Navigating Imperfect Automation: Automation's Impact on Operator Dependence Behaviors, Response Strategies, and Adaptations

Abstract

Automation has become an integral aspect of modern work environments, promising enhanced efficiency, safety, and accuracy across various domains. Despite this, automation is still imperfect, and the human operator is ultimately responsible for outcomes. Operators have been inappropriately using automation, which has resulted in documentation of various incidents and accidents. Researchers have extensively explored the influence of automation reliability on human dependence behaviors and collaborative performance in human-automation interactions. A limited body of research has explored the impact of automation errors on operator cross-checking behaviors or strategies. While existing trust research (i.e., attitudes toward automation) explores operator adaptations, there is a notable gap in the literature regarding how operators adjust their dependence behaviors and strategies. The projects in this dissertation contribute knowledge by (1) examining how operators’ dependence behaviors (i.e., compliance and reliance rates) and cross-checking rates are affected by automation performance; (2) evaluating how operators’ adapt their dependence behaviors, cross-checking rates, and response strategies to varying degrees of imperfect automation; (3) investigating a design intervention focusing on the incorporation of likelihood information, specifically, to compare the effects of predictive values with a frequency format, additionally, a baseline condition where no a priori information was examined. A meta-analysis was utilized to address aim 1, where we systematically extracted dependence and cross-checking behavior data. We found that the human operators not only varied their compliance and reliance behaviors to the automation, but also varied how often they used additional information to verify the automation’s recommendation. Human operators’ blind compliance (β1 = .74) and reliance (β1 = .89) rates increased as automation’s Positive Predictive Value (PPV) and Negative Predictive Value (NPV) increased. Alternatively, the operators were more likely to cross-check automation’s recommendation when automation performed worse. Operators’ cross-checking behaviors were marginally more sensitive (p = 0.08) to non-alarm errors (β1 = −.90) than alarm errors (β1 = −.52). To address aim 2, we utilized a dual-task laboratory experiment to evaluate how operators adapt their dependence behaviors, cross-checking rates, and response. Automation performance influenced dependence behaviors and response strategies. More specifically, operators adapted using trial-by-trial feedback during alarms and non-alarms; their behaviors and strategies were independently adapted to the automation’s PPV and NPV. We introduced a novel optimal decision-making strategy that considers operator access to alarm validity information. In the experiment, adjustments in behaviors converged towards the theoretical optimal behavior. However, more research is needed to empirically validate the proposed optimal strategy. Aim 3 was addressed through a data reanalysis and a human-subject study. Results indicated that automation performance influenced operator dependence behaviors, response strategies, and adaptations. More specifically, operators used trial-by-trial feedback to adjust to alarms and non-alarms; their behaviors and strategies were independently adapted to the automation’s PPV and NPV. Participants strategically changed their behaviors to improve performance; they accepted a loss in accuracy for to allocate more attentional resources to a compensatory tracking task. Participants with likelihood information made slightly faster behavioral adjustments than those without information. The findings of this dissertation enrich our understanding of how operators depend on, validate, or ignore automated systems during dual-task performance. The introduction of a theoretical optimal standard can serve as a benchmark, enabling operators to calibrate their dependence behaviors. Insights into how information affects changes to operator behaviors can facilitate an accelerated learning process for operators and support more effective solution implementation.