What is being proposed
Melanie Sisson, a senior fellow at the Brookings Institution, and Tianjiao Jiang, an associate professor at Fudan University, published a set of recommendations in mid-September for limits on military AI that the United States and China could agree without settling anything else between them.
Four of the five are prohibitions. AI should not autonomously decide to use nuclear weapons. AI should not autonomously attack the other side’s nuclear command and control systems. Humans should retain sole authority to initiate significant AI-enabled cyberattacks on the other country’s nuclear command, control and communications, or on strategically important infrastructure such as energy, finance and healthcare. And both sides should agree a shared definition of meaningful human control, a phrase arms control documents use constantly without pinning down.
The fifth item
The fifth is a dedicated military channel between the two countries for reporting AI incidents — a hotline, in the sense the Cold War used the word.
That is the proposal most likely to survive contact with either government, because it commits neither side to a capability limit. It is also the one with the clearest theory behind it: if a model does something unexpected inside a military system, the dangerous interval is the one before the other side knows whether it was an accident.
Where it comes from
Both authors take part in the US–China AI and National Security Dialogue, running since 2019 under the Brookings Institution and Tsinghua University’s Center for International Security and Strategy. This is track-two diplomacy: scholars talking in a structure that governments watch without being bound by.
The substantive precedent is the November 2024 agreement between Joe Biden and Xi Jinping on maintaining human control over nuclear weapons decisions. These recommendations extend that principle outward, from the decision to launch to the systems that support it.
The timing
The paper landed ahead of fresh official US–China AI talks and a Trump–Xi meeting expected in Washington on 24 September. Neither government has endorsed any of it.
That is the honest measure of where this sits. Track-two work produces the text that officials can pick up if they decide they want one, and its function this month is to have something drafted before the meeting rather than after it.
What to watch
Whether AI appears in whatever the two governments say on 24 September, and in particular whether the incident channel is mentioned. A hotline is the cheapest confidence-building measure available, which makes its absence from any communiqué more informative than its presence.
