AISA: AI Safety Assistant Framework for Continuous Improvement of Highway Construction

AI systems for safety planning need local, deterministic processing: The framework prioritizes "deterministic, local inferencing" rather than cloud-based solutions, recognizing that safety-critical decisions in highway construction require reliable, auditable systems that don't depend on external services or black-box decision-making.

Classification accuracy has real limits in safety applications: The system achieved 75% accuracy on incident classification, meaning roughly 1 in 4 cases were misclassified—a significant error rate when workers' safety depends on correctly identifying hazard types and past incidents.

AI quality metrics can fail on rare but critical cases: The quality scoring system worked well on routine incidents but "was distorted on out-of-distribution fatalities," highlighting a key responsible AI challenge: models trained on common scenarios may fail precisely when encountering the rare, severe incidents that matter most for safety planning.

Unstructured safety data represents untapped knowledge: The paper identifies that historical accident narratives are "difficult to consult at the point of planning," suggesting that better AI systems could help workers access relevant past incidents during daily safety planning—but only if the underlying data quality and classification are reliable.

Human oversight remains essential for safety applications: The framework was benchmarked against "majority-vote LLM ensemble" and human-labeled records, indicating that responsible deployment requires multiple verification layers rather than trusting AI recommendations alone.