Diablo Canyon adopts generative AI tool for nuclear paperwork

PG&E has started using a generative-AI tool called Neutron at California’s Diablo Canyon nuclear plant to search through “millions of pages” of technical and regulatory documentation, according to Business Insider. The company says the ChatGPT-like system cuts the time required to locate procedures, incident logs, and other records that staff need to answer questions and prepare submissions. The rollout is being described as the first on-site use of a generative-AI tool at a U.S. nuclear power plant.

The immediate problem Neutron is built to solve is not reactor control but paperwork throughput. Nuclear operations are surrounded by a dense perimeter of NRC-mandated procedures, corrective-action records, work orders, training materials, and configuration documentation; in practice, much of the plant’s “work” is proving to regulators and internal auditors that work was done. When a tool promises to turn a document hunt from hours into minutes, it changes the bottleneck in the organization: less time spent searching means more time spent acting, but also more reliance on whatever the system chooses to surface.

That shift matters because generative AI does not behave like a deterministic index. A traditional search system returns a list of documents; a model can return an answer that sounds like a document, and the user may treat it as one. In a safety-critical setting, the difference is not philosophical—it is about traceability. If a maintenance decision is later challenged, the audit trail needs stable references: which revision of which procedure, which engineering basis document, which approved exception. A probabilistic assistant that summarizes and paraphrases can speed up human work while simultaneously making it harder to prove, after the fact, what the operator actually relied on.

It also changes incentives around compliance. If the internal “cost” of producing regulator-ready narratives falls, organizations can expand the volume of compliance activity without adding headcount. That can be good—better documentation hygiene, faster responses to safety questions—but it can also inflate the surface area of paper controls that look complete while remaining thinly tested in the real plant. Generative systems are especially attractive in environments where the output is text and the success metric is “did we answer the auditor’s question,” not “did the machine run.”

Diablo Canyon is scheduled to remain in operation beyond its originally planned retirement, which means its documentation base is both unusually large and unusually long-lived. Neutron’s value proposition is that it makes that accumulated institutional memory searchable at the speed of a chat window.

At Diablo Canyon, the first high-profile use of generative AI is not to run the reactor but to find the page number that explains how it is run.