Classifying crash severity by SAE automation level using transformer models: Insights for automated vehicle policy

The rise of automated vehicle (AV) technologies has emphasized the need to understand how crash severity varies across different SAE automation levels. This study analyzes 3891 crash narratives from the Texas Department of Transportation's (TxDOT) Crash Records Information System (CRIS) 2024 using transformer models (BERT, RoBERTa, and DistilBERT) to classify injury versus non-injury crashes across SAE Levels 1 through 5. Results show that BERT performed better for Levels 1–2, while DistilBERT was more effective for Levels 3–5. For Level 1, lane changes and right turns were associated with non-injury crashes, while illegal U-turns and severe damage increased injuries. At Level 2, controlling speed and engaging automation reduced injury risks, but visibility challenges impaired sensor performance. For Level 3, emergency braking reduced severity, whereas driver inattention and loss of control significantly increased injuries. At Levels 4–5, sensor data quality was critical, with poor data or interference elevating crash risks, particularly during maneuvers such as lane changes or left turns. To enhance AV safety, this study emphasizes mandatory driver monitoring for L3 vehicles, regular sensor calibration for L2, and stricter standards for L4–5 sensors. These findings reinforce existing policy initiatives, including NHTSA's Standing General Order 2021-01 on crash reporting, the Infrastructure Investment and Jobs Act mandate for driver monitoring, and the emphasis on Operational Design Domain (ODD) disclosure in A Vision for Safety 2.0. By linking technical evidence with regulatory priorities, the study demonstrates how machine learning can inform evidence-based AV policy and deployment.