Pilots and Pixels: A Comparative Analysis of Machine Learning Error Effects on Aviation Decision Making

Despite immense improvements in machine learning (ML)-based decision support systems (DSSs), these systems are still prone to errors. For use in high-risk environments such as aviation it is critical, to find out what costs the different types of ML error cause for decision makers. Thus, we provide pilots holding a valid flight license with explainable and non-explainable ML-based DSSs that output different types of ML errors while supporting the visual detection of other aircraft in the vicinity in 222 recorded scenes of flight simulations. The study reveals that both false positives (FPs) and false negatives (FNs) detrimentally affect pilot trust and performance, with a more pronounced effect observed for FNs. While explainable ML output design mitigates some negative effects, it significantly increases the mental workload for pilots when dealing with FPs. These findings inform the development of ML-based DSSs aligned with Error Management Theory to enhance applications in high-stakes environments.