The Trust Cost of AI Errors: Examining Cognitive and Affective Trust Dynamics in Human-AI Collaboration

Building on existing theories of trust, this study examines the dynamic evolution of cognitive and affective trust over time through a series of human-AI collaborative tasks in an experimental setting. The results from a time-series random-effects panel model confirmed the “easy-error hypothesis,” demonstrating that AI errors on easy tasks led to a greater reduction in cognitive trust compared to errors on difficult tasks. Interestingly, AI self-disclosure, such as expressions of encouragement, apologies, or self-doubt, did not significantly enhance affective trust. Furthermore, the findings revealed that cognitive trust significantly and positively predicted trusting behaviors over the course of the trials, whereas affective trust did not. Lastly, the study identified a decreasing trend in both cognitive and affective trust in AI over time. Theoretical and practical implications are discussed.