The old machinery of nuclear arms control is no longer doing the work it once did. For decades, treaties, inspectors and verification practices helped the United States and Russia reduce global nuclear arsenals. Now, researchers are asking whether satellites, remote sensing and artificial intelligence could become part of a new verification model.
The proposal is not a replacement for trust. It is a response to a world where traditional nuclear treaties have weakened, on-site inspections may be politically harder to accept, and countries still need some way to check whether arms control commitments are being followed.
Why nuclear verification is looking for a new model
The immediate backdrop is New START, an Obama-era treaty that limited the amount of nuclear weapons the United States and Russia deployed. It expired last week, on February 5. The source notes that the countries reportedly still plan to maintain the status quo for now, but the treaty’s end leaves a major gap in formal nuclear arms control.
That gap matters because previous arms control work took decades. In 1985 there were more than 60,000 nukes. The number is now just over 12,000. Reducing roughly 50,000 nuclear weapons required sustained work by politicians, diplomats and scientists.
The concern is not only about one treaty. The source describes a broader environment in which both the United States and Russia are spending billions to build new and different kinds of nuclear weapons, China is building new intercontinental ballistic missile silos, and countries like South Korea are eyeing the bomb. Trust between nations is described as being at an all-time low.
The satellite-and-AI proposal
Matt Korda, an associate director at the Federation of American Scientists, and coauthor Igor Morić have outlined a possible approach in a report called Inspections Without Inspectors. Their term for the model is “cooperative technical means.” The basic idea is that satellites and other remote sensing technology could take on some of the verification work that inspectors once performed on the ground.
The system would not simply be passive surveillance. It would require countries to cooperate with each other in specific ways. For example, a country could be asked to open a silo hatch at a particular time when another country’s satellite is passing overhead. That would allow a remote check without sending inspectors into the territory.
In this model, AI and machine-learning systems would help process the incoming data. Korda argues that artificial intelligence is useful for pattern recognition. With a large enough and well-curated dataset, such systems could in theory help identify small changes at known sites or potentially identify individual weapon systems.
The targets could include:
- intercontinental ballistic missile (ICBM) silos
- mobile rocket launchers
- plutonium pit production sites
The appeal is clear. Countries may reject intrusive inspections, but they may still accept a verification process that relies on existing satellites and remote sensors. Human review would remain part of the process, with AI sorting and flagging data rather than independently deciding whether a treaty has been violated.
Cooperation is still the hard part
The proposal sounds technical, but its biggest obstacle is political. A remotely enforced treaty would still need nuclear powers to agree to participate. They would have to decide what is being monitored, when data is collected, what counts as relevant evidence and how each side can verify the process.
That means the system would not remove negotiation from nuclear arms control. It would create new subjects for negotiation. Countries would need agreements not only about weapons, but also about the AI systems, remote sensing methods and verification tasks used to monitor them.
This is why the proposal sits between two unsatisfying options. On one side is no arms control at all. On the other is an older model built around on-site inspections that may no longer be politically viable. Cooperative technical means is an attempt to find space between those choices.
The data problem for AI arms control
AI verification also faces a basic training problem. Competent AI systems need large datasets for the task they are meant to perform, and nuclear-weapon data is limited. Korda says that bespoke datasets would be needed for each country because Russia and the United States build ICBM silos differently, and even within a single country there can be variation.
Sara Al-Sayed of the Union of Concerned Scientists has built one such dataset as part of a forthcoming study about AI systems for arms control verification. Her study focuses on missiles, but she says a broader nuclear monitor would need to track many more things: missiles, launchers, bombers, submarines, production sites, testing, storage, maintenance and dismantlement, including objects present at those sites.
That level of detail shows why the task cannot be reduced to simply pointing AI at satellite images. The parties would first need to define the task. Is the system checking whether an object is present or absent? Is it classifying what appears in an image? Is it identifying changes over time?
Each version of the task creates different technical and diplomatic problems. A system built to classify objects is not the same as one built to track changes. A system trained on one country’s infrastructure may not work reliably on another country’s facilities.
Trusting the machines may be the biggest barrier
Al-Sayed’s critique goes deeper than dataset size. She argues that AI systems can be more complex and less dependable than their advocates suggest. Data curation, labeling, model behavior, performance variation and lack of explainability all introduce uncertainty.
That matters because nuclear arms control is not an ordinary monitoring problem. A false signal, an unexplained model output or an opaque classification could become politically dangerous if governments treat it as evidence of cheating. Human review can reduce that risk, but it does not erase the underlying problem of relying on systems that may be difficult to explain.
Al-Sayed’s central question is whether machines themselves can be made trustworthy. The source’s answer is stark: at the moment, they cannot. AI systems fail, can ship with major security flaws, and are often difficult even for their designers to explain.
That does not make the satellite-and-AI proposal irrelevant. It makes it a warning as much as a plan. Remote verification could help preserve some form of nuclear arms control when older treaties are gone, but it cannot substitute for political agreement, transparency and trust. If AI becomes part of nuclear verification, it will need to be treated as a tool for humans to scrutinize, not as an authority that can solve the crisis on its own.