Redazione RHC : 17 November 2025 16:50
In the technological field, the idea that the future development of generating systems depends inexorably on the expansion of nuclear capacity is taking root, pushing the sector’s players to identify strategies to speed up the construction of new reactors.
In this regard, neural networks have recently been integrated into the preparation of documentation required for launching new projects. However, research institutes warn of the potential negative repercussions associated with this approach.
According to a report by AI Now , an initiative by Microsoft and the American energy company Westinghouse aims to use models to expedite the licensing of nuclear power plants. The report emphasizes that algorithms cannot replace the multi-step process designed to minimize risks and prevent errors that lead to serious accidents.
AI Now emphasizes that licensing is not a simple set of forms, but a complex process that requires analysis, discussion of design solutions, and justification of the future facility’s reliability . The report’s authors emphasize that attempting to reduce this process to simply generating documents could result in some control steps being treated as a formality.
Microsoft proposes to train the model using materials archived by regulatory agencies and data on future plant sites. The algorithm will generate design documents, which will then be reviewed by company employees . Idaho National Laboratory is already using a similar approach, hoping to use a neural network to speed up the preparation of permit application materials. Lloyd’s Register is following a similar approach, and Westinghouse is promoting its own system called Bertha , promising to reduce documentation preparation time from months to minutes.
AI Now experts consider this scenario a threat to nuclear safety. They emphasize that even a small inaccuracy in the software version or equipment specifications can lead to incorrect conclusions about system behavior and trigger a chain reaction of errors. They cite the 1979 Three Mile Island accident as an example, in which a combination of malfunctions and misinterpretations of readings led to a partial core meltdown. The report’s authors believe that models are prone to subtle errors, so the risk of a similar scenario repeating itself increases when part of the authorization process is automated.
In addition to technical risks, the issue of nuclear confidentiality is also raised. According to AI Now, technology companies’ requests for access to extensive data sets demonstrate an interest in information that is typically not disclosed to the public. This information could facilitate the creation of dual-use resources, so access to it is strictly regulated. The report’s authors emphasize that transferring such data to cloud services poses additional threats.
The political context is particularly worrying. The US administration is promoting regulatory reform, seeking to reduce project review times. Representatives of the Nuclear Regulatory Commission have already warned Congress that decisions and staff reductions could compromise safety oversight capabilities. At the same time, the White House is linking nuclear infrastructure development to the need for high-density energy sources from AI, considering accelerated reactor construction a national security imperative.
The nuclear industry, however, has a different perspective. Some analysts believe that the careful use of algorithms to systematize large volumes of documents can improve efficiency and facilitate the work of regulators. However, even proponents of such technologies emphasize that automation cannot replace human oversight and that blindly relying on model conclusions poses a threat to the industry , where errors have irreversible consequences.
The authors of the AI Now report conclude with a warning: attempting to subject nuclear design to the logic of an AI race and reduce the level of oversight to speed up the process could undermine confidence in nuclear technology and simultaneously increase the risks of proliferation of the data needed to create dual-use resources.
Redazione