Can large language models automate the HAZOP process without human intervention?

This peer-reviewed study evaluates whether four leading LLMs (GPT-4o, GPT-4o-mini, LLaMA, and Gemini) can autonomously perform a complete HAZOP analysis from a P&ID without human intervention. Each model generated full HAZOP worksheets that were benchmarked against an expert-prepared reference. Results showed that only 19–37% of AI-generated scenarios were technically valid, and safeguards suggested were predominantly procedural rather than engineering-based. The authors conclude that LLMs can serve as a useful starting point for HAZOP but cannot replace expert-led studies. This paper is directly relevant to PSM practitioners evaluating AI augmentation of PHA workflows, providing empirical evidence of both the promise and current limitations of LLM-assisted HAZOP, a cornerstone of RBPS Element 7 (Hazard Identification & Risk Analysis).