Nuclear Power

U.S. Nuclear Plants Turn to A.I., Mostly for Paperwork

co-produced by AI12. October 2026, 14:00

Cooling towers of a nuclear power plant. © Lukáš Lehotský / Unsplash

A.I. has found its first big job in America’s nuclear power plants, and it is surprisingly mundane: searching files, sorting defect reports, drafting applications. Nearly all 94 reactors in the United States have now been offered A.I. tools, and most operators have taken them up, IEEE Spectrum reports. They are not letting the systems anywhere near plant controls, however.

What A.I. Actually Does at the Plant

Few industries produce as much documentation as nuclear power. Every repair, every anomaly and every change must be documented, justified and traceable for the Nuclear Regulatory Commission. That is where the new tools come in:

Searching the rulebook: The Nuclear Industry Virtual Assistant (NIVA) from the start-up Atomic Canyon has been available across the North American fleet since August. It was developed together with the industry organizations EPRI, INPO and NEI. NIVA answers questions in plain language based on N.R.C. guidance, industry standards and technical reports, and cites its sources for every answer.
Learning from past mistakes: A second assistant searches the industry’s operating experience, meaning reports on how other plants solved similar problems. If a pump has a problem, for instance, the A.I. summarizes maintenance guides and tens of thousands of incident reports, and engineers then check the sources.
Sorting defect reports: The start-up Nuclearn categorizes so-called condition reports, which flag anomalies and defects, and routes them to the right people. According to the investor SJF Ventures, about half of the U.S. fleet uses the system. In one case, 87 percent of more than 38,000 reports were processed automatically, saving more than 5,600 hours of engineering time.
Writing drafts: Assistants summarize long documents and produce first drafts of evaluations and work plans.
Supporting inspections: In a pilot project by the research institute EPRI, A.I. flags anomalies in weld inspection data. A human reviews every recommendation.
Preparing license applications: In a demonstration by the U.S. Energy Department with Microsoft and the start-up Everstar, A.I. turned a safety analysis for a new reactor type into 208 pages corresponding to sections of an N.R.C. license application. It took one day instead of the usual four to six weeks, and an expert reviewed the result.

What A.I. Is Not Allowed to Do

The tools have nothing to do with running the reactors. NIVA has no access to plant systems, writes nothing to controls and approves nothing. “You don’t want machines making decisions that could impact the operations,” said Trey Lauderdale, Atomic Canyon’s chief executive. Licensed operators and engineers retain their decision-making authority. Security rules are strict as well: Nuclearn keeps each plant’s data on site, and NIVA only searches databases that already require plant-level clearance.

Atomic Canyon trained the models behind NIVA together with Oak Ridge National Laboratory on the Frontier supercomputer, using more than 53 million pages of N.R.C. documents. The industry cites an aging workforce that will soon retire and take its knowledge with it, as well as the planned nuclear revival. The first small modular reactor in the United States recently got the green light, and Microsoft and Nvidia are also working on A.I. tools for the entire life cycle of power plants, from permitting to operations.

Criticism: Paperwork Is Safety Work

Critics consider the use of A.I. in licensing particularly risky. In a report for the AI Now Institute, the safety experts Heidy Khlaaf and Sofia Guerra warn that licensing documents are not an end in themselves. They are meant to demonstrate that those responsible understand and control a plant’s risks. Even the smallest errors in the licensing process could have catastrophic consequences, they argue, and language models still get factual questions wrong too often. People also tend to trust convincingly written A.I. text, and finding subtle errors in it could take longer than writing the documents themselves. On top of that come new attack surfaces, such as manipulated training data. Atomic Canyon itself urges caution: Experts should never take an A.I. system’s word for something, Mr. Lauderdale said.