Building Appropriate Trust in Artificial Intelligence: Examining a Dynamic Trust Calibration Model and an Intervention

This study conceptualizes trust in artificial intelligence as a calibrated response to the system’s performance, aimed at fostering appropriate levels of trust in AI. We refine the trust calibration model and propose four types of trusting behaviors: under trust, over-trust, validated trust, and discerning trust, and examine their antecedents, temporal trajectory, and the impact of performance feedback as an intervention strategy. Two experiments assessed the dynamic evolution of these trust behaviors across 24 human-AI collaborative tasks. Time-series random-effects modeling revealed that in both studies, over-trust increased with task difficulty, with individuals of higher cognitive ability showing a sharper rise in over-trust on more difficult tasks. In Study 1, AI’s chain-of-thought explanations effectively reduced under-trust and increased validated trust, though these effects did not hold in Study 2, where a performance feedback intervention was introduced. In Study 2, performance feedback helped participants reduce over-trust and fostered discerning trust over multiple trials.