Generative Artificial Intelligence (GenAI) is rapidly finding its way into engineering processes. In safety engineering, however, the challenge is not simply generating plausible answers. The key question is whether AI can support safety-critical tasks in a way that remains transparent, reviewable, and trustworthy.
October 8, 2026
© Fraunhofer IKS/Andreas Jacob
At the Fraunhofer Institute for Cognitive Systems IKS, researchers have been investigating the question of GenAI in safety engineering as part of its work on trustworthy AI and safety-critical systems. The focus lies on hazard analysis and risk assessment, an essential but often time-consuming activity in the development of complex technical systems. As systems become increasingly interconnected and autonomous, the effort required for comprehensive hazard analyses continues to grow.
An important milestone in this work is the publication of the paper by Fraunhofer IKS’ researchers called “Enhancing Reliability of LLM-Supported Hazard Identification” at EDCC 2026. The paper explores how large language models can assist engineers during hazard identification while maintaining human oversight. Rather than relying on generic chatbot interactions, the research shows that structured workflows, domain-specific knowledge, and human review are essential for achieving reliable and useful results.
The research is not intended to replace safety engineers. Instead, the research team aims to provide tools that reduce manual effort and support experts in identifying relevant hazards more efficiently while keeping responsibility and final decision-making in human hands.
Today, the focus is increasingly shifting from research prototypes to industrial application. The Fraunhofer IKS researchers are currently evaluating AI-supported hazard identification approaches on real industrial use cases and larger hazard analyses. Initial results are encouraging. In first trials with an industry leader in its field, a single engineer supported by the tooling identified approximately 95 percent of the hazards that had previously been found through a full expert workshop. While such results do not eliminate the need for expert review, they highlight the potential of AI-assisted workflows to significantly improve efficiency in early analysis phases.
These findings reinforce the belief that generative AI can become a valuable assistant for safety engineers when embedded into controlled and reviewable engineering processes. At the same time, many questions remain open. How can these approaches be integrated into industrial workflows? How well do they scale to large real-world projects? And how can we ensure that AI-generated results remain understandable and auditable?
Further information on the publication is available in Fraunhofer Publica: "Enhancing Reliability of LLM-Supported Hazard Identification."
Readers interested in trying the technology themselves can explore an alternative version at the Fraunhofer IKS Playground.
The publication at EDCC marks an important step on this journey, but the work continues. Through ongoing industrial evaluations and research with real project data, Fraunhofer IKS researchers are further exploring how trustworthy generative AI can support the next generation of safety engineering.
