Abstract
We propose a machine learning framework to predict human decision failures in AI-assisted systems using automated analysis of behavioral patterns. Ensemble classifiers trained on 127 engineered features achieve 89% accuracy (AUC = 0.89) in identifying accountability deficits across 2.3 million humans–AI interactions drawn from three studies: controlled laboratory experiments (n = 847), high-fidelity flight simulations (n = 134), and longitudinal operational observations at 29 aviation and ground sites over 60 months. The feature set operationalizes five behavioral markers—response latency profiles, confidence calibration deviations, override propensity patterns, situational awareness degradation metrics, and cognitive load redistribution indicators—capturing complementary aspects of human–AI interaction. The ensemble, combining Random Forest, Gradient Boosting and Logistic Regression, also supports early prediction (AUC = 0.76 using the first 10 decisions), enabling pre-emptive intervention before performance degrades. SHAP-based interpretability analysis indicates that situational awareness degradation contributes the largest share of predictive power, and the overall framework is applicable to real-time monitoring in AI advisory systems that require human supervision.