On Sunday in New York, US Treasury Secretary Scott Bessent sat down with Chinese Vice Premier He Lifeng and proposed something that sounds eminently sensible: a "notification mechanism" between the two countries, so each can warn the other when an artificial intelligence incident becomes serious enough to threaten national security.

An AI hotline. It goes to Presidents Trump and Xi at a summit in Washington this week.

The case for it is not really about trust. It is about attribution.

Sun Chengdao, a professor at Tsinghua University in Beijing, explained it to Al Jazeera better than any official did:

That is the whole argument in a paragraph. As AI systems act more autonomously, the gap between accident, crime and act of war stops being obvious from the outside. And the window in which someone has to decide which one they are looking at is short.

Bessent's version: "moving from opaque to more transparency between the number one and the number two AI powers in the world is very important." Trump's science and technology adviser Michael Kratsios said the key thing Washington wants to discuss is "avoiding shared risks."

One clarification, because it is being reported loosely. That vivid example about cyber-operations and cross-border model behaviour is Sun Chengdao's illustration, not a description of the proposal. Al Jazeera is explicit:

Here is the part that should shape how you read the headline.

This would not be the first crisis channel between Washington and Beijing.

In 1997, Bill Clinton and Jiang Zemin agreed to create a "direct secure telephone link" between the two countries.

Since 2008, they have had a military hotline, built precisely so senior defence officials could talk during dangerous incidents. It gives both sides 48 hours to respond.

And then Al Jazeera's sentence, which is the most useful thing in the article:

It gets more concrete. In 2023, after the US shot down a suspected Chinese surveillance balloon, then-Defense Secretary Lloyd Austin sought a call with his Chinese counterpart.

Beijing declined the request.

That was not a technical failure. The line existed and worked. Someone decided not to answer it - because in a confrontation, declining the call is itself a message. It communicates displeasure, withholds legitimacy, and refuses to let the other side set the tempo.

Nothing about that calculation changes because the subject is AI.

So the difficult problem was never building the channel. It is whether anyone answers during the exact category of incident the channel was built for, at the exact moment answering is most costly.

Most coverage will blur this. The proposal is American. It has not been accepted.

China's official Xinhua news agency confirmed that He and Bessent discussed AI, describing "frank, in-depth and constructive exchanges" on economic and trade issues - and did not say whether Beijing had accepted the notification mechanism. He Lifeng and chief trade negotiator Li Chenggang left the talks without speaking to reporters.

Xi's state visit runs Wednesday to Friday. Trump says he expects to discuss "almost everything."

It would be easy, and wrong, to write this off.

Scott Singer, codirector of the China AI Initiative at the Carnegie Endowment for International Peace, told Al Jazeera:

Jiang Tianjiao of Shanghai's Fudan Development Institute called it "a positive signal, showing that despite intense competition, there is still a willingness to cooperate."

Singer's framing is the realistic one: he does not expect the "fierce competition for AI to slow down," but "even competitors can find space to cooperate... The key for right now is for both sides to act in their own self-interest."

That is right. A channel that is answered sixty per cent of the time is enormously better than no channel. And the cost of building one is close to zero.

The reason neither side will simply concede the point is that both think they can win.

The US leads on money and chips. Stanford's 2026 AI Index reports US companies invested $285.9bn in AI in 2025, against $12.4bn in China, and the US retains an advantage in the most advanced semiconductors.

But in January, Chinese startup DeepSeek released a model built for a fraction of what its competitors had spent, and US tech share prices fell sharply. China also holds the rare earth minerals used in advanced technology, and generates more than twice as much electricity as the US - which matters enormously when the constraint on AI is increasingly cheap power for data centres.

Meanwhile the industry itself keeps asking for brakes. Al Jazeera cites Anthropic CEO Dario Amodei's essay warning that within 6 to 12 months an AI agent swarm could be "capable of taking over the entire internet with a persistent botnet," potentially causing "hundreds of billions of dollars in damage" - an argument endorsed by Sam Altman and Elon Musk.

Trump's position has been consistent: the US must keep its lead, and "Whoever wins AI, wins."

So a government that has rejected a slowdown, and told the UN Security Council days earlier that dialogue "cannot be allowed to drift towards global governance," is nonetheless proposing a bilateral safety channel. That is not a contradiction. A phone line between two capitals is not a world government. It is the narrowest possible form of cooperation - which is exactly why it might happen.

There is no child or classroom angle in this story, and we are not going to invent one. But there is a thread worth pulling, because it connects to something we covered a week ago.

When a false AI-generated intelligence report nearly led to US forces boarding a Chinese vessel, the failure was not a rogue machine. It was people unable to tell, fast enough, what they were actually looking at.

That is the same problem this hotline is aimed at, one level up. As AI systems act with more autonomy, more of what goes wrong will be genuinely ambiguous - not obviously an accident, not obviously an attack. Our institutions are built to answer "who did this and did they mean it," and that question is getting harder to answer at exactly the speed it needs answering.

A hotline does not fix that. It buys minutes in which someone can ask.

Whether anyone picks up is a different question, and the record is not encouraging.

Source: Al Jazeera, "What's the US-China AI 'hotline' that Trump plans to pitch to Xi Jinping?", 21 September 2026. https://www.aljazeera.com/news/2026/9/21/whats-the-us-china-ai-hotline-that-trump-plans-to-pitch-to-xi

Quotations verbatim. This is a US proposal that China has not publicly accepted, and the criteria that would trigger a notification have not been disclosed. The cyber-operation example is Sun Chengdao's hypothetical, not a description of the mechanism. The reading of declined calls as deliberate signalling, and the connection to the AI intelligence-report story, are ours.

Disclosure: Anthropic CEO Dario Amodei is quoted in the source article, and this package was drafted using Claude, made by Anthropic.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter