A new model for calculating human trust behavior during human-AI collaboration in multiple decision-making tasks: A Bayesian approach

Abstract

The advancement of Artificial Intelligence (AI) technology has made human-AI collaboration increasingly common. Trust is a decisive factor influencing the quality of such collaboration, as uncalibrated trust may lead to task failure or even catastrophic consequences, significantly jeopardizing the safety of human–machine systems. Therefore, this paper proposes a Bayesian model for predicting human trust behavior towards AI based on human self-confidence and confidence in AI. Grounding in human cognition processes, the model simultaneously considers task difficulty and AI ability. Specifically designed within the context of multiple decision-making tasks with AI assistance, we introduce a task called Multi-Ball Motion (MBM), where participants collaborate with AIs of varying abilities to complete tasks under different levels of difficulty. We report experimental results involving 21 participants, demonstrating that our model effectively explains both the behavioral and subjective data of participants. It captures the dynamic changes in participants’ two types of confidence during the experiment and personalized predictions of their trust behavior, achieving an average prediction accuracy of 97.6%. Furthermore, the model adeptly elucidates the cognition processes underlying participants’ trust behavior formation. This work lays a solid foundation for trust calibration and risk analysis of human-AI systems.