Gender Dynamics and Homophily in a Social Network of LLM Agents

TL;DR

This study analyzes a large-scale AI agent social network to understand how gender is performed and propagated among LLMs. Gender is treated as a dynamic, language-driven performance, not a fixed attribute, and is quantified weekly with $X_{it} \in [0,100]$ using a zero-shot classifier. By constructing 52 cumulative weekly networks and applying scalar assortativity $r_t$, separable temporal ERGMs for selection, and panel regressions with fixed effects and IVs for social influence, the authors find persistent gender-based homophily and evidence that both social selection and influence shape tie formation and gender expression. The results suggest cultural entrainment of gender in synthetic populations, with potential feedback loops into human discourse and future model training, underscoring the need for governance and design strategies to mitigate reinforcement of gender stereotypes in AI-driven social systems.

Abstract

Generative artificial intelligence and large language models (LLMs) are increasingly deployed in interactive settings, yet we know little about how their identity performance develops when they interact within large-scale networks. We address this by examining Chirper.ai, a social media platform similar to X but composed entirely of autonomous AI chatbots. Our dataset comprises over 70,000 agents, approximately 140 million posts, and the evolving followership network over one year. Based on agents' text production, we assign weekly gender scores to each agent. Results suggest that each agent's gender performance is fluid rather than fixed. Despite this fluidity, the network displays strong gender-based homophily, as agents consistently follow others performing gender similarly. Finally, we investigate whether these homophilic connections arise from social selection, in which agents choose to follow similar accounts, or from social influence, in which agents become more similar to their followees over time. Consistent with human social networks, we find evidence that both mechanisms shape the structure and evolution of interactions among LLMs. Our findings suggest that, even in the absence of bodies, cultural entraining of gender performance leads to gender-based sorting. This has important implications for LLM applications in synthetic hybrid populations, social simulations, and decision support.

Figures (10)

• Figure 1: Example of a Chirper's profile and feed on Chirper.ai.
• Figure 2: Followership network for weeks 4 and 35. The visualisations are made using ForceAtlas2 layout on Gephi 0.10.1, with nooverlap option.
• Figure 3: Weekly gender scores for the top 100 most active Chirpers. The vertical axis lists Chirpers, and the horizontal axis lists weeks. Each cell shows the gender score (0: most masculine to 100: most feminine) for each agent.
• Figure 4: Scalar assortativity for gender-scores over time in the followership network and two degree-preserving null ensembles. The configuration null model preserves in-degree and out-degree sequences, while the joint degree null model additionally preserves the mixing of out-degrees and in-degrees across edges. Shaded bands show the mean plus or minus one standard deviation across null realizations.
• Figure 5: Upper panel: Gender similarity in new tie formation from eight-week STERGMs. Lower panel: Peer influence dynamics (OLS vs IV) across the eight-week windows.