Effect of AI Reliability Disruptions and Cueing Method on Trust and Performance in Visual Search

AI-based cueing in visual search can improve detection performance (Goh et al., 2005; Chavaillaz et al., 2018), but operators must calibrate their trust and adapt their behavior when system reliability changes (Lee & See, 2004). In applied domains such as aerial search and rescue, many complex factors (e.g., environmental conditions, sensor quality, target variability) prevent AI systems from maintaining constant reliability, making appropriate trust calibration and sustained task performance persistent challenges. Previous research demonstrated that AI cueing method affects this calibration: bounding boxes enhanced sensitivity, while contextual blur produced higher compliance and better skill retention after AI removal (Gruner et al., 2026, submitted). However, it remains unclear how these differential interaction patterns affect trust and performance when AI reliability degrades or AI support is temporarily unavailable. We investigate how the cueing method moderates operator responses to two types of AI disruption in a visual search task: a temporary reliability drop (85% to 70% to 85%) versus a complete system breakdown (85% to no AI to 85%). The reliability drop reduces AI to the boundary of usefulness (Wickens & Dixon, 2007), while the breakdown removes AI support entirely. These disruption types may produce qualitatively different trust trajectories: reliability changes affect trust calibration over time (Wiegmann et al., 2001), and the nature of the disruption (degraded vs. absent) may determine whether trust damage is lasting or reversible. Reliability drops provide negative experiences with incorrect AI recommendations, whereas breakdowns remove AI support without exposing operators to incorrect recommendations. Although removing automation incurs performance costs, it avoids the negative trust associations that arise from experiencing unreliable AI. However, automation removal creates out-of-the-loop effects that impair situation awareness and independent performance (Endsley & Kiris, 1995), suggesting that breakdowns are harmful through a different mechanism than drops. Trust recovery after reliability failures is typically incomplete and asymmetric (Yang et al., 2023; Rittenberg et al., 2024), but whether the type of disruption and the type of AI cueing produce different trust and performance trajectories remains an open question. Using aerial maritime imagery, we examine how different forms of AI disruption influence trust calibration, how the cueing method moderates responses to different types of disruption, and whether trust recovery trajectories differ as a function of both factors.