[4J1-GS-5] Agents:

[4J1-GS-5-02] The Herd Immunity Effect of Prebunking Methods for Preventing Misinformation: Verification through LLM Multi-Agent Simulations

〇Miyoshi Kazuki¹, Ito Takayuki¹ (1. Graduate School of Informatics, Kyoto University)

Keywords:Fake News, Multi Agent System, Prebunking

The risk of social decision-making being influenced by misinformation has grown, leading to increased attention on “prebunking,” a method that enhances individuals’ ability to evaluate misinformation by providing relevant information in advance. However, existing research on misinformation has not sufficiently examined the actual reduction of sharing behaviors or the propagation of misinformation through diverse channels beyond social networking services. In this study, we employ a multi-agent simulation using a LLM to investigate how prebunking affects both the act of sharing misinformation and the practice of pointing out others’ misinformation, which is essential for building herd immunity. The results indicate that prebunking significantly decreases the frequency of sharing misinformation, thereby suppressing its spread in discussions. Moreover, the propensity to flag misinformation shared by others increased notably only under a fact-based intervention, whereas a technique-based approach often lost its rhetorical cues during the sharing process, making it more difficult to detect misinformation.