One notable outcome from President Donald Trump’s visit to China is that Beijing agreed to formal discussions around artificial intelligence safety. Dialogue between rivals over contentious issues is welcome, but Washington could do more harm than good if American negotiators bring the nuclear arms control playbook to AI talks.
Treasury Secretary Scott Bessent recently announced that the world’s “AI superpowers” will begin talks to develop a framework for responsible use. He said one priority will be making sure that “non-state actors don’t get ahold of these models.”
Bessent also made clear that the U.S. lead in AI is an advantage Washington intends to keep: “The reason we are able to have wholesome discussions with the Chinese on AI is because we are in the lead.” He predicted that they wouldn’t be having talks if the Chinese firms were far ahead of American ones.
The secretary is right to push for the U.S. to stay ahead, but China won’t assent to any guardrails that set it further behind in the competition. That’s the kind of fantasy Sen. Bernie Sanders (I-Vermont) talked up during a recent event with Chinese academics. Hard-nosed realists know better.
The history of negotiating nuclear arms control agreements in the 20th century is instructive and suggests China won’t be too forthcoming over AI.
In 1963, the U.S., Britain and the Soviet Union signed the Limited Test Ban Treaty just as China was preparing for its first-ever nuclear test. Beijing called it “a complete scam” and claimed the deal was intended to strangle its program in the cradle.
Three decades later, when China signed the Comprehensive Nuclear-Test-Ban Treaty in 1996, the U.S. had conducted more than 1,000 nuclear tests to China’s 45. The freeze was real, and it locked China at a permanently lower technical level. Chinese leaders came to understand arms control treaties as tools the stronger party used to lock in its advantage.
Every subsequent arms control initiative has been filtered through the same lens: when a rival proposes restraint, ask who benefits. Beijing participates, but it does not constrain itself on anything that matters. There is no reason to expect AI safety talks to be different — especially when the American side opens by declaring that its own advantage will be preserved.
Atomic weapons were instruments of destruction that both sides had an interest in limiting, at least in theory. AI is not primarily a weapon — it will be the primary driver of productivity in the economy of tomorrow. Whoever leads in AI will lead in drug discovery, materials science, logistics and finance, as well as defense.
Asking China to cooperate on constraining AI development is asking it to accept permanent second-tier economic status. The real stakes are not existential risk or robot overlords — they are market share, industrial capacity and strategic autonomy. No country will volunteer to fall behind on those.
Still, dialogue is not useless. The most dangerous thing about AI in a military context is the compression of time to make decisions. The risk is that one side misreads the other’s intentions at machine speed.
Those are problems of trust and predictability that can be mitigated. Working groups and communiqués can build personal relationships and a level of baseline understanding. In a world where both sides are building systems that can act faster than either side can think, being in touch can be helpful so long as it does not create a false illusion of reliability.
The Chinese Communist Party remains untrustworthy. Beijing has previously demanded U.S. concessions in exchange for creating deconfliction channels, but military leaders haven’t always picked up when their American counterparts call during a crisis.
The lesson from these vexing experiences is that talking is fine but not in exchange for concessions. The Chinese won’t give up anything that might allow them to leapfrog America on AI.
The post How to get the most out of AI talks with China appeared first on Washington Post.
