Too Much is as Bad as Too Little: The Nonlinear Impact of AI Risk Perception on Trust in AI

This study goes beyond the dominant linear assumption, developing a dual-path theoretical framework to explain how perceived AI risk shapes trust in AI through cognitive assessment and emotional response pathways. Using threshold models and multiple linear regression, we investigated the direction, form, and heterogeneity of this relationship. The results reveal a significant non-linear association (inverted U-shaped pattern) between perceived AI risk and trust in AI; moderate levels of perceived risk enhance trust, while excessively high risks weaken it. Furthermore, this non-linear relationship exhibits systematic differences across different population groups. Younger individuals and men show higher risk tolerance thresholds, resulting in a clearer inverted U-shaped pattern, while middle-aged, older, and women are more sensitive to rising risks. These findings suggest that perceived AI risk does not always undermine trust, but can promote its formation under certain conditions. This study highlights the constructive role of moderate risk perception and its implications for trustworthy AI development.