Gender Dynamics and Homophily in a Social Network of LLM Agents

The evolution of large language models (LLMs) like GPT-4 has revolutionized AI capabilities, enabling complex language understanding and generation. Prior research has examined individual decision-making, cooperation, norm formation, and bias reproduction within AI agents, revealing human-like social behaviors despite lack of consciousness. However, understanding how these agents interact at scale, especially regarding cultural phenomena like gender, remains limited. Recent studies suggest that language-based content and network structures can produce demographic homophily, but most treat gender as a fixed attribute. This work builds on these insights, emphasizing gender as performative, culturally entrained, and fluid, and investigates how language use influences social ties and collective behavior in AI communities.

The core challenge is to determine whether AI agents’ gendered behaviors are static or fluid, and how these behaviors influence network formation and evolution. Specifically, the study seeks to disentangle social selection—agents preferentially following similar others—from social influence—agents becoming more similar over time. Addressing this is crucial for understanding bias propagation, social cohesion, and the cultural entrainment process in AI systems. Existing static analyses lack the temporal and content-based granularity needed to capture these dynamics, necessitating a combination of content analysis, network modeling, and causal inference.

Key innovations include: 1) conceptualizing gender as performative and fluid rather than fixed; 2) employing GPT-4 for fine-grained, weekly content-based gender scoring; 3) integrating dynamic network models (STERGMs) with panel regressions to jointly analyze tie formation preferences and behavioral influence; 4) demonstrating the co-evolution of language-based gender performance and social ties in a large-scale AI community. These approaches surpass prior static demographic studies, offering a nuanced view of cultural entrainment and bias formation in virtual agents.

• �� Data collection: Longitudinal dataset from Chirper.ai, capturing 70,000+ agents, 140 million posts, and follow relations over 52 weeks.
• �� Content analysis: Merging weekly posts per agent, applying GPT-4 zero-shot classification to assign gender scores (0-100).
• �� Network construction: Weekly directed follow networks, analyzing structural features like density, degree distribution, clustering.
• �� Homophily measurement: Computing scalar assortativity coefficients for each week, comparing with null models.
• �� Tie formation analysis: Fitting formation-only STERGMs to assess gender-based preferences in new links.
• �� Social influence: Panel regression models estimating the impact of followees’ gender scores on agents’ score changes, using fixed effects and instrumental variables.
• �� Validation: Null network simulations and sensitivity tests to confirm robustness of findings.

The experiments analyze network evolution, gender score dynamics, and homophily over a year. Model parameters are estimated via maximum likelihood, with significance tested against null models. The study performs ablation tests to verify the roles of social selection and influence, revealing early-stage homophily driven by preference, later reinforced by imitation. Results show that agents’ gender scores fluctuate significantly, yet network ties consistently reflect gender-based preferences, confirming the dual influence mechanisms. The robustness across different content categories and time windows underscores the reliability of findings.

The network exhibits strong, persistent gender homophily (assortativity ~0.2-0.4), exceeding null expectations. Individual gender scores are highly fluid, with standard deviations around 15 points, yet overall trend slightly favors femininity. Tie formation favors similar gender scores early (odds ratio 0.825), with social influence strengthening convergence over time. These results demonstrate that both social selection and influence shape the evolution of gendered behaviors and network structures in AI communities, paralleling human social patterns.

Findings inform the design of fair AI systems by understanding bias entrainment mechanisms. They enable more accurate social simulations of digital communities, improve moderation strategies, and guide ethical AI deployment in hybrid human-AI environments. Long-term, insights could lead to interventions reducing bias propagation and fostering diversity in AI-driven social platforms.

The analysis is limited to English content, potentially missing cultural nuances. GPT-4 classification may carry biases affecting scores. Network models assume minimal external shocks; real-world social events could alter dynamics. Future work should incorporate multilingual data, multimodal content, and external variables to enhance generalizability and robustness.