OA

Large Language Model-driven Multi-Agent Simulation for Fake News Diffusion Under Different Network Structures

The spread of misinformation threatens societal trust and democratic processes, motivating extensive research on how misinformation diffuses through social networks in communication research. While simulation has long been a central tool for studying such processes, existing agent-based models typically rely on hand-crafted contagion rules that limit behavioral expressiveness. In this work, we propose a novel approach of using generative AI (GenAI) as a methodological tool for simulating complex dynamics in information ecosystems. Specifically, we introduce a large language model (LLM)-driven multi-agent simulation framework in which LLM-driven agents make forwarding decisions conditioned on psychological traits, while misinformation propagates through given network structures. Simulations conducted across network topologies indicate that LLM-driven agents generate diffusion patterns that are both internally coherent and sensitive to structural properties of the network. Moreover, they exhibit emergent behavioral phenomena that are not replicated by conventional rule-based models. To assess external plausibility, we further conduct simulations on an empirical community derived from the Higgs dataset and show that the resulting diffusion patterns exhibit key qualitative regularities observed in real-world rumor propagation. Finally, we evaluate several intervention strategies and find that their effectiveness varies across network structures. Taken together, our results suggest that LLMs can serve as a flexible and expressive simulation component for studying information diffusion, enabling network-aware and behaviorally grounded analysis of misinformation dynamics beyond traditional modeling approaches.