Splitting Source Evaluation and Compliance: AI Avatars and Behavioral Intention in Health Communication

Artificial intelligence (AI) avatars are moving into preventive and mental-health communication faster than the research needed to evaluate them. A central assumption in this literature is that source credibility drives compliance, but it has rarely been tested against a transparently disclosed AI source across health domains of contrasting emotional intensity. This study asks whether the negative perceptions typically associated with AI sources (greater eeriness, lower trustworthiness, lower perceived accuracy) actually translate into reduced behavioral intention. We ran a 2 (source: human vs disclosed AI) × 2 (domain: mammography vs psychotherapy) between-subjects experiment with 338 women aged 40–80, recruited through Prolific. Stimuli presented an identical health recommendation attributed either to a human physician or to a hyper-realistic AI avatar explicitly identified as artificial; there was no concealment. Two-way ANOVAs confirmed that participants perceived the AI source as significantly more eerie, less trustworthy, and modestly less accurate. Yet the source had no effect at all on behavioral intention. Domain, by contrast, mattered considerably: mammography screening generated much higher compliance intention than psychotherapy. The dissociation between how the source was evaluated and how participants intended to act suggests that in health contexts the personal stakes carried by a message can override source- credibility judgments entirely. The findings challenge the classical credibility-to-compliance chain and indicate that disclosed AI avatars can function as effective channels for preventive health communication, without any need for artificial humanization. Limitations include self- reported intention and a female 40–80 sample; field experiments are needed to verify whether the paradox holds for actual behavior.