AI Generates Traceable Hazard Scenarios for Aviation Safety from ASRS Reports
A new arXiv preprint (2608.04697) introduces an AI-assisted method for generating candidate hazard scenarios in aviation operations, using NASA's Aviation Safety Reporting System (ASRS). The approach targets adverse outcomes and produces structured hypotheses with categorical factors and narrative scenarios, each scored for plausibility based on historical co-occurrence and traceability to similar ASRS reports. A hybrid variant conditions narrative generation on a structured hypothesis from evolutionary abduction, improving correctness and reducing variability. The study evaluates multiple large language models, focusing on operational hazard analysis that accounts for weather, ATC actions, airspace constraints, aircraft operations, and human factors. This work addresses the gap between functional hazard assessment at the aircraft-system level and operational-level safety analysis, offering a traceable, data-driven tool for safety analysts.
Key facts
- The paper is arXiv:2608.04697, announced as new.
- The method uses NASA's Aviation Safety Reporting System (ASRS) reports.
- It generates structured hypotheses and narrative scenarios for adverse outcomes.
- Each scenario includes a plausibility score from historical co-occurrence evidence.
- Traceability to the most similar held-out ASRS reports is provided.
- A hybrid variant uses evolutionary abduction to condition narrative generation.
- The hybrid approach improves correctness and reduces variability.
- Multiple large language models are evaluated.
- The focus is on operational hazard analysis, not functional hazard assessment.
- Factors include weather, ATC actions, airspace constraints, aircraft operations, and human factors.
Entities
Institutions
- NASA
- arXiv