Enhancing Reliability of LLM-Supported Hazard Identification
Year: 2026
Type: conference-paper
Abstract: Safety-critical systems, such as industrial robots, require thorough safety engineering during their development, including hazard analysis and risk assessment (HARA). However, as applications and systems are becoming increasingly complex, the effort required to perform these analyses increases rapidly. Due to advances in generative artificial intelligence (GenAI) and large language models (LLMs), these tasks can be partially automated. A lot of attention is already being paid to using LLMs for automating general engineering tasks. However, applying AI to critical development tasks such as hazard analysis raises safety concerns. In this paper, we propose a human-in-the-loop workflow with an LLM-supported approach for hazard identification, targeting to reduce the associated residual risk. Our approach makes use of state-of-the-art techniques for prompt engineering and retrieval augmented generation, in order to improve the reliability and consistency of the hazard identification. For the evaluation, we use a simplified collaborative manufacturing use case to compare our solution to an off-the-shelf LLM and to the reference results of a hazard identification workshop with safety engineers. In this setting, the results indicate that our proposed approach can improve the quality and structure of the identified hazards. Further, in this use case, the results show reduced expert effort compared to the reference workshop, suggesting potential efficiency gains for the hazard identification.
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7201849792