Proceedings of the Annual Meeting of the Cognitive Science Society

Large language models (LLMs) enable tailoring of persuasive messages at scale, raising hopes and concerns that persuasion may become a problem of optimizing message content. However, people can resist persuasion for qualitatively different reasons, including doubts about the persuader's motives. Here, we reanalyze a public dataset of multi-turn human-LLM persuasion dialogues to examine whether such resistance can be detected early from recipients' own replies. We use an LLM as a proxy forward model of human response: we specify different stances in the system prompt and compute how likely the model is to generate the observed rebuttal. We find that skepticism about the sender's incentives, where recipients attend to possible incentive misalignment and discount message content, is common. Such skepticism predicts lower subsequent persuasion success and a higher risk of backfire. These findings highlight limits of content optimization and suggest the methodological potential of seeing LLMs as superpositions of human cognition.