AI Now Drafts Nuclear Paperwork for Plants and the NRC, and Microsoft Sits in Three Chairs

Nearly the entire U.S. fleet of 94 commercial reactors was offered AI tools this year, and most plants took them, IEEE Spectrum reported on Oct. 8. The tools search and draft the records a plant has to keep, from maintenance histories and incident reports to the filings that go to the Nuclear Regulatory Commission.

The biggest rollout is NIVA, an assistant from California startup Atomic Canyon built with the three industry groups that hold the fleet’s shared operating records (INPO, EPRI and the Nuclear Energy Institute). It went fleetwide on Aug. 18 after a six-month pilot that included Constellation Energy, which runs 21 reactors at 12 plants. Nvidia is one of the startup’s new investors. Rival Nuclearn says it works with more than 65 U.S. partners.

Nuclearn cofounder Jerrold Vincent told Spectrum why it moved this fast. The tech companies betting on AI also want nuclear plants to power their data centers, and the plants don’t have the staff to keep up with the paperwork regulation requires. “If you think of it as ChatGPT for nuclear,” he said, “it’s actually not a bad place to start.”

In July 2025, Idaho National Laboratory announced a Microsoft tool built on Azure AI to generate the engineering and safety analysis reports that go into construction permit and operating license applications. INL says the tool assembles documents for human verification and does not run the safety analyses itself.

In a February case study, INL ran the tool on OpenAI’s GPT-4o to draft a safety analysis chapter for a reactor test bed at the lab. Reviewers called it an acceptable first draft and estimated it would cut drafting effort by about 50%. They also found several conversations between the AI agents pasted into the final document. One reviewer wrote: “Probably shouldn’t use I statements in this document.”

The NRC’s own AI inventory lists SimplifAI, an in-house tool for finding and drafting text in licensing and oversight work, as deployed, next to Microsoft 365 Copilot. In June, NRC chief data officer Basia Sall told Nextgov/FCW that SimplifAI is built on Azure OpenAI, Microsoft’s cloud service for OpenAI’s models, and that one type of licensing review that used to take four years had just finished in nine months. She said the agency also let some industry players curate NRC public data for their own AI tools. “We receive a much better application than we have in the past,” she said. “We don’t have as many questions.”

The NRC’s environmental sign-off on restarting the Crane Clean Energy Center, the former Three Mile Island Unit 1, ran in the Federal Register on Sept. 25. Constellation cites a 20-year deal, signed in 2024, to sell the plant’s power to Microsoft as the reason to bring 835 megawatts back. The NRC still has to rule on an exemption and three license amendments, and a final restart decision is expected in 2027.

Microsoft shows up three times in that chain: the drafting tool on the applicant side, the cloud models under the regulator’s own tool, and the 20-year buyer waiting on Crane. All three roles are public.

The people building these tools say AI stays away from the controls. Atomic Canyon CEO Trey Lauderdale told Spectrum he can’t see AI making operating decisions “in the short or the long term.” The exposure sits in the file. An application drafted on Azure OpenAI and checked by a reviewer whose tool also runs on Azure OpenAI can carry the same blind spot through both ends, and two busy humans reading two fluent documents are the only thing in between.

The NRC’s review of NuScale’s small reactor design took more than 250,000 staff hours and about 2 million pages, according to an INL report citing NuScale’s own lessons-learned paper. At 2,000 working hours a year, that’s about 125 years of reviewer time for one design. Shrinking that number is how restarts like Crane and the reactors the data centers want get built this decade.

The NRC should require every applicant to disclose which sections of a filing were AI-drafted and on what model, and log the same for its own review tools. INL’s reviewers found agent chatter in their draft because they knew a machine had written it, and the next NRC reviewer deserves that same warning on page one of the application.

Sources

• IEEE Spectrum: AI Steps Into Nuclear Power Plants, and Workers Are Letting It In
• Atomic Canyon: NIVA launches fleetwide
• Idaho National Laboratory: INL collaborates with Microsoft to streamline nuclear licensing
• Idaho National Laboratory: Evaluation of AI-Enabled Digital Documented Safety Analysis, A Case Study (PDF)
• Idaho National Laboratory: Evaluation of AI-Enabled Digital Documented Safety Analysis (PDF)
• NRC: How We Are Using AI
• Nextgov/FCW: AI has helped to slash nuclear licensing review times, NRC official says
• Federal Register: Crane Clean Energy Center environmental assessment and finding of no significant impact