A Research Agenda for Studying LLM-Powered Persuasion Technology to Mitigate Misinformation

Abstract

Emerging research on the use of large language models (LLMs) to counter false beliefs, including conspiracy theories, vaccine hesitancy, and climate denial, has relied predominantly on experimental methods rooted in social psychology, operationalizing persuasion as short-term attitudinal change measured through randomized controlled trials. We argue that this approach, while offering valuable causal inference, is insufficient for understanding how LLM-powered persuasion actually works. Drawing on communication theory and a qualitative analysis of publicly available transcripts from the “DebunkBot study” (Costello et al., 2024), we identify four constructs largely absent from existing research: users’ mental models of the LLM communicator, anthropomorphism (including negative forms such as attributions of naivety or gullibility that can drive resistance), folk theories about how AI systems operate, and epistemic literacy regarding machine-generated knowledge claims. We outline a research agenda organized around three priorities: methodological pluralism suited to the dynamic and personalized nature of LLM interactions, transparency and reproducibility standards for prompt design and transcript analysis, and ethical frameworks for assessing potential harms, including belief reinforcement among resistant participants and privacy risks in large-scale transcript data.