AI Chatbot’s Self-Confidence Cues Shape Users’ Impressions of the Chatbot, but Not Accuracy Judgments of Hallucinated Responses

A 2 × 2 × 2 (confidence level × confidence variability × confidence modality) between-subjectsonline experiment (N = 738) examined how an AI chatbot’s self-confidence influences users’ evaluations of the chatbot and its hallucinated outputs. When confidence variability was low, more dominant participants evaluated the verbal (vs. numerical) chatbot more positively and indicated greater willingness to use it, whereas less dominant participants favored the numerical chatbot. In the high-variability condition, these patterns disappeared. Post-hoc analyses show that attributing greater dominance to a verbal chatbot triggered the similarity-attraction heuristic, but it did not affect users’ perceived accuracy of the chatbot’s outputs.