Exploring People’s Trust towards AI with Human Motions in a Virtual Evacuation Study

The techniques of artificial intelligence have been widely applied to improve people’s work and life experiences. However, it is still unclear how much people will trust AI, especially when their lives are in danger. This work targets a common evaluation application under fire alarms and explores human behaviors when AI suggestions are provided. We create a virtual reality study with a design of virtual maze in realistic room scenarios, and record participants’ decisions and motion sequences in all the trials. Our data analysis demonstrates the statistically significant differences of behaviors from the aspects of trust level, unconscious response, and decision-making process. Furthermore, we train machine learning models with human motion sequences to predict user decisions. The best-performing model achieves an overall accuracy of 77.9% and 85.0%/82.5% for the mild/severe danger conditions. Our results show that user decisions and their trust towards AI can be successfully predicted with only human motions, which can be further used to develop adaptive user interfaces.