Automated diagnostic aids: The effects of aid reliability on users' trust and reliance

We examined the effects that different levels of, and changes in, automation reliability have on users' trust of automated diagnostics aids. Participants were presented with a series of testing trials in which they diagnosed the validity of a system failure using only information provided to them by an automated diagnostic aid. The initial reliability of the aid was either 60, 80 or 100% reliable. However, for participants initially provided with the 60%-reliable aid, the accuracy of the aid increased to 80% half way through testing, whereas for those initially provided the 100%-reliable aid, the aid's reliability was reduced to 80%. Aid accuracy remained at 80% throughout testing for participants in the 80%-reliability group. Both subjective measures (i.e. perceived reliability of the aids and subjective confidence ratings) and objective measures of performance (concurrence with the aid's diagnosis and decision times) were examined. Results indicated that users of automated diagnostic aids were sensitive to different levels of aid reliabilities, as well as to subsequent changes in initial aid reliabilities. However, objective performance measures were related to, but not perfectly calibrated with, subjective measures of confidence and reliability estimates. These findings highlight the need to distinguish between automation trust as a psychological construct that can be assessed only through subjective measures and automation reliance that can only be defined in terms of performance data. A conceptual framework for understanding the relationship between trust and reliance is presented.