The most cautious industry in America has quietly adopted a chatbot. According to a recent report, almost all of the 94 operating nuclear reactors in the United States now have the ability to bring AI into their workflows in 2026, and most have started doing so. Nuclear power, where "move fast" is not a phrase anyone puts on a poster, turns out to have a very specific use for software that reads quickly.
What the reactors are actually getting
The headline suggests something dramatic, so it is worth saying plainly what this is not. Nobody is handing the control room to an algorithm. The tools in question are, in essence, specialised cousins of ChatGPT, trained on the vocabulary and records of the nuclear industry.
Two companies feature prominently:
- Atomic Canyon introduced NIVA in August, a virtual assistant for the nuclear sector built together with representatives of energy companies. It was piloted at plants run by Constellation Energy, and has since spread to reactors across the country.
- Nuclearn says it works with more than 65 American companies. They include traditional plant operators and NuScale Power, which develops small modular reactors.
Microsoft and Nvidia are also working on a project to apply AI across the whole life of a nuclear plant, from choosing a site and designing it, through construction and operation.
The job is mostly paperwork
A nuclear plant produces electricity, and also a remarkable volume of documents. Operators have to prove to regulators that they comply with the rules, record incidents, and describe what they did about them. Somewhere behind every kilowatt-hour is a filing cabinet, digital or otherwise, that would make a tax lawyer wince.
This is where the AI earns its keep. According to the report, the systems help in three ways:
- Search. They find information in enormous document archives faster than a person can.
- Reporting. They help prepare reports for government bodies.
- Precedent. They surface solutions to similar problems that came up before.
It is not glamorous, but it is a sensible place to start. A tool that reads a mountain of text and points to the relevant paragraph is useful, and its output can be checked by the person who asked for it.
Why now
Jerrold Vincent, chief financial officer and co-founder of Nuclearn, gives two reasons for the speed of adoption.
First, data centres want a great deal of electricity, which has renewed interest in nuclear generation. There is a certain symmetry here. The industry that needs the power is also selling the software that helps produce it.
Second, AI happens to fit several of the sector's hard problems, namely regulatory requirements, management and maintenance. All three are document-heavy, and document-heavy is what language models do best.
The knowledge gap
There is a quieter motive too. The pool of people with deep nuclear expertise is shrinking. As experienced staff leave, the industry risks losing part of what they know, and some of that knowledge was never written down in a convenient place.
To build NIVA, Atomic Canyon loaded 53 million pages of publicly available documents from the US Nuclear Regulatory Commission in 2025. The material taught the model the terminology, rules and quirks of how nuclear plants work. A model that has read 53 million pages of regulatory material is not a veteran engineer, but it is at least a patient colleague who has done the reading.
Humans stay in charge
The report is firm on one point. AI is not seen as a replacement for the people who decide how reactors are run. Trey Lauderdale, chief executive of Atomic Canyon, stresses that these decisions carry high risk and fall under strict regulation, so human involvement in running nuclear plants will remain necessary for the foreseeable future.
That is the reassuring part, and also the realistic one. In an industry where a wrong answer is expensive, an assistant that drafts, searches and suggests is acceptable, and one that decides is not.
What we do not know
The report is brief, and several things are left open. It does not say how "using AI" is measured across the 94 reactors, or how deeply the tools are embedded at any given plant. It gives no performance figures, such as hours saved or error rates. The claims about Microsoft and Nvidia describe a project's ambition rather than a delivered result.
So the fair reading is modest. Nuclear plants are adopting AI where the work is reading, searching and writing, and they say they are keeping people in the loop for everything that matters. For an industry famous for caution, that is about as bold as a reactor operator gets.