Read how Canada's NWMO and international experts plan to navigate the 40-generation AI paradox while protecting critical nuclear infrastructure from threat actors.
Inside the Nuclear Industry’s Drive to Erect Digital Guardrails for the AI Revolution
VIENNA — It is being billed by international delegates as the structural alliance of atoms and algorithms. At the International Symposium on Artificial Intelligence and Nuclear Energy, hosted by the International Atomic Energy Agency (IAEA) at its headquarters in Vienna, global power brokers and technology innovators reached a unanimous conclusion: the artificial intelligence boom and the nuclear energy renaissance are irrevocably fused.
On Day 2 of the summit, held on Thursday, December 4, the morning plenary shifted to a far more sobering logistical inquiry during Panel 3: Nuclear Capacity Deployment Consistent with Safety, Security and Safeguards Objectives. For 90 minutes, a high-level council of global regulators, industrial vendors, and state-backed engineering executives took the stage to answer a critical industry question: how can the nuclear sector aggressively accelerate capacity deployment using machine learning without compromising the industry’s strict zero-fail safety mandates?
The consensus from the summit floor is that while the nuclear sector is eager to exploit automated data processing, it is moving with the extreme deliberation typical of a heavily regulated, procedure-driven culture. Leading that charge was Canada’s controversial Nuclear Waste Management Organization (NWMO), which mapped out how multi-decade civil engineering projects can survive the volatile, hyper-accelerated lifecycle of modern software.
Building the Digital Foundation
Opening the panel, moderator Lisa Berthelot of the IAEA Department of Nuclear Safety and Security set the parameters for the session. She noted that expanding global nuclear capacity requires modernizing activities entirely separate from physical construction sites—specifically regulatory frameworks, technological innovation, and international cooperation.
Delivering the opening keynote, Christina Van Drunen, senior director of operations at Canada’s NWMO, framed the dialogue not as an abstract technical exercise, but as an immediate operational necessity.
“I think of this panel as taking the promise of AI that we discussed yesterday and now beginning to build the foundation,” Van Drunen told international delegates. “Taking a grounded view of how to make best use of AI while considering the safety, security, and safeguard objectives that are fundamental to our license to operate in a nuclear industry.”
While preceding conference sessions focused heavily on uranium mining and front-end reactor deployment, Van Drunen anchored her presentation at the absolute conclusion of the fuel loop.
“So where yesterday Cameco talked about the beginning of the nuclear life cycle, what we are looking at is the conclusion, the resolution,” Van Drunen explained. “So, NWMO is tasked with implementing Canada’s plan for the long-term management of nuclear fuel. This is essential to responsibly manage the fuel from our existing reactor fleet and also to enable future expansion whether we’re talking large or small modular reactors.”
The 40-Generation AI Paradox
Canada’s approach to dealing with radioactive byproduct is dictated by a long-term strategy known as Adaptive Phased Management, a framework that balances a definitive technical resolution with an iterative oversight process.
“The end point of the technical method is a centralized containment and isolation of Canada’s used nuclear fuel in a deep geological repository with suitable geology and informed and willing host communities,” Van Drunen said, adding that the deep geological repository (DGR) blueprint aligns with international best practices established in Finland, France, Sweden, Switzerland, and the U.K.
However, it is the second, fluid half of the strategy—the management approach—that faces an existential collision with the digital age.
“The management approach, the second part of NWMO’s plan includes phased and adaptive decision making supported by public engagement and continuous learning throughout the process,” Van Drunen said. “A fundamental tenet is incorporating new knowledge and adapting to new technology, which is of course where AI fits in.”
The core hurdle for the nuclear sector is a severe technological mismatch. Van Drunen presented the assembly with a staggering timeline paradox showing how infrastructure projects operate on a completely different scale than computational development.
“Before I can describe NWMO’s approach to AI, we first need to understand time scales,” she cautioned. “So last year, NWMO selected this site for our future DGR. We are currently entering regulatory decision making. We expect to be in construction or beginning construction in the mid 2030s and in operation in the early 2040s.”
By the time heavy machinery begins excavating the Canadian repository site a decade from now, the digital landscape will have gone through dozens of iterations.
“By contrast, I understand that AI basically changes substantially every 3 months,” Van Drunen said. “So to put it in perspective, that’s on the order of 40 AI generations before we are even going to start construction.”
The ‘User’ Strategy and the Guardrails of Literacy
Given these vast time scales and the specialized nature of her agency, Van Drunen revealed that the NWMO—an organization of roughly 270 people that leverages an external supply chain—will strictly avoid trying to develop proprietary software.
“When we consider AI, a few things become clear for us,” she stated. “NWMO will not be developers of AI. But we need to be proficient and well-informed users of the technologies that are available at any given time to the extent it makes sense.”
The pivot from developer to consumer, however, does not mean importing commercial code blindly. Van Drunen issued a blunt warning regarding the limitations of standard commercial software: “The tools out of the box do not deliver value. To do this, our workforce and particularly our leadership must be AI literate.”
For Canada’s nuclear waste program, this literacy is a defensive necessity against an emerging technological double-edged sword.
“AI literacy will enable us to understand how AI can be used as an advantage for sure, but also frankly a threat,” Van Drunen warned. On one hand, she noted, the organization must “prepare us for tomorrow when AI is increasingly used, frankly, by threat actors to target NWMO systems and facilities.” Conversely, when managed properly, she noted the technology represents an unprecedented asset, “for when AI is a highly sophisticated, highly reliable tool for enabling NWMO operational excellence and sustainable nuclear operations.”
The Global View: Risk vs. Deployment Acceleration
As the keynote concluded, the floor opened to the international panel, exposing a diverse spectrum of strategies separating tech adopters from cautious overseers.
Representing the regulatory frontline, Mr. Petteri Tiippana, Director General of Finland’s Radiation and Nuclear Safety Authority (STUK), acknowledged that independent oversight agencies must adapt to society’s calls for efficiency. “Regulators do recognize the expectation from the industry and from societies to be more effective and efficient in licensing and regulation and oversight,” Tiippana said. He detailed that safety and speed do not have to be opposing forces if early engagement between vendors, operators, and regulators is established.
However, Tiippana was candid about the regulator’s stance on adopting the technology internally: “STUK is not in the front line of the AI users. We are following the developments, piloting something, maybe waiting that somebody else comes out with tangible results—but we are getting ready.” He advocated for a strictly “risk-informed approach,” emphasizing that AI has massive potential in low-risk administrative categories, thereby freeing up valuable human hours to double down on highly critical risk areas. “In any case, I think a human has to be there, at least in the beginning, and we should always understand what happens if the AI fails,” Tiippana warned.
Offering a vendor’s historical perspective, Mr. Bertrand Negrello, Vice President of Digital & Industrial Performance at Framatome, surprised the room by presenting a company article dating back to 1991 to prove that AI is a decades-long narrative in nuclear computing rather than a sudden novelty. Negrello outlined a tiered architectural structure for implementing machine learning based entirely on safety criticality.
According to Negrello, if a system governs a safety-related critical function, the industry must rely on deterministic AI systems. For less critical operations, probabilistic AI can be authorized, provided it is bound by rigid governance. “Let’s be clear that it’s not one unique AI that is to be implemented, and AI could fit the different situations including highly safety-critical features,” Negrello stated. He underscored that operators and regulators must maintain distinct software suites to avoid shared single-point failures: “If we share the same tools between operators and regulators, there is a risk. Different tools, different perspectives, but I think both are fully compatible.”
The conversation then shifted to physical plant design and construction monitoring, led by Mr. Huo Xiaodong, Vice President of the China Nuclear Power Engineering Corporation (CNPE). Huo, a reactor physicist who leads the design of China’s flagship Hualong 1 series plants, argued that safety and AI acceleration are inherently complementary.
“If we use AI properly, we can improve safety,” Huo asserted. He provided concrete operational examples currently in use, such as deploying smart acoustic and vibrational sensors on primary circuit components. “We use a good sensor to predict the secluded pump,” Huo explained. “It can take the abnormals very earlier to tell the operator you can make some maintenance earlier in case of the accident.” Furthermore, Huo highlighted CNPE’s use of computer vision and image recognition on “smart construction sites” to track unsafe physical behaviors among construction crews and run automated cyber-attack simulations to find design weaknesses before plants are ever built.
Bringing the commercial small modular reactor (SMR) perspective from the United States, Ms. Carrie Fosaaen, Vice President of Regulatory Affairs & Services at NuScale, focused heavily on using automated systems to dismantle the bureaucratic drag of licensing. Fosaaen described how NuScale is already utilizing machine learning to cross-screen massive technical submittals for procedural compliance, a job that previously swallowed days of engineering labor.
More notably, Fosaaen spotlighted AI’s capability to drive international regulatory harmonization, using advanced tools to run comparative analyses between US Nuclear Regulatory Commission (NRC) rules and foreign frameworks. “Within a matter of minutes, [AI] can provide a very quick screening on these are areas that I think you need to spend some more time… whereas that process for comparison used to take weeks to months,” Fosaaen said. She noted that nuclear-specific software excels because it provides absolute traceability. “It’s telling us this is not only what I found, but this is where I found it and how I came to my conclusion, so that when the human’s involved, they’re able to quickly assess the validity.”
Rounding out the panel’s geopolitical framework, Ms. Caroline Leadbeater-Hart, Nuclear Specialist Advisor to the UK Department for Business and Trade, urged delegates to treat the AI transition as a fundamental cultural shift rather than a software upgrade. Invoking the recent legislative findings of the UK’s Fingleton Report on mega-infrastructure delivery, Leadbeater-Hart noted that the civil nuclear sector must overhaul its legacy institutional mindsets to stay relevant and trusted.
“When AutoCAD came on in the 1980s, that was just a complete step change for us all to move away from hand-drawn schematics,” Leadbeater-Hart reminded the audience. “It made everyone feel uncomfortable at the time… Is it going to remove humans from the loop? And it has completely been a tool for the positive.”
She proposed an immediate, high-value global use case: assigning AI to ingest and analyze the historical operating and design data compiled by the nuclear industry over the last 60 years to calculate the exact margins of conservatism built into old plants. “Do we know what those safety margins are really? I don’t know if we do,” Leadbeater-Hart remarked. By exposing where designs are over-engineered or disproportionately restrictive, she argued, the industry can safely build at pace while remaining credible to public taxpayers and state treasuries.
Keeping the Human Accountable
To bridge the cultural chasm between fluid technological experimentation and a rigid, zero-mistake nuclear safety ethos, the session’s speakers repeatedly returned to the absolute requirement of human governance.
“We’ll be able to leverage AI to improve performance, efficiency, productivity… but at the same time ensuring that humans are accountable for AI systems and accountable for the outcomes,” Van Drunen said during the panel’s closing remarks.
The ultimate success of the “Atoms for Algorithms” era, the panel concluded, will not be judged by the raw complexity of the software deployed, but by the training of the personnel overseeing it. As Van Drunen summarized, the true hurdle for the international nuclear sector is a leadership paradox: fostering an environment of “responsible experimentation within a procedure-driven, healthy nuclear safety and security culture,” while ensuring the industry equips its workers “to be competent, ethical humans in the loop.”
About the Author
