| As machines' cognitive capacities approach those of humans, one may consider human-automation interactions in light of social psychology concepts such as trust. Past work shows trust is affected by past performance, or perceived performance, of the automation, and by user parameters such as general biases about machines. In turn, trust combines with self-confidence to affect decision to use the automation; whichever is trusted more, the self or the machine, is more likely to be used. Automation performance can depend on the situation, and successful human-automation interaction may depend on training the human operator to recognize the factors impacting automation's reliability.;Air traffic controllers were required to detect conflicting (closely approaching) aircraft pairs in a computer simulation of their workstation, assisted by an automated tool listing pairs of aircraft that would conflict. The tool was set to be less reliable in trials where certain aircraft deviated from their flight plans. One group of subjects was reliability-cued, i.e., told that the automation's reliability would suffer in deviating scenarios; the others were not cued. Subjectively-rated trust was lower in deviating scenarios for the cued participants but the difference did not exist for the non-trained participants. The cued group was more likely to unquestioningly accept the automation's judgments, and more likely to detect both real conflicts (more correct detections) and perceived conflicts (more false alarms). Across all participants, higher subjective trust predicted propensity to be misled by the automation into missing conflicts. After several hours of experience, participants settled into a pattern either of trusting the tool at least as much as themselves, or vice versa, and adopting a corresponding usage pattern.;Automation trust predicts usage and performance, especially ability to detect critical events missed by the machine. Training for variable-reliability automation may therefore have to depend on the person's pre-existing attitudes about automation, and on whatever attitudes develop as they gain experience. Also, users may exhibit discrepant attitudes about the likelihood and importance of machine missed detections vs. machine false alarms. |
