View / The US and China are hurtling toward Kissinger’s worst AI nightmare - Semafor
The late Henry Kissinger initially made his name as an academic writing about nuclear arms control during the Cold War, and latterly forged his reputation as the architect of US-China detente.
In his final years, however, Kissinger was fixated on what he believed to be an even greater danger: For the first time in human history, he observed, a technology — self-learning machines — could potentially target anything, anywhere, with any kind of destruction, making the world “100% vulnerable.”
He had become an AI doomer. And he pleaded for the US and China to open talks to avoid “catastrophic consequences.” The template would be nuclear arms control agreements with the Soviet Union, which he himself negotiated.
Alas, Kissinger may have been uncannily prescient about technology and its impact on global power, but his prescriptions on how to avert AI disaster haven’t aged well. As US President Donald Trump prepares for a summit this month with Chinese leader Xi Jinping, Cold War analogies have become pointless — and the reasons why are instructive.
Beijing refuses to consider nuclear arms-limitation talks for precisely the reasons it would balk at joining the US in slowing AI progress, or “pacing the frontier” in tech jargon.
The US has many more nuclear warheads. It also has at least a six-month lead in AI models, and vastly more compute. From Beijing’s perspective, restraint in both domains would only lock in Washington’s advantages, leaving China open to bullying and blackmail. No Chinese leader would perform that kind of kowtow.
For similar reasons, China spurns guardrails. Moscow and Washington had elaborate protocols to avoid inadvertent escalation during the Cold War, but Beijing thinks they would simply encourage US recklessness: The US and China can’t even agree on “rules-of-the-road” to prevent accidents involving ships and planes in geopolitical flashpoints, like the South China Sea, from spiraling into armed conflict.
Some analysts argue that the Obama administration set a hopeful precedent on AI cooperation by coaxing China to join the Paris Climate Agreement in 2016. But that, too, is a stretch. In fact, China had its own powerful reasons to sign up for climate action, not least a public revolt against killer smog blanketing its megacities, and a security imperative to swap out imported oil for home-grown green technologies.
On the present trajectory, the US and China are hurtling toward the very calamity that Kissinger warned about.
US President Donald Trump is no Nixon. In the run-up to next week’s meeting, he’s framed the stakes in zero-sum terms. “We’re leading China in AI,” he told reporters over the weekend. “And frankly, I want to keep it that way because whoever wins AI, wins.”
Scott Bessent, the US Treasury Secretary who’s preparing the summit, put it even more starkly, warning that if the US loses the AI race to China, “there is no day after tomorrow.”
This is classic game theory that calls to mind the chilling 1950s debates over second strikes and counterforce. Both sides have a clear interest in slowing the development of the most advanced AI models — the Hugging Face attack was a wake-up call in the US; China has been confronted with the prospect of rogue agents taking over WeChat, its ubiquitous chat and payments platform. But they don’t trust each other enough to do so. Silicon Valley labs assume that if they pause, China will race ahead anyway.
In an interview with The Economist in 2023, just before his 100th birthday, Kissinger warned that the US and China had five to 10 years to avoid an AI disaster. The clock is ticking. But, for now at least, neither side is watching the dials.
