Answers are grounded in this paper and cite their sources.
Evaluation of how exposure and error type impact reliability estimates in automated systems, suggesting implications for user training.
Objective We evaluated the sensitivity and calibration of users when estimating the reliability of an automated decision support system as exposure to the system and the saliency of errors varied. Background Although users are typically sensitive when estimating the reliability of an automated system, it is not known how exposure to a system or error saliency influences the estimate. Viewpoints of the “leaky integrator” and “intuitive statistician” suggest competing outcomes regarding estimates over time. Methods Participants identified matching images of MegaBlocks with assistance from an automated system. Participants experienced three levels of reliability and three levels of exposure (number of trials). They provided estimates of reliability, trust, and confidence in their estimates for each system. In Experiment 2, the saliency of false alarms and misses was manipulated. In Experiment 3, the saliency of these error types was matched. Results Reliability estimates were sensitive to true reliability but were lower with more salient errors. Exposure had no impact on the estimates of reliability or trust. Trust and confidence in the estimate of reliability increased with true reliability. Conclusions Estimates of reliability are generally sensitive to the true reliability of a system, although they can be hindered by memory limitations and/or saliency effects. Estimates of reliability are made very quickly and do not tend to change with exposure. Applications Training users on a new system may not require extensive exposure, but early interactions with a system should accurately represent the amount and types of errors that the system will make.
No takes yet. Share an insight, caveat, or question.
Patton et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: