How Does AI Persuasion Depend on Prior Beliefs?

Conversational generative artificial intelligence (AI) is persuasive, and growing evidence attributes much of this persuasive power to the facts and evidence that AI conveys. The simplest version of this informational account implies that AI conversations produce approximately constant changes in latent belief strength across levels of baseline belief. We formalize and test this account with a cognitive response process model. Using this model, we examine persuasion effects in a two-wave experiment in which GPT-4 held personalized conversations with 1,968 Canadian voters, and in reanalyses of five publicly available datasets (75,238 observations from the United States and the United Kingdom). Examining treatment effects across pre-treatment belief levels, we find that the model accounts for about half to two-thirds of explainable variance and closely tracks observed effects at each level. Counter-attitudinal conversations consistently produce the most variable latent updates, followed by pro-attitudinal conversations, and then placebos. In most datasets, counter-attitudinal conversations also produce larger mean updates in latent belief. These results suggest that there are information-processing differences in pro-attitudinal and counter-attitudinal conversational interactions with AI.