On Modeling Human Trust in Automation: Identifying distinct dynamics through clustering of Markovian models

Abstract Intelligent and autonomous technology is performing tasks from driving to healthcare, in which calibration of human trust in automation is critical. In this paper, we use a model-based clustering algorithm to identify and model dominant dynamics of human trust in automation among a general population. In doing so, we seek to balance the tradeoffs between a single generalized, or several individualized, models of human trust. We show that two models optimally represent the sampled population and denote the participants in the two clusters as “Followers” and “Preservers”—those with high and low propensities to trust automation, respectively. We compare the dynamics of each partially observable Markov decision process model as well their associated control policies that calibrate human trust while reducing workload in a reconnaissance mission context. The resulting control policies for varying the transparency of an automated decision aid suggest that the use of a generalized control policy would sub-optimally calibrate human trust dynamics in the general population.