TEPCO Tests Generative AI to Check Nuclear Licensing Paperwork
Restarting a nuclear reactor in Japan takes an enormous amount of paper, and Tokyo Electric Power Company Holdings wants AI to carry some of the load. TEPCO, ITOCHU Techno-Solutions (CTC) and Stockmark said on October 5 that they have installed a proof-of-concept version of a generative AI system in TEPCO’s nuclear division to help staff draft and check the applications the utility files with government regulators. Evaluation of the system begins this month.
The scale of the task explains the interest. According to the companies, an application can run to about 42,000 pages, with supplementary explanatory material of around 27,000 pages more, and several departments cross-check figures and diagrams for consistency and omissions. Nikkei Asia reported that more than 100 TEPCO specialists spent months on roughly 70,000 pages of documents tied to one reactor restart, and that the new tool is expected to cut that workload by about 80 percent.
The system combines generative AI with optical character recognition (OCR) and image recognition, and uses retrieval-augmented generation (RAG) built on Stockmark’s SAT data-structuring platform. Past documents and diagrams are structured so the AI can search them, and the relevant passages are fed into its answers so reviewers can quickly check where a statement comes from. Because the material is confidential, the system runs on premises rather than in a public cloud. TEPCO defined the business requirements and runs the evaluation, CTC built the system and Stockmark supplied the platform and technical support.
What it means in practice
The partners are explicit that people stay in charge: humans make the final call on whether an AI flag is valid and whether a document needs to change. That line matters in a field where errors in safety filings carry serious consequences and where responsibility for each application has to rest with the engineers who prepare it.
The project also speaks to a wider labour problem. Japan’s nuclear sector is short of specialists, much like the workplaces in our report on firms deploying AI robots to ease the labour shortage. Power is also an AI story in its own right: JERA, the generator TEPCO co-owns with Chubu Electric, is a partner in the $15 billion Chiba AI infrastructure project backed by Apollo.
What to watch: the results of the proof of concept, whether the Nuclear Regulation Authority comments on AI-assisted filings, and whether the tool spreads to other TEPCO divisions and other power companies, as the three partners say they intend.