The DOE’s plan for AI in NRC licensing

The Department of Energy announced the completion of a proof-of-concept demonstration of the use of Everstar's AI tool to generate chapter 5 of an NRC license application from preliminary safety documents.

The 208-page document was created by the AI tool in approximately one day. According to the DOE, it would typically take a team of people between four and six weeks to complete this work.

According to spokesperson at the DOE's Office of Nuclear Energy, a subject-matter expert who reviewed the document described the results as "consistent with what would be expected of a Revision 0 document."

A pilot program for using approved documented safety analysis (DSA) documents from willing participants is planned as a future phase, which would move this capability closer to production-grade regulatory use. This would require extending the work to a secure operating environment, compared with the proof-of-concept project, which used open-source documents.

The project was a collaboration among the DOE, Everstar, Idaho and Argonne National Laboratories, and Microsoft.

AI accelerated: This project has emerged in the context of the Genesis Mission, which was launched in November 2025 as a national effort to accelerate the use of AI in scientific research.

A DOE-NE spokesperson said the office intends to use AI tools for draft authorship, shifting expert labor to refinement and review.

The document was created for the National Reactor Innovation Center's Generic High Temperature Gas Reactor (HTGR) at INL. The team used the INL site-level DSA, the Generic HTGR preliminary DSA (PDSA), and the Demonstration of Microreactor Experiments safety analysis-documents created by subject-matter experts-as input to the AI. It also had access to some standard reference corpus, such as the NRC Agencywide Documents Access and Management System database and publicly available federal regulations.

Everstar's AI tool, named Gordian, uses retrieval-augmented generation architecture that the generative model uses to produce a cited output, structured for NRC compliance. This approach reduces hallucinations by anchoring responses to source documents, which serve as the technical foundation.

A DOE-NE spokesperson said that experts also authored "skill modules," which are structured instruction sets that govern AI behavior at each step of the document generation process, enforcing cite-only discipline, defining regulatory mapping, specifying section-by-section requirements, and including self-check and quality assurance routines.

"Because PDSAs are inherently preliminary and incomplete documents, the completeness of the AI output was correspondingly limited, a direct reflection of the inputs, not a failure of the tool," said a DOE-NE spokesperson.

The spokesperson also said that Gordian's ability to identify missing, derived, or inconsistent information across source documents is a capability advantage, surfacing issues that might otherwise persist through multiple human review cycles.

The tool showed areas for refinement in correctly distinguishing between a parameter's safety limits and its operational ranges, a distinction that is technically and regulatorily significant in nuclear licensing, they said.

What's next: Several next steps were identified by DOE-NE:

A near-term priority is to develop a dedicated review tool to systematically document and resolve discrepancies.

The team is pursuing a formal NRC working group on AI-assisted licensing.

Document standards may evolve alongside AI tooling, ensuring source documents are written in a manner that AI can properly parse, source, and cite.

Longer term, the same modular architecture is envisioned for additional final safety analysis report chapters, PDSA generation, NQA-1 (nuclear quality assurance) compliance documentation, and fuel fabrication facility licensing.