Crash severity in automation-engaged vehicles: A comparative analysis across SAE levels

Driving automation is increasingly being deployed in real-world traffic environments, yet empirical evidence on how automation engagement shapes crash severity remains limited. Using crash data from 2024 and 2025, this study examines factors associated with crash severity under different levels of driving automation. The dataset includes 18,413 automation-engaged crashes, with automation levels ranging from Society of Automotive Engineers (SAE) Level 1 to Level 5. AutoGluon, an automated machine learning (AutoML) framework was employed to model crash severity separately for lower-level (SAE 1–2) and higher-level (SAE 3–5) driving automation, with model interpretation conducted using SHapley Additive exPlanations (SHAP) and permutation importance. Results indicate that harmful event type, collision configuration, and object struck are the dominant predictors of crash severity across automation levels, while roadway geometry, environmental conditions, and person-level demographics play secondary but meaningful roles. Comparisons across automation levels reveal differences in the relative importance and functional effects of roadway and contextual factors. In addition, a qualitative analysis of crash narratives involving fully automated vehicles identified recurring patterns related to maneuver execution failures, lane-keeping issues, and interactions with human-driven vehicles, despite the absence of injury outcomes in these cases. The results inform policy decisions on infrastructure readiness, operational design domain management, and safety oversight for automated driving systems.