Influence Message Wargaming in the Metaverse: Towards LLM-Based Persuasion and Counter-Persuasion

The future metaverse will not only offer immersive environments, but also populate them with interactive agents—virtual characters that mirror real-world users or social groups. A key application use-case for such agents will be to test social persuasion techniques (for such purposes as marketing, advertisement, and calls to action) by first exploring their effects on simulated target populations in the metaverse environment. Inoculation techniques against undue persuasion can be studied using the same approach. Such applications call for designing believable agents that faithfully reproduce realistic and diverse social behaviors at scale, without requiring large observational data. In this work, we explore whether large language models (LLMs), trained on massive web-scale corpora that embed latent social patterns, can serve as social proxies—agents capable of simulating nuanced group behaviors in sociotechnical information systems with a focus on “wargaming” both persuasion and mitigation techniques. Our approach generates differentiated, socially plausible responses to persuasive content across different simulated population segments. We pose three central questions: (1) Can LLMs adapt their responses to persuasive content when conditioned on specific community profiles? (2) Can they predict impact of mitigation policies aimed at exposing undue persuasion attempts? (3) Do their predictions align with known social influence theories? While large-scale human validation is beyond the scope of this study, our experiments demonstrate that LLM-generated community-specific responses are both consistent and theory-congruent. These findings position LLMs as promising tools for building adaptive, socially grounded agents in next-generation digital ecosystems, advancing the realism and credibility of social simulations in the metaverse, and calling for research on technical and ethical implications.