Human-Agent Interaction in Synthetic Social Networks: A Framework for Studying Online Polarization

Online social networks have dramatically altered the landscape of public discourse, creating both opportunities for enhanced civic participation and risks of deepening social divisions. Prevalent approaches to studying online polarization have been limited by a methodological disconnect: mathematical models excel at formal analysis but lack linguistic realism, while language model-based simulations capture natural discourse but often sacrifice analytical precision. This paper introduces an innovative computational framework that synthesizes these approaches by embedding formal opinion dynamics principles within LLM-based artificial agents, enabling both rigorous mathematical analysis and naturalistic social interactions. We validate our framework through comprehensive offline testing and experimental evaluation with 122 human participants engaging in a controlled social network environment. The results demonstrate our ability to systematically investigate polarization mechanisms while preserving ecological validity. Our findings reveal how polarized environments shape user perceptions and behavior: participants exposed to polarized discussions showed markedly increased sensitivity to emotional content and group affiliations, while perceiving reduced uncertainty in the agents' positions. By combining mathematical precision with natural language capabilities, our framework opens new avenues for investigating social media phenomena through controlled experimentation. This methodological advancement allows researchers to bridge the gap between theoretical models and empirical observations, offering unprecedented opportunities to study the causal mechanisms underlying online opinion dynamics.