Using Large Language Models (LLMs) to Complement Expert-Based Hazard Identification under Limited Operational Data for Remotely Piloted Ships

Remotely piloted ships are still under development, and therefore, real operational experience and data are limited. This limited real-world evidence makes it challenging to comprehensively explore potential hazards during the early stages of design. However, a substantial amount of collected data already exists through previous expert intuition, questionnaires and workshops with maritime pilots that provides a significant qualitative benchmark. Recently, synthetic data generation using Generative Artificial Intelligence (GenAI) has emerged to deal with problem of real-world data scarcity in various studies. This study investigates the potential of Large Language Models (LLMs) as a source of synthetic data to complement existing expert-based hazard identification processes. In this feasibility study, structured prompts are designed to extract possible hazards related to communication, remote control, and human-machine interaction. The LLM-generated hazard lists are then qualitatively compared with the hazards identified through expert workshops to highlight differences and potential synergies. Initial results indicate that LLM-generated outputs can complement existing human-derived insights by expanding the diversity of identified hazard scenarios. Nevertheless, careful consideration is needed when interpreting and integrating AI-generated content to ensure consistency, occurrence probability, feasibility, and reliability. The study aims to motivate discussion on how GenAI can responsibly contribute to structured hazard identification in the development of remote pilotage services.

Keywords: Remote Pilotage, Large Language Models, Generative AI, Hazard Identification, Expert Validation, Maritime Design.