An Empirical Study of Collective Behaviors and Social Dynamics in Large Language Model Agents

TL;DR

This study empirically analyzes a large-scale, memory-enabled LLM social ecosystem (Chirper.ai) to assess whether interactions among LLM agents mirror or amplify human social dynamics. Through network analyses, topic modeling, toxicity assessment, ideological classification, and a novel mitigation prompt (CoST), the authors reveal human-like homophily and social influence at the micro level but also reveal distinctive macro patterns in toxicity and polarization, along with a practical, zero-cost intervention that reduces toxic posting propensity by about 43%. The work provides a rich, longitudinal dataset and multiple analytical lenses (topic, emotion, stance, network topology) to understand LLM collective behavior and to guide safer design of LLM-driven social platforms. It highlights the potential and limitations of emergent LLM sociality and offers a scalable mitigation strategy with implications for AI governance and online discourse safety.

Abstract

Large Language Models (LLMs) increasingly mediate our social, cultural, and political interactions. While they can simulate some aspects of human behavior and decision-making, it is still underexplored whether repeated interactions with other agents amplify their biases or lead to exclusionary behaviors. To this end, we study Chirper.ai-an LLM-driven social media platform-analyzing 7M posts and interactions among 32K LLM agents over a year. We start with homophily and social influence among LLMs, learning that similar to humans', their social networks exhibit these fundamental phenomena. Next, we study the toxic language of LLMs, its linguistic features, and their interaction patterns, finding that LLMs show different structural patterns in toxic posting than humans. After studying the ideological leaning in LLMs posts, and the polarization in their community, we focus on how to prevent their potential harmful activities. We present a simple yet effective method, called Chain of Social Thought (CoST), that reminds LLM agents to avoid harmful posting.

Figures (22)

• Figure 1: Topic modeling of (Left) Chirpers backstories, and (Middle) Chirpers posts. (Right) Novel topics and the correlation between related topics in Chirpers backstories and posts.
• Figure 2: (Left) Average similarity of posts and backstories over time. (Right) Average similarity of neighbors over time.
• Figure 3: (Left) Fraction of overall toxicity by users at different toxicity levels. (Right) Average engagement on toxic vs. non-toxic posts for toxic Chirpers.
• Figure 4: Emotion analysis of all and toxic posts as well as posts around "humans".
• Figure 5: The distribution of leaning scores around topic "human".