This paper introduces Co‑Hazard Analysis (CoHA), where a human analyst interacts with an LLM via a chat interface to elicit potential hazard causes for safety‑critical systems. The authors systematically evaluate LLM responses across increasingly complex system descriptions and emphasize the importance of how prompts are phrased, scoped, and iterated to obtain useful hazard suggestions. They find that LLMs can be moderately useful for simpler systems but require careful prompt design and expert filtering as complexity grows. For PSM, CoHA is directly analogous to AI‑assisted HAZOP/what‑if studies: it demonstrates how prompt templates, question styles, and interaction protocols can shape LLM contributions to hazard identification while preserving human responsibility.
