Action Over Words: Predicting Human Trust in AI Partners Through Gameplay Behaviors
Inferring trust in human-AI teams from gameplay behavior rather than questionnaires. Participants collaborated with AI agents of varying competence in Overcooked-AI, and behavioral models matched the predictive accuracy of self-report without asking players a single question.
In the burgeoning field of human-AI interaction, trust emerges as a cornerstone because many think that it is critical to the effectiveness of collaboration and the acceptance of AI systems. Traditional methods of assessing trust have predominantly relied on self-reported measures, requiring participants to articulate their perceptions and attitudes through questionnaires. However, these explicit methods may not fully capture the nuanced dynamics of trust, especially in real-time and complex interaction environments. This paper introduces an innovative approach to evaluating trust in human-AI teams, pivoting from the conventional reliance on verbal or written feedback to analyzing gameplay behaviors as implicit indicators of trust levels. Utilizing the Overcooked-AI environment, our study explores how participants' interactions with AI agents of varying performance levels can reveal underlying trust mechanisms without a single query posed to the human players. This approach not only bypasses the efficiency challenges posed by repetitive and lengthy trust assessment methods, but also provides insights comparable to them. We highlight the potential of non-verbal cues and action patterns as reliable trust indicators by comparing the predictive accuracies of questionnaire-based models with those derived from gameplay behavior analysis. Furthermore, our findings suggest that these implicit measures can be integrated into adaptive systems and algorithms for real-time trust assessment.
Published in the 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), pages 563–568, Pasadena, California. Joint work with Matthew L. Bolton at the University of Virginia and Peter A. Beling at the Virginia Tech National Security Institute.
An open-access preprint is available on TechRxiv for anyone without IEEE access.