A workshop organised by the Halden Project and embedded in the PSAM Conference.

While challenges in the transparency and consistency of AI-generated material currently preclude it from use in safety-critical domains, investigations of LLM-assisted safety analysis indicate significant potential for supporting analysts in tasks such as scenario generation, error identification, and structured documentation. However, known limitations, including hallucination, inconsistent reasoning, and sensitivity to prompt formulation, require careful consideration of how these tools are integrated into analytical processes. This workshop aims to provide participants with both conceptual grounding and practical insights in approaches that may mitigate these limitations.

The workshop will begin with a brief technical introduction covering foundational technologies relevant to improving LLM reliability in analytical contexts. Topics include retrieval-augmented generation (RAG), knowledge graphs, and structured automation workflows that constrain and verify model outputs. The workshop with then include hands-on examples of Human Reliability Assessment (HRA) analyses using LLM tools. These exercises will allow participants to directly observe the effects of different mitigation strategies on output quality, consistency, and traceability.

Participants will benefit from exposure to current research approaches and practical techniques for working with generative AI in domains where accuracy and reproducibility are important. The workshop format encourages discussion of both technical solutions and methodological considerations for evaluating AI-assisted analyses.