Same Actions, Different Endings: Does Outcome Severity Change How AI Models Apply Reason's Culpability Decision Tree to Aviation Flight Crews?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Just culture holds that people should be judged on what they knew and did at the time, not on how events turned out. AI models are increasingly used to help triage and analyse aviation safety reports. This study tested whether general-purpose AI models judge identical crew actions as more culpable when an occurrence ends badly than when it ends safely. 48 occurrences were drawn by a pre-specified random rule from NASA's Aviation Safety Reporting System (ASRS) for 2025. Each exists in two versions, word for word identical except for the outcome: the safe ending as reported, or a standard, plausible bad ending. Claude Sonnet 5.5, GPT-6.1 Sol and DeepSeek V4 Pro applied James Reason's (1997) culpability decision tree to each version, independently, under two prompt wordings, five times each: 3,240 judgements in all. This project holds the materials, data, code and paper. Pre-registration: https://osf.io/s8q24

Last synced 2026-10-04 02:30 UTC