How do humans infer silent reliability degradation in automated decision aids? A computational modelling study of first failure reports

Automated decision aids may undergo silent reliability degradation, in which performance declines without explicit warnings, confidence cues, or reliability-state feedback. In such situations, humans must infer whether automation has degraded from sequences of correct and incorrect outputs. This study examined the computational mechanisms underlying first failure reports in an industrial visual inspection task. We conceptualized silent-failure perception as a training-baseline-referenced first-report problem and compared five candidate mechanisms within a shared discrete-time hazard framework: Bayesian reliability learning, Delta-rule learning, CUSUM cumulative-deviation inference, recent-window heuristic monitoring, and a no-learning/reporting-tendency baseline. Trial-level data from 60 participants were analyzed across three levels of failure severity and two levels of initial reliability, yielding 720 participant-scenario observations with first-report times or right-censored outcomes. Participant-grouped 10-fold cross-validation showed that the Delta-rule learning model received the strongest predictive support among the five candidate mechanisms for participants whose data were entirely excluded from model estimation. This model-relative result is consistent with a baseline-referenced current-state tracking account: trial-level outcomes recursively update an estimate of the automation's current error state, and first-report probability increases when that estimate shows sufficient deviation from the training baseline. Failure severity primarily shaped the density and persistence of degradation evidence, whereas initial reliability shaped the experiential baseline against which later errors were interpreted. These findings provide a mechanistic account of silent-failure perception and inform the design of reliability-transparent automated systems.