Investigating the Potential of Generative AI for Documentation Management

Sammanfattning

Safety documentation in the nuclear industry is essential but labour-intensive, with long lead times, strict regulatory demands, and a high risk of inconsistencies as facilities pursue digitalization and automation. This thesis investigates whether Generative Artificial Intelligence (GAI) can assist in authoring safety documentation for a regulated Swedish nuclear facility without compromising information security, traceability, or compliance. The work is framed as a conceptual proof of concept and structured using the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework, covering business and data understanding, data preparation, modelling, and process evaluation. The study combines a literature and benchmark review with interviews with subject matter experts and a detailed analysis of existing Safety Analysis Reports (SARs) to map available input sources to their target sections and establish a data backbone for automation. A technical architecture is proposed based on an on-premises, open-source large language model fine-tuned with Low-Rank Adaptation (LoRA) adapters and supported by a retrieval-augme-nted generation pipeline over vetted internal documents. Evaluation is planned along two dimensions: text quality (readability, tone, structure, and factual completeness) and process impact, modelled through scenarios for authoring, revision, and audit rework. The analysis indicates that the proposed integration is technically feasible and can substantially strengthen traceability, while identifying governance and hardware constraints as primary deployment bottlenecks. Overall, the thesis contributes a structured, domain-aware roadmap for introducing AI-assisted safety documentation in the nuclear sector and outlines prerequisites for a future pilot implementation.