The US and China agree on the need for bilateral talks to reduce the risks from advanced artificial intelligence. Such talks could lead to restrictions on using AI to control nuclear weapons command and control systems.
The affirmation of the need for government talks on AI safety between US president Joe Biden and China’s president Xi Jinping during the Asia-Pacific Economic Cooperation summit in San Francisco on 15 November. It may provide some common ground at a time when the US has in an effort to limit China’s AI development.
“It’s pretty low-hanging fruit, because we all agree that nuclear weapons should not be used and we all agree that we should responsibly manage them,” says at the Asia Society Policy Institute, a think tank based in New York and Washington DC. “And then we also can agree that having a human in the loop is important for [managing] escalation of nuclear weapons.”
Advertisement
Despite and on the possibility of an agreement on the issue of AI control over nuclear weapons leading up to the meeting, no formal agreement was immediately achieved. But the political and military leaderships of both countries have considered the issue independently.
In October 2022, the US Department of Defense committed to over nuclear weapons usage. Similarly, Chinese expert discussions and papers suggest no rush to apply AI or autonomous systems to nuclear weapons.
“While overall [Chinese military] staff is much more in favour of seeking aid from machines, the political leadership is more sceptical about handing too much control to software,” says at the Peace Research Institute Oslo, a think tank based in Norway.
There are many ways in which integrating AI into nuclear forces – especially full automation of nuclear command and control systems – could go wrong, says at the Stockholm International Peace Research Institute, a think tank based in Sweden.
“The core problem with AI is that it is difficult for states to evaluate each other’s capabilities and intentions,” says Boulanin. “Potential misperception of capabilities or intentions could be destabilising and increase nuclear risk.”
He and other experts described any possible US-China discussion on the topic as a positive development. But they also cautioned that it would be difficult to verify any eventual agreements unless the US and China took the unlikely step of allowing outsiders to scrutinise their nuclear launch systems’ software code.
Not all automation is necessarily bad if used to reduce the risk of accidental nuclear war, says at the RAND Corporation, a think tank in California. He cited the Soviet Union’s cold war development of the automated Perimeter system that would automatically grant launch authority to field commanders in possible nuclear war scenarios.
That system – which did not initially use modern AI – was designed to reduce pressure on Soviet leaders to launch nuclear weapons immediately if they thought they were under nuclear attack, while keeping a human in the loop. By comparison, the Soviet Union rejected the idea of implementing a fully automated system.
“Nobody is asking for a Skynet to begin with,” says Geist.
