Abstract
Previous research has shown that advice sources influence individuals’ risk perceptions and health decision-making. We conducted two experiments to examine differences in health risk assessment between AI algorithms and human peer groups, and how these assessments influence individuals’ judgments of behavioral health risks. In Experiment 1, 60 participants (gender-balanced) and 30 GPT-4 samples (from independent runs with varying temperature settings) rated the perceived risk of 60 health behaviors. The results revealed that AI systematically overestimated health risks by inflating outcome severity rather than risk probability. In Experiment 2, 60 participants compared higher- or lower-threat health behaviors to judge which posed lower risk, then revised judgments after receiving advice from AI or human peer groups. The results indicated that participants preferred human advice over AI in the lower-threat condition. However, this preference disappeared in the higher-threat condition, and participants accepting AI-disagreeing advice showed greater belief updates than those following human advice. Collectively, these findings highlight how the threat context influences human-AI advice integration, offering insights for the design of effective AI-based health decision support systems.