Can conversational AI influence confidence in pre-existing political beliefs?

Conversational artificial intelligence (AI) may influence political polarization by shaping the strength with which political attitudes are held. Generative AI systems are increasingly capable of conducting adaptive, human-like conversations, raising questions about their ability to strengthen or weaken belief certainty. We report results from a preregistered, two-wave randomized experiment (N = 1,968) testing whether GPT-4 can strengthen or weaken individuals' attitudes on salient, evolving political issues. Participants engaged in personalized, multi-turn conversations in which the AI was instructed to either strengthen or weaken their stated views, or to engage in non-political discussion. Across issues including housing, immigration, and economic policy, AI-mediated conversations consistently reduced attitude strength. Specifically, belief confidence, while showing no evidence of strengthening. These effects persisted for up to 21 days. Our findings indicate that generative AI can durably influence political attitudes, primarily by weakening rather than strengthening belief certainty. By examining both pro and counter-attitudinal persuasion among individuals with existing views in dynamic political contexts, this study provides a demanding test of AI’s persuasive capacity and clarifies its implications for political polarization and democratic discourse.