The U.S. Nuclear Plant Fleet Is Leaning into AI
IEEE Spectrum Andrew Moseman
U.S. nuclear plants are testing AI for paperwork and searches, not control rooms. That matters because the industry is short on people and buried in documents.
Based on reporting by IEEE Spectrum, Andrew Moseman — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The U.S. nuclear industry, famous for moving carefully, is quietly letting AI into its day-to-day work. This year, nearly all 94 U.S. nuclear reactors were offered AI tools, and most accepted. The pitch is not that a machine should run a reactor. It’s that the machine might finally help humans keep up with the endless paperwork, maintenance records, and regulations that keep those reactors legal and alive.
Atomic Canyon’s NIVA, launched in August and built with nuclear industry groups, is now available across the U.S. reactor fleet after pilot testing at plants run by Constellation Energy. Nuclearn says its products are already in use with more than 65 U.S. partners. Microsoft and Nvidia, meanwhile, have teamed up on a project aimed at the whole nuclear plant lifecycle, from site permitting and design to construction and operations.
The real pressure point is labor. Jerrold Vincent, Nuclearn’s CFO and cofounder, says the industry is being squeezed by aging hardware and a shrinking workforce. Jobs stay open. Regulations still have to be met. And nuclear plants generate a mountain of documentation for everything from routine compliance to oddball problems, like a crack in the sidewalk or a fault in the reactor. Nuclearn’s software is built to search those archives faster and help staff find how a problem was handled before.
Atomic Canyon took a similar route. Rob Austin of the Electric Power Research Institute says the project began after Constellation chairman Joseph Dominguez asked industry groups to make their knowledge accessible like a large language model, but secure. Atomic Canyon then trained on 53 million pages of public Nuclear Regulatory Commission data and partnered with Oak Ridge National Laboratory to build nuclear-specific models that understand the field’s terminology. The point is not flashy autonomy. It’s retrieval: find the right form, the right rule, the right past fix.
Security still puts hard limits on all of this. Nuclear data is proprietary or protected by federal law, so Nuclearn keeps solutions at a single site and NIVA only searches databases that already require plant-level security clearances. Even the bolder pilot at EPRI, where AI is being tested on inspection data from welds, keeps a human reviewing every recommendation. No plant is handing AI the controls. Not yet, and probably not soon.
My take — AI-written commentary, not fact-checked reporting
This is the least sexy AI story in the world, which is exactly why it may be the most useful one. Nuclear plants do not need a chatbot with swagger; they need a very expensive librarian that never gets bored. The industry’s real problem is not robot overlords. It’s missing people, missing time, and an absurd amount of paperwork.
Read more about this at: IEEE Spectrum