For half a century, the world’s nuclear powers relied on an intricate and complex series of treaties that slowly and steadily reduced the number of nuclear weapons on the planet. Those treaties are gone now, and it doesn’t appear that they’ll be coming back anytime soon. As a stopgap measure, researchers and scientists are suggesting a bold and weird path forward: using a system of satellites and artificial intelligence to monitor the world’s nukes.
“To be clear, this is plan B,” Matt Korda, an associate director at the Federation of American Scientists, tells WIRED. Korda has written a report at FAS that outlines a possible future for arms control in a world where all the old treaties have died. In Inspections Without Inspectors, Korda and coauthor Igor Morić describe a new way to monitor the world’s nuclear weapons they call “cooperative technical means.” In short, satellites and other remote sensing technology would do the work that scientists and inspectors once did on the ground.
Korda says AI could help this process. “Something that artificial intelligence is good at is pattern recognition,” he says. “If you had a large enough and well-curated dataset, you could, in theory, train a model that’s able to identify both minute changes at particular locations but also potentially identify individual weapon systems.”
New START, an Obama-era treaty that limited the amount of nuclear weapons the United States and Russia deployed, expired last week, on February 5. (Don’t worry, the countries reportedly still plan to maintain the status quo—for now.) Both countries are spending billions to build new and different kinds of nuclear weapons. China is building new intercontinental ballistic missile silos. As America withdraws from the world stage, its nuclear vouchsafes mean less, and countries like South Korea are eyeing the bomb. Trust between nations is at an all-time low.
In this environment, Korda and Morić’s pitch is to use existing infrastructure to negotiate and enforce new treaties. No country wants “on-site inspectors roaming around on their territory,” Korda says. So, failing that, the world’s nuclear powers can use satellites and other remote sensors to monitor the world’s nuclear weapons remotely. AI and machine-learning systems would then take that data, sort it, and turn it over for human review.
It’s an imperfect proposal, but it’s better than the literal nothing the world has now.
For decades, the US and Russia have worked to reduce the amount of nuclear weapons in the world. In 1985 there were more than 60,000 nukes. That number is down to just over 12,000. Eliminating roughly 50,000 nuclear weapons took decades of dedicated work from politicians, diplomats, and scientists. The death of New START represents the refutation of those decades of work. These on-site inspections fostered trust between Russia and the US and laid the groundwork for a drawdown of tensions during the Cold War. That era is over now, replaced by an age of acrimony and a renewed nuclear arms race.
“The idea we had in this paper was, what if there was a sort of middle ground between having no arms control and just spying, and having arms control with intrusive on-site inspections which may no longer be politically viable?” Korda says. ”What can we do remotely if the countries cooperate with each other to facilitate a remote verification regime?”
Korda and Morić’s proposal is to use the web of existing satellites to monitor intercontinental ballistic missile (ICBM) silos, mobile rocket launchers, and plutonium pit production sites. One big hurdle is that a good implementation of a remotely enforced treaty regime would require a certain level of cooperation. The nuclear powers would still need to agree to participate.
“They could transmit a message saying, ‘Hey, for the purposes of verifying this arms control treaty, we want you to open up silo hatch X on this day, at this specific time, because that’s when our satellite is going to be flying over,’” Korda says. “And so that enables you to have this sort of mutual verification using technology that currently exists. It doesn’t require anyone on-site, and it doesn’t require you to do this persistent spying-style coverage, where you kind of just hope to catch glimpses of things. Instead, you’re sort of working together to do a mutual verification.”
AI systems, verified by humans, could watch over it all. There are, of course, many problems with this. One is that competent AI systems need large datasets to be good at a trained task, and there’s just not a lot of nuclear-weapon training data. “You have to build these bespoke datasets for each country,” Korda says. “Here’s how Russia builds ICBM silos. Here’s how the United States builds ICBM silos, that sort of thing. But even within countries, there can be differences.”
Sara Al-Sayed of the Union of Concerned Scientists has built one such dataset as part of a forthcoming study about the use of AI systems for arms control verification. The focus of her study is missiles, but she says that there’s a lot more data a hypothetical AI-powered nuclear monitor would need to track. “You could think of all manners of things like missiles, the launchers, the bombers, the submarines, the sites of their production, the testing, the storage, the maintenance, and the dismantlement, including for any and all objects that are present at those sites,” she tells WIRED. “So you really need to think at that granular level of all the objects.”
And then there’s the question of what, exactly, these AI systems would be doing. “What is the task?” Al-Sayed says. “Is it that you want to detect the presence or absence of an object? Classify what you’re seeing? Or do you want to be identifying and tracking changes over time?”
In a hypothetical world where AI and remote sensing replaces on-site inspections, the parties agreeing to a new treaty regime would need to agree on how the AIs work and what they’re tracking. That means new rounds of negotiation and new agreements at a time when countries aren’t even honoring their old agreements.
Al-Sayed says that if we get to a place where countries are negotiating the specifics of AI arms control, then the world is already in trouble. “Why would you want to rely on an AI-based verification regime?” she says. “If you believe that automation is necessary, then you are in this paradigm where you feel like you need to catch every instance of your adversary or arms control treaty partner cheating. How is it that two parties or more could come together to even agree to negotiate an arms control agreement or treaty if the assumption is going to be that every single action could be suspicious?”
Al-Sayed’s research into AI and arms control has also shown her that these systems are more complex than their boosters would have you believe. “There’s an inherent stochasticity of these techniques, starting from the process of curating the data that you are going to use in order to train your model, then the labeling, and then the model itself, and the random performance of that model and its lack of explainability, and then the randomness that comes out,” she says
“How can we make the machines themselves trustworthy?” she asks.
At the moment, we can’t. AI systems fail every day. They ship with massive security flaws, and the people who design them often can’t explain how they work. The technology would be hard to hang a new nuclear arms control regime on.
But all treaty regimes are imperfect, and no country is going to allow a foreign power to tour its nuclear facilities these days. For Korda, Morić, and others at FAS, it’s about triage—using satellites and imperfect AI systems to keep watch on the world’s nuclear weapons could be a small bridge to a better world.
“A successor to New START is not going to put us on the path towards disarmament,” Korda says. “It’s just going to help us prevent a real spiral into hundreds more additional nuclear weapons being deployed.”
The post AI Is Here to Replace Nuclear Treaties. Scared Yet? appeared first on Wired.
