Just the facts: How dialogues with AI reduce conspiracy beliefs

Recent research shows that personalized dialogues with AI substantially reduce conspiracy beliefs. To identify the mechanisms driving this effect, we conducted two complementary experiments (total N = 2,348). Belief reduction effects proved robust across 7 of 8 variations, including when participants were told the AI aimed to change their minds or were asked to debate the AI; when the AI was prompted to offer factual information without otherwise seeking to persuade or to be concise in its exposition; and when the AI was restricted to a single static message (rather than a dialogue). The only condition that substantially undermined the effect was prompting the AI to persuade without presenting any facts. Content analyses of the AI's output paint a similar picture: When the AI relied more on facts and evidence, participants changed their minds more often. Furthermore, participants who were persuaded overwhelmingly cited the AI's factual arguments as the reason. These results suggest that AI-mediated dialogues are able to reduce conspiracy beliefs primarily because of their ability to supply belief-relevant facts and evidence that counter the conspiracy, rather than due to rhetorical, social, or source-based features of the interaction - consequently, humans might be similarly effective if they were able to produce the same quantity and quality of factual evidence as the AI.