The group examined “typical AI applications designed to detect anomalies in real-time operational data” and considered the potential benefits and risks of adopting this technology. Participants were international and included regulators, licensees, and engineers.
According to the report, the group recognized the potential of AI in improving safety margins, detecting deviations earlier and reducing operational costs. It also highlighted the need to address AI explainability challenges, ensure layered defense measures are maintained, and support robust data assurance.
Nuclear facilities already routinely collect large amounts of data for maintenance and operational purposes, so in some ways they are a natural fit for deploying AI tools. As data becomes more digital, nuclear power plants will become even more ready to deploy AI tools.
However, the report states that “AI is only as reliable as the information used to train it,” so standardization and quality assurance of data across the industry is important. Participants suggested the use of independent data verification, data traceability and auditing, and strict boundaries that prevent AI applications from expanding beyond the conditions for which they were trained.
This exercise showed that deploying AI tools in nuclear power plants raises many of the same concerns as in other industries. Workers fear being replaced, the “black box” nature of tools, and concerns about declining human skills. However, the nuclear industry also needs to consider how AI tools can change its approach to regulation and how they fit into the industry’s rigorous safety culture.
For example, when considering what it would be like to use AI to “justify plant behavior that allows operation near operating limits and conditions,” participants asked The need for diverse and redundant barriers has become even greater. Maximizing the explainability of AI was generally considered important for all use cases, but when it comes to high-stakes decisions, participants emphasized the importance of maintaining a defense-in-depth approach. This could include hybrid modeling, where the AI output is backed by a deterministic rule-based layer or physical model, or acts as secondary analysis to supplement traditional methods.
Participants confirmed that the nuclear sector has extensive practice and knowledge of human decision-making during operations, and that this expertise can be used as a basis for setting expectations for AI decision-making.
While access to real-time data and plant status was seen as an advantage for both operators and regulators, it was also highlighted as a potential challenge if it leads to changes in legal liability or regulatory expectations.
The report suggests the nuclear industry should take a proactive approach, recommending the creation of a group that can develop technical guidance on best practices for using AI in nuclear applications. Topics to consider include expectations regarding AI safety envelopes, explainability, uncertainty standards, and validation protocols.
Participants also recognized the need for greater overlap in nuclear and AI expertise. AI developers must have a deep understanding of nuclear operations, and nuclear workers must develop expertise in the AI tools employed.
The report concludes that while participants see significant potential value in using AI for real-time monitoring in nuclear operations, technical justification remains an ongoing challenge and that the explainability of AI alone is insufficient for applications that result in greater safety.
