Nuclear energy regulators in the United Kingdom, the U.S. and Canada in September jointly released a paper that outlines guiding principles to consider when using artificial intelligence to ensure the safe and secure operation of nuclear facilities and other nuclear materials.
The principles discuss the need to clarify and address the challenges arising from these fast-developing technologies while encouraging the beneficial uses of AI.
The principles paper, "Considerations for Developing Artificial Intelligence Systems in Nuclear Applications,” was published by the Canadian Nuclear Safety Commission, the UK Office for Nuclear Regulation and the U.S. Nuclear Regulatory Commission.
AI could support outcomes that are not readily attainable via non-AI techniques. For example, AI could analyze a much larger volume of data than conventional approaches, potentially allowing better management of plant risks and improved efficiency, the paper notes.
“Similarly, sufficiently developed AI could accomplish tasks that have until recently only been possible using humans, reducing the need to enter hazardous areas, and potentially reducing error.“
Additionally, the ability to retrain AI to benefit from updated information “provides flexibility to rapidly learn from previous experience to improve over time. The unique capabilities of AI also make feasible activities that are not achievable by either conventional technologies or humans, creating opportunities to directly and indirectly improve safety, security, and efficiency.”
Even though each country is operating with a different regulatory framework, “it is nevertheless important to examine some foundational tenets all parties should consider when AI is used in nuclear applications.”
This document is split into several sections that the Canadian, UK, and US nuclear regulators consider may be important in managing the risks arising from the use of AI:
- Use of existing safety and security engineering principles
- Human and organizational factors
- AI architecture
- AI lifecycle management
- Documenting AI safety and security
“Standards for nuclear-specific AI design and substantiation do not currently provide sufficient assurance in high consequence applications or where societal harms might exist,” the paper said.
“The fast pace of AI development means it is unlikely that AI-specific consensus standards for the nuclear domain will be available to support regulatory activities within the near future. In the interim, existing nuclear-specific standards remain a starting point coupled with considering the unique attributes introduced by AI.”
