The News Central

Cooperation in Words, Controls in Practice: Can the US and China Really Talk AI?

When Xi Jinping sat down with Donald Trump in the Oval Office last week, one line from the Chinese readout stood out. On artificial intelligence, Xi said, there is more room for cooperation than competition between the two countries. He went further, saying both sides have "the capability and responsibility to develop and manage AI for good" and must ensure it stays "always under human control".

The words sound reassuring. Yet they were spoken at a time when Washington still restricts the chips that power advanced AI, and Beijing is working hard to make sure it never needs them. That raises an obvious question. Can two countries hold a genuine conversation about AI while each treats the technology as the central prize of their rivalry?

What Was Actually Agreed

The concrete outcome was modest but real. As part of an eight-point consensus that also covered a $30 billion reciprocal tariff cut, the two sides agreed to launch a formal dialogue on AI. China's Foreign Ministry said this would include a channel for communicating about AI-related incidents, with the first exchange expected in November.

The groundwork had been laid quietly. Treasury Secretary Scott Bessent had already proposed a notification mechanism for national security-related AI incidents when he met Vice Premier, He Lifeng in New York. Ambassador David Perdue called the new dialogue and crisis mechanism "historic".

What was left out matters just as much. Nobody has explained what counts as an "AI incident", which agencies will take part, or whether military systems and cyber operations fall within scope. Most tellingly, chip export controls were never on the table. US Trade Representative Jamieson Greer said plainly that those conversations "are not really happening in these negotiations", because controls are "the way we stay ahead in the race for superintelligence or AI".

The Export Control Problem

This is where the optimism runs into hard reality. American policy on AI chips has swung back and forth over the past year. Trump approved limited sales of Nvidia's H200 processors to China in January, and licences for a handful of Chinese firms followed in May. The result was underwhelming. Nvidia said H200 sales in China made up less than 1% of its data centre revenue in the second quarter.

Part of the reason lies in Beijing. After Washington eased the rules, China first held up H200 shipments at customs and later told companies to put domestic chips first. Huawei's Ascend processors, along with chips from Cambricon and Biren, are central to that push. China has concluded that relying on American hardware is a strategic weakness, whatever Washington offers.

On the American side, Congress is pulling in the opposite direction to the White House. Lawmakers from both parties are backing bills that would require location tracking on advanced chips, tighten controls on chipmaking equipment in coordination with allies, and write the ban on the most powerful AI chips into law.

So, each side is building walls around its technology while offering to talk about the dangers of that same technology. The contradiction is not fatal, but it limits what any dialogue can achieve. Cooperation on AI safety needs trust and some transparency. Export controls exist precisely because neither side trusts the other.

Why Talking Still Makes Sense

It would be a mistake to dismiss the dialogue as empty theatre. Rivals have long found ways to manage dangerous technologies without ending their competition. The Cold War hotline between Washington and Moscow did nothing to slow the arms race. It did make it less likely that a misunderstanding would turn into a catastrophe.

There is also a recent precedent. At the APEC summit in Lima in November 2024, Joe Biden and Xi agreed that decisions to use nuclear weapons should be controlled by humans, not by AI. It was the first time China had made such a statement. Xi's words at the White House about keeping AI "under human control" echo that commitment.

The realistic aim, then, is narrow. An incident channel could help the two sides clarify a suspicious cyber event, an AI-enabled military accident, or the misuse of a model by a third party before tempers flare. That is useful even if the two countries keep racing each other on everything else.

The bigger doubt is on the American side. Trump told the UN General Assembly that the US rejects any "globalist scheme" to control AI, and has said he wants development left "exactly where it is". Without a settled view in Washington on what AI safety even means, it is hard to see what firm commitments the US could ask of Beijing.

Conclusion

Is a US-China AI dialogue realistic amid export controls? Yes, but only on narrow terms. The two sides can talk about incidents, misuse and keeping humans in charge of the most dangerous decisions. They will not talk away the chip war, because both now see control of AI hardware as a matter of national power.

Xi's claim that there is more room for cooperation than competition is best read as diplomacy rather than description. The real test comes in November, when the first exchange is due. If the two sides agree even on a shared definition of an "AI incident", the dialogue will have earned its place. If not, it will join a long list of summit promises that looked good in a readout and faded soon after.