The JRC explains: How does AI strengthen nuclear safeguards? JRC explains | 19 May 2026 | Joint Research CentreWhat if checking that nuclear material is used exclusively for peaceful purposes no longer meant reviewing thousands of hours of surveillance footage? A new AI tool now cuts events for once-manual review by 98% while still spotting everything that matters. Developed by the Joint Research Centre (JRC), the tool is being deployed with high efficiency gains in the video surveillance reviews performed by Euratom nuclear safeguards at the Directorate-General for Energy. This saves inspectors’ time without reducing the quality of the results. Nuclear safeguards in practice Nuclear safeguards in place across the EU ensure that all nuclear material within power plants and other nuclear installations never falls into the wrong hands. This is checked through regular on-site inspections by Euratom nuclear inspectors and inspectors from the International Atomic Energy Agency. Video surveillance captures what happens at the installations in between the inspections. This means that inspectors need to regularly check video surveillance footage from nuclear installations across the EU. Sitting in a video review room in Luxembourg, they systematically scrutinise the footage to confirm that material is used as declared and nothing is diverted from peaceful use. With around 100 nuclear power plants in the EU and many more nuclear fuel cycle installations, this means trawling through thousands of hours of video footage. Though labour-intensive, this task is a crucial element of upholding nuclear safeguards in Europe and making sure that nuclear material stays accounted for. New AI tool cuts manual review by 98% A new AI tool ‘next generation model trainer’ developed by the JRC has been trained by inspectors to identify time periods where relevant activities may have occurred. This can cut the number of events for focused manual review by 98% - from 700 in a 6-month period to just 11. This means that inspectors can focus on the events that matter rather than hours of nothing or false alarms, while maintaining full control over the outcomes of the review. © EU Training AI in a secure environment Creating an AI tool and training it in a very secure environment with very restrictive data security constraints presents unique challenges. And the result must be totally reliable, and a tool the inspectors can trust. For this reason, the next generation model trainer is the result of close collaboration between the European Commission’s scientists and its Euratom inspectors. The models are initially trained quickly with limited data and then subject to a lengthy testing and validation process with the inspectors. The next generation model trainer integrates seamlessly into the existing safeguards verification workflows and complements established safeguards practices rather than replacing them. Its deployment demonstrates how AI can be responsibly developed, trained and applied in a sensitive domain, delivering efficiency gains while ensuring complete reliability, transparency and full human oversight. What’s next? The technology is being rolled out by Euratom inspectors in nuclear installations across the EU. In this way, the JRC is using technology to support the Commission’s broader efforts to strengthen nuclear safeguards and security, reinforcing Europe’s role in global nuclear non-proliferation and the safe and peaceful use of nuclear energy.
A new AI tool ‘next generation model trainer’ developed by the JRC has been trained by inspectors to identify time periods where relevant activities may have occurred. This can cut the number of events for focused manual review by 98% - from 700 in a 6-month period to just 11. This means that inspectors can focus on the events that matter rather than hours of nothing or false alarms, while maintaining full control over the outcomes of the review. © EU
