM agents -- encompassing behaviours, mechanisms, and emergent phenomena -- raise concerns about their capacity to amplify societal polarization, but also hold the potential to serve as a valuable testbed for identifying plausible strategies to mitigate polarization and its consequences.

Figures (48)

• Figure 1: Political polarization in a networked system of LLM agents.a, A networked system of LLM agents, where agents operate on three basic stages: (1) self-expression, (2) communication, and (3) opinion update. In the self-expression stage, agents are required to generate reasons supporting their opinions. In the communication stage, agents decide with whom and what to communicate. In the opinion update stage, agents update their opinions based on the messages received from their socially connected agents. b, Opinion dynamics of LLM agents on the political issues of partisan alignment, gun control, and abortion ban. c, Opinion distributions in the initial and final states. d, Proportion of left-leaning, neutral, and right-leaning camps in the final state, where the left-leaning camp consists of agents with left and moderate left opinions, the neutral camp includes those with neutral opinions and the right-leaning camp contains those with right and moderate right opinions.
• Figure 2: Human-like polarization emerges from self-regulated LLM agents.a, Evaluating the self-inconsistency of LLM agents through pairwise interaction-based experiments. b, Opinion transition probability in pairwise interaction-based experiments, where agents with right-leaning opinions occasionally switch to opposing opinions while those with left-leaning opinions do not. c, Performances of the self-regulation strategy across political issues, where the self-inconsistency problem is largely mitigated. d, Opinion dynamics of self-regulated LLM agents on the political issues. Human-like polarization emerges from free-form social interactions among self-regulated LLM agents.
• Figure 3: Mechanisms behind the emergence of human-like polarization among LLM agents.a, Changes in the proportion of homophilic interactions over time. Agents are increasingly likely to interact with those holding similar opinions. b, Evolution of social networks among LLM agents, where agents with similar opinions are more likely to interact with one another, exhibiting the tendency toward homophilic clustering. Each network visualization corresponds to the circled points in (a). c, The echo chamber effect, where radical homophilic interactions intensify agents' polarization level. d, The backfire effect, where interactions with agents holding opposing opinions can also increase polarization. In (c, d), bars represent the average and error bars represent the corresponding 95% confidence intervals (CIs). e-g, Effects of individual-level social mechanisms, including selective exposure, confirmation bias, and elite signaling. In (e-g), bars show average levels of polarization in the last five timesteps, and error bars show the corresponding 95% CIs. When the system consists of more agents with traits of (e) selective exposure or (f) confirmation bias, and (g) influencers adopt non-neutral opinions, the level of polarization increases.
• Figure 4: Intervention strategies for reducing polarization.a, Intervention experiments, where two types of strategies are applied to the original polarized system: (i) network interventions, which directly modify LLM agents' social network, and (ii) individual interventions, which adjust agents' traits and behaviours. b, c, Network intervention strategies of (b) random interaction, where agents randomly interact, and (c) moderate opposing interaction, where agents receive messages only from those with opposing moderate opinions. d-f, Individual intervention strategies of (d) no selective exposure, where agents tend to interact with those holding diverse opinions, (e) no confirmation bias, where agents are open-minded to diverse opinions, and (f) neutral elite signaling, where agents receive non-personalized neutral messages. In (b-f), the upper sub-figures show the comparison of polarization levels between the original and the intervened systems, while the lower sub-figures illustrate the comparison of proportions of homophilic interactions. Compared with network interventions, individual-level strategies with no confirmation bias and neutral elite signaling contribute to the greatest reduction in polarization.
• Figure 5: Comparison between empirical and simulated opinion distributions.a, Empirical results, where Twitter-Politics is based on Flamino et al. flamino2023political, Twitter-Gun Control and Twitter-Abortion are based on Cinelli et al. cinelli2021echo. b, Simulation results of self-regulated networked systems. Here we take a coarsen division of left-leaning, neutral, and right-leaning camps.