Homophily-induced Emergence of Biased Structures in LLM-based Multi-Agent AI Systems

TL;DR

This work investigates how autonomous LLM-driven agents shape network topology through partial information and prompt-guided decisions. The authors introduce a growth framework starting from a fully connected seed and grow to $|V_f|=1500$ nodes, comparing Highest Degree, Preferential, Partial Preferential, and LLM-Assisted attachment across four LLMs. They reveal that introducing binary social attributes yields strong attribute assortativity and homophily, with political orientation and religious participation driving pronounced fragmentation and asymmetric heterophilous ties, while degree-based effects remain influential. The study highlights model-specific differences in bias strength and demonstrates potential structural inequalities emerging from directional tie formation, underscoring the need for ethical foresight and interventions in AI-driven network architectures. These findings advance understanding of how AI agents co-evolve socially biased networks and offer guidance for designing fairer, more cohesive AI-mediated digital ecosystems, while outlining avenues for formalizing attachment dynamics and validating against real-world data.

Abstract

This study examines how interactions among artificially intelligent (AI) agents, guided by large language models (LLMs), drive the evolution of collective network structures. We ask LLM-driven agents to grow a network by informing them about current link constellations. Our observations confirm that agents consistently apply a preferential attachment mechanism, favoring connections to nodes with higher degrees. We systematically solicited more than a million decisions from four different LLMs, including Gemini, ChatGPT, Llama, and Claude. When social attributes such as age, gender, religion, and political orientation are incorporated, the resulting networks exhibit heightened assortativity, leading to the formation of distinct homophilic communities. This significantly alters the network topology from what would be expected under a pure preferential attachment model alone. Political and religious attributes most significantly fragment the collective, fostering polarized subgroups, while age and gender yield more gradual structural shifts. Strikingly, LLMs also reveal asymmetric patterns in heterophilous ties, suggesting embedded directional biases reflective of societal norms. As autonomous AI agents increasingly shape the architecture of online systems, these findings contribute to how algorithmic choices of generative AI collectives not only reshape network topology, but offer critical insights into how AI-driven systems co-evolve and self-organize.

Figures (16)

• Figure 1: Flowchart of the node addition process in the LLM-driven evolving network model.
• Figure 2: Visualization of 4 networks, each illustrating different node attribute configurations. Node colors indicate attribute categories, while node sizes are proportional to degree. Each network comprises $n = 1500$ nodes, where each new node connects to $m = 2$ existing nodes selected from a random subsample of $s = 50$ nodes, based on recommendations generated by Gemini 1.5 Flash.
• Figure 3: Comparison of degree distributions across different attachment strategies in network growth for $n = 1500$ nodes and two levels of connectivity ($m = 2$ and $m = 5$) and subsample sizes of $s = 10$ and $s = 50$. The strategies include: (i) highest-degree attachment , (ii) preferential attachment, (iii) partial preferential attachment, and (iv) LLM-assisted attachment generated using the Gemini 1.5 Flash model. The figure illustrates how each strategy shapes the emergent degree distribution. Results are averaged over 30 realizations
• Figure 4: Average attribute assortativity ($r_{\text{attr}}$), the tendency of nodes to connect with others sharing similar attributes, in LLM-generated networks with $n = 350$, $m = 3$, $s = 50$: (a) per model across Claude 3.0 Haiku, GPT-4o Mini, Gemini 1.5 Flash, and Llama-4-Scout; (b) per attribute across political orientation, gender, age, education, ethnicity, religious practice, and socio-economic status. Error bars denote standard error of the mean (SEM) and indicate variability in $r_{\text{attr}}$ across attributes and models (right). A Welch's ANOVA indicated a highly significant main effect for model, $F(3, 101.65) = 7.14$, $p < .001$, and for category, $F(7, 71.75) = 3073.58$, $p < .001$.
• Figure 5: Attribute assortativity in LLM-generated networks with $n = 350$, $m = 3$, $s = 50$. This plot captures the extent to which nodes connect based on similarity in assigned node attributes. Error bars denote standard error of the mean (SEM)