Chatbot's confident tone shaped how people liked it, not how accurate they judged its errors

Where confidence varied little, more dominant participants rated the verbal chatbot more positively and were more willing to use it, while less dominant participants preferred the numerical one. Those differences disappeared when confidence varied a lot.

Follow-up analyses suggest that attributing greater dominance to a verbal chatbot triggered a similarity-attraction effect. That effect did not extend to how accurate users thought the chatbot's outputs were, so confidence cues did not alter judgments of the hallucinated answers.

The experiment manipulated three features of a chatbot's self-confidence: how high it was, how much it varied, and whether it was expressed in words or numbers. Participants were assigned to one condition each, and the study measured their evaluations of the chatbot and their willingness to use it.

The results depended on the user's own dominance. When the chatbot's confidence stayed steady, more dominant participants preferred the verbal version and less dominant participants preferred the numerical version. When confidence varied widely, neither pattern held.

Post-hoc analyses tied the verbal chatbot effect to the similarity-attraction heuristic, in which people like others who seem similar to themselves. This shaped impressions of the chatbot but not perceived accuracy of its hallucinated outputs.