How are we doing?: Presenting System Confidence Information to Support Trust Calibration and Adaptive Function Allocation

The “strong but silent” design of many decision support systems (DSS) has contributed to problems such as automation bias and trust miscalibration. The present study examined whether these difficulties can be overcome by providing continually updated information regarding a system's confidence in its own ability to perform its task(s) accurately and reliably. The application domain for this research was in-flight icing. Two groups of pilots flew a motion-based simulator in simulated icing conditions and were assisted by a neural net-based DSS that detected, and identified the location of, ice accretion. One group of pilots received information about overall system reliability only, whereas a second group was presented with a trend display of system confidence. Pilots in the latter group were better calibrated in terms of when to follow the system's advice. They were also more likely to reverse their actions when system-recommended actions were unsuccessful. Consequently, they experienced fewer icing-induced stalls. The findings from this research will be discussed in terms of their implications for the design of decision aids and for adaptive function allocation in human-machine teams.