Hazard analysis in the era of AI: Assessing the usefulness of ChatGPT4 in STPA hazard analysis

This paper investigates the usefulness of LLM, such as GPT4 from OpenAI, in applying the STPA hazard analysis in socio-technical systems. This was achieved by recreating the process of STPA on a verified analysis applied to the ROLFER search and rescue drone system. Firstly, the system was discussed in detail with the LLM, defining its purpose, goals, components, and operations. Then, the LLM was talked through the steps of STPA and prompted to generate UCAs, loss scenarios, and safety specifications. These outputs were compared with those from the human safety team that applied STPA to the same system. The team discovered that LLM can be beneficial in certain aspects, such as loss scenario and safety specification generation and helping the researchers to reach a better understanding of the system, but have drawbacks in others, such as the generation of UCAs and being used as a verification tool of the STPA results of the human analysts.