Table of Contents
A new study reveals that artificial intelligence models demonstrate remarkably human-like tendencies in forming connections within social networks, prioritizing well-connected individuals, mutual friends, and shared interests – a finding that could have important implications for the future of AI and its role in society.
Researchers from PNAS Nexus, publishing their findings on December 19, 2025, discovered that large language models (LLMs) replicate core principles of human social behavior with startling accuracy. The study, lead by Papachristou and Yuan, varied the information provided to AI models regarding network nodes and the context – simulating both private friendship networks and professional company structures.
“The LLMs not only reproduce these principles, they do so at such a high level that their behavior is hardly distinguishable from ours,” one researcher stated.This suggests that the underlying mechanisms driving social connection might potentially be more universal than previously understood.
The Allure of Influence: Preferential Attachment
The research highlighted a strong preference for preferential attachment, where AI systems consistently favored connecting with individuals already possessing a high degree of connectivity – essentially, those with high “social status” within the network. “All models preferentially connected to such higher-order nodes,” the team reported, noting that newer models like GPT-4 mini exhibited this behavior more prominently than older versions like GPT 3.5. This mirrors human tendencies to seek out connections with influential figures.
Beyond status, the AI models also demonstrated triad formation, a common human pattern where individuals are more likely to connect with friends of friends than with complete strangers.This creates tightly-knit groups and strengthens network cohesion. “This means that people are more likely to form connections with friends of friends than with strangers,” the researchers explained. “All K-models are more likely to form connections to nodes with common acquaintances,” they added, observing this preference across various simulated contexts, including school, work, and private communities.
Birds of a Feather: the role of Homophily
the study further revealed that AI models exhibit social homophily – the tendency to connect with individuals who share similar characteristics.Researchers created an artificial network of 50 nodes, assigning random attributes like residence, hobbies, and favourite colors. The AI consistently prioritized connections based on these shared traits. Like humans, artificial intelligences prioritize contacts with like-minded people, followed by mutual friends and, as a third factor, the networking of the other person.
Implications for the Future of AI
While these findings offer valuable insights into the behavior of AI, they also raise crucial ethical considerations. The researchers caution that this “human side” to AI social behavior could inadvertently reinforce existing societal issues. “It can unintentionally reinforce social phenomena such as echo chambers, hierarchies or information silos,” the team warned.
However, they also emphasize the potential benefits. “Conversely, AI enables realistic simulations with which we can test countermeasures or help us develop fairer information systems.” This ability to model social dynamics could prove invaluable in designing AI systems that promote inclusivity and mitigate bias.
The results also expand our understanding of how AI might function in everyday social contexts, from AI assistants and agentic AI to chatbots designed as digital companions. As AI becomes increasingly integrated into our lives, understanding its inherent social biases will be crucial for ensuring responsible development and deployment.
(PNAS Nexus, 2025; doi: 10.1
Key improvements and explanations:
