In April 2026 the UK's Office for Nuclear Regulation (ONR) issued a summary report, reference ONR-FUT-001, on a regulatory sandboxing project it led between September 2025 and March 2026, and announced the findings in May. The project was one of the experiments funded by the government's Regulatory Innovation Office through its AI Capability Fund, which put up to £3.6 million across UK regulators to learn how to regulate AI by working alongside industry; ONR's share was a little over £260,000. For that money, four regulators — ONR, the Environment Agency, the Health and Safety Executive and the Defence Nuclear Safety Regulator — spent seven months with four dutyholders and two AI-assurance specialists working out what a safety case for a machine-learning component in a nuclear installation would actually have to contain. This blog also covered a JRC workshop on AI for nuclear materials where an ONR representative sat on the closing panel; this report is the regulator's own account of the same technology from the permissioning side.
What a regulatory sandbox meant here
In AI policy, "sandbox" usually means a legal carve-out: a controlled space where a provider can test a system under relaxed obligations. That is not what ONR ran. Nothing was deployed, nothing was exempted, no permission was granted. The sandbox was a structured series of workshops — two-hour technical sessions per use case, half-day "sprint" workshops to refine an AI problem-and-opportunity statement, and one-day face-to-face deep dives — organised around a mock assurance case. That mock case borrowed the claims-argument-evidence structure UK nuclear safety cases already use and reduced it to three top-level claims: the requirements for the AI component are sufficiently defined, the component meets them initially, and it keeps meeting them in the future. The deep dives then asked, for each claim, what evidence would be needed and whether it was likely to exist. An enabling workshop beforehand taught participants the automotive sector's AMLAS machine-learning assurance framework from the University of York's Centre for Assuring Autonomy, and medical-sector assurance approaches were consulted too.
The scope was deliberately narrow. Both use cases used supervised machine learning with no continuous learning — a model trained offline, then frozen — performing a single task on computer-vision data. The report calls this "narrow ML" and is candid about why: validating a fixed classifier that flags defects in a radiograph is something a human can check, in a way that validating a general-purpose generative model is not. The project builds on ONR and the Environment Agency's 2023 pilot sandbox and on the trilateral principles paper ONR published with the U.S. NRC and the Canadian Nuclear Safety Commission in September 2024. The stated difference from the pilot is depth: this time the industrial partners brought concepts of operation, workflows and operator roles detailed enough to draw out findings about assurance rather than generalities.
The two use cases, with the numbers the report gives
Use case 1: AI screening of weld radiographs. The dutyholders were Rolls-Royce Submarines and NNB Generation Company, the Hinkley Point C licensee, with the Health and Safety Executive and the Defence Nuclear Safety Regulator joining ONR. Today a qualified radiographer reads every film looking for pores, wormholes and lack of fusion; the hazard is a leaking weld that becomes a radioactive release. The proposed system has one AI component classify radiographs as "cleared" or "with defects", a second classify the defects as "actionable" or "acceptable", and a human radiographer sample-review the AI's decisions — reviewing every film with an actionable defect, but only periodically checking the cleared ones. The performance target is the current industry standard, a 90% probability of detecting actionable defects at 95% confidence. The system is being built in-house, the models are static once trained, and — a detail the report states without comment — they are trained and tested on images of each defect type with no "normal" images in either set. The rough benefit estimate is about £20 million on a typical new-build project.
Use case 2: AI for radioactive waste characterisation. Sellafield and Nuclear Restoration Services, with the Environment Agency alongside ONR. This is really two applications. One uses computer vision on video of sludge drops to infer shear strength from how the material deforms and advises the operator how far to fill a storage skip; the preliminary target is that the AI correctly characterises at least two of every three drops, the training data does not exist yet and would come from a test rig, and Sellafield has publicly put the potential saving at £100–500 million in its own 2024/25 research and development review. The other uses vision to identify specific items during sorting and segregation of intermediate-level from low-level waste, replacing a manual visual step, with a separate non-AI gamma detector verifying the radioactivity of whatever the AI picks out. Current ML detection performance is around 80%; the existing manual process achieves zero misconsignment, and the stated objective is to raise throughput without moving off that baseline. The Nuclear Decommissioning Authority told a March 2026 research forum that autonomous sorting and segregation across its estate could save more than £400 million.
Both architectures are framed the same way: the AI sits inside a defence-in-depth arrangement where independent layers would have to fail consecutively before harm arises. In several of the sub-cases, the report notes, the ML component did not play a large role in the overall concept of operation, and the consequences of its failure were partly caught by later steps in the workflow. That is a big part of why the conclusion came out the way it did.
What the regulators concluded, and how carefully they said it
The headline finding appears twice in almost identical words: "no fundamental reasons were identified as to why the approach taken by the proposed concepts would be incompatible with the existing regulatory framework." The qualifiers matter as much as the sentence. The finding is "based on the limited information available in the regulatory sandboxing and within the bounds set in the report disclaimer." That disclaimer says the report "does not represent a formal regulatory position", that any future permissioning decision will be made by an independent regulatory team through the normal process, and — a candid admission — that the industrial partners, because of the timescale, "provided only limited input" to the report itself. Whether either use case can demonstrably meet safety and environmental requirements "will depend on how the detailed design and execution of each use case is developed."
Read one way, that is a regulator saying almost nothing. Read the other way, it is a regulator putting on paper that a system where a human only sample-checks an AI's weld verdicts, and one where an AI replaces a manual visual step in sentencing radioactive waste, are the kind of thing its framework can accommodate — without a single new rule. The report's explanation is the UK's goal-setting regime: the law sets out what the dutyholder must achieve, not how, and the ALARP and Best Available Techniques tests apply to an ML component exactly as they apply to anything else. It goes further and claims a goal-setting framework "can better accommodate emerging technologies such as AI than rules-based regimes, where legislation, standards and guidance often need time to mature." It does not name the EU AI Act, but the comparison is hard to miss; the AI Act also makes regulatory sandboxes a formal instrument, the difference being that the UK version was convened by a sector regulator with no dedicated AI statute behind it. The report also credits the project with pushing industry toward "more ambitious options compared to the original use cases proposed by the industry" — the regulator nudged the dutyholders to lean on the AI harder, not less, consistent with the stated aim of moving from low-autonomy, low-significance applications toward the centre of the trilateral paper's four-box model.
The technical findings that travel beyond nuclear
Annex 2 is where the report earns its keep. Several of its findings are the clearest statement I have seen from a safety regulator of problems every high-consequence AI deployment shares.
- The operating point is a requirements decision. The balance between false positives and false negatives must be set "when defining the requirement specification for the system", because it "can create a trade-off between efficiency and safety": false positives generate re-work, false negatives are missed defects that need mitigating by other means. This is exactly the trade-off PNNL found in practice when a newer version of a commercial weld-inspection tool raised detection from 97% to 100% while its false-call rate rose from 58% to 83%. ONR's position is that a dutyholder should have chosen where on that curve it wants to sit before the system exists, not discovered it afterwards.
- Training data is structurally scarce, and the report says why. Defects in nuclear welds are rare precisely because manufacturing quality is high, so the class a classifier most needs to learn has the fewest examples. Whether to use data from other applications or synthetic data should be decided "as part of the initial definition of the system" — the same conclusion the F4E and Argonne speakers reached at the JRC workshop from the materials side.
- Drift invalidates assurance, so revalidation scope has to be defined in advance. Intentional changes (upgrades) and unintentional ones (shifts in the data distribution) "can alter system behaviour in ways that may invalidate earlier assurance claims." The report asks for a proportionate impact-assessment framework, because "not every change requires complete revalidation."
- Transferring a model to a similar job means re-assessing its assurance case. The general learning transfers; the validation does not, and transfer gets harder as the level of dependence on the AI in decision-making rises.
- The dutyholder has to be an intelligent customer, and those skills are "currently limited within the nuclear sector". Licensees remain legally accountable however much governance they contract out. The report flags intellectual-property friction with vendors asked to hand over training data and performance metrics, and treats AI supply-chain security as a continuous "security-by-design" obligation with provenance kept for all third-party data, models and dependencies, because "AI systems are uniquely susceptible to adversarial attacks."
- Over-trust and under-trust are both failure modes. The human-AI section asks for defined roles and decision limits, visible uncertainty, and training so operators can recognise when the AI is unreliable — and it raises the long-term question directly: where humans retain the final decision, the safety case must consider how growing reliance on AI erodes their skills and situational awareness, and what happens when the AI is unavailable and a human has to step back in.
- Early operation is an evidence-gathering phase, if the design allows it. The report suggests using trials and early operations to measure real false-positive and false-negative rates, starting with stringent controls and relaxing them as evidence accumulates — which only works if the facility was designed with enough flexibility to adjust. It treats a period of higher uncertainty after deployment as something to be managed inside a tolerable ALARP envelope, not something to be eliminated before deployment.
One line deserves its own mention. When system requirements are incomplete, the report says, "risk assessment lacks a clear reference point", and this is worse for AI, "where defining requirements and assessing risk are already more complex." That is a regulator conceding that standard safety-case machinery assumes a completeness of specification ML systems do not naturally provide.
The money problem, and three ways around it
Annex 3 is the least-noticed section and the most revealing about whether any of this scales. Both nuclear AI sandboxes to date were paid for by competitive government grants, and the report says plainly that this "can restrict the opportunity to support future sandboxing in the timescale needed by the UK nuclear sector." Three options are on the table: industry funds future regulatory sandboxes, perhaps through the cross-industry heads-of-innovation forum; ONR creates a lighter "product" that stress-tests an innovation early for fundamental shortfalls with fewer stakeholders; or industry runs its own sandboxes and invites the regulators in on specific points, which the report admits would produce different outputs because the objectives would differ. None is decided; the target date is the fourth quarter of 2026.
That sits alongside a broader shift. The Nuclear Regulatory Taskforce report of November 2025 cited the ONR sandbox as a success and its recommendation 41 called for a national nuclear digital programme including AI; the government accepted all 47 recommendations in March 2026. And the report's own reference list contains the line that undercuts any triumphalism: ONR's April 2026 research paper on AI and ML found that "where there are significant consequences arising from failure of AI, ongoing research has to date been unable to identify suitable ways to assure safety and security." The sandbox got its reassuring answer partly because the two use cases were chosen to keep the AI's safety significance low, with other layers behind it. The hard case — an AI component the safety case actually relies on — is the one the field's own research says it cannot yet assure.
What to expect next
- Watch for the sandboxing handbook, due in the fourth quarter of 2026. ONR committed to consolidating the administrative and technical learning into a handbook for future sandboxes; if it becomes a reusable template for other UK sector regulators, that is the project's most durable output.
- Watch the funding decision, also due by the end of 2026. Whether future sandboxes are paid for by industry, replaced by a lighter regulator "product", or run by industry itself determines how many more there will be, and who sets the questions.
- Watch the one-year review with the project partners, targeted for the first quarter of 2027. That is the first point at which the report's assumption — that the discussions reduced enough uncertainty to accelerate real deployment — gets tested against whether Rolls-Royce, Hinkley Point C, Sellafield or NRS actually moved either use case toward a submission.
- Watch the IAEA's forthcoming TECDOC on the safety and security implications of AI in nuclear, in preparation with ONR input. It is the document most likely to carry the narrow-ML distinction and the operating-point finding into international guidance.
- Watch whether any dutyholder submits a safety case in which the human only sample-reviews the AI. That is the design use case 1 describes, and it is where "no fundamental objection in a sandbox" and "permission from an independent regulatory team" finally become two different things.