US, China Experts Want Nuclear-Style Rules for Military AI

As Washington and Beijing prepare for high-stakes talks on artificial intelligence, two experts — one American, one Chinese — have put the same radical idea on the table: treat military AI the way the superpowers once treated nuclear weapons.

Melanie Sisson, a senior fellow at the Brookings Institution, and Tianjiao Jiang, an associate professor at Fudan University in Shanghai, published separate papers last week proposing explicit red lines around the use of AI in warfare — borrowing directly from the Cold War playbook that kept nuclear competition from tipping into catastrophe. Their proposals come from the US-China dialogue on AI and national security convened since 2019 by Brookings and Tsinghua University's Center for International Security and Strategy.

The timing is deliberate. Both governments are preparing for government-level AI talks, and President Trump is scheduled to meet Chinese President Xi Jinping in Washington on September 24 — a summit where AI and trade are expected to dominate the agenda. Neither government has publicly endorsed the experts' proposals. But the fact that American and Chinese scholars are converging on the same framework, independently, is itself a signal.

A U.S. Air Force Minuteman III test launch. (File photo — illustrative, not connected to the proposals.)
Photo: U.S. Air Force, public domain, via Wikimedia Commons

What the experts are proposing

Sisson's paper argues for keeping humans firmly in control of the most consequential military decisions. Her central red line: humans must retain "sole authority" to initiate AI-enabled cyber operations against nuclear command, control and communications networks and against strategically important infrastructure. In other words, an AI system should never be the thing that decides to attack the networks through which nuclear orders flow.

Jiang's paper reaches for the same destination from the Chinese side. He calls for explicit, jointly recognised red lines that would bar AI systems from independently deciding to use nuclear weapons or from autonomously attacking nuclear command systems. He also proposes protections for critical civilian infrastructure — energy grids, financial systems, healthcare — against AI-driven disruption, and a dedicated hotline for AI-related military incidents.

The logic both papers share is Cold War logic. The United States and the Soviet Union never agreed on much, but they agreed that certain kinds of escalation served nobody's interest — and they built treaties, hotlines and norms around that shared fear. Sisson and Jiang are arguing that AI creates a new category of shared risk that demands the same treatment: systems that can act at machine speed, across domains, with consequences no human fully anticipated.

The scenario that keeps them up at night

The core fear is speed colliding with ambiguity. Imagine an AI system interfering — deliberately or accidentally — with the networks that carry nuclear command signals, or launching a military cyber operation on its own authority. Washington and Beijing would then have minutes to decide whether the other side had just attacked, working from information that might itself be incomplete, corrupted or AI-generated. It is the classic recipe for accidental escalation, compressed from hours to minutes by machine-speed decision-making.

This is not abstract theorizing. AI is already woven into military operations on both sides. This week, Chen Yixin — China's State Security Minister and one of Beijing's top arms-control officials — said AI has "fundamentally transformed the form of military struggle," moving from a supporting role to a key role in battlefield operations. He specifically cited AI-directed strikes in the U.S.-Israeli war with Iran as evidence of the shift.

Chen's remarks are significant for another reason: they confirm that AI is now formally on China's arms-control agenda. On the American side, AI policy remains divided across agencies, with the Pentagon, the White House and Congress pulling in different directions. Both governments, Reuters noted, remain wary of any framework that might constrain their own AI development — the same tension that has defined every arms-control negotiation in history.

Why the Cold War analogy works — and where it breaks

The nuclear comparison is powerful because the underlying game theory is the same. In the Cold War, both sides understood that a surprise attack on nuclear command systems could look like the prelude to a first strike — and that the victim, unable to be sure, might retaliate. So both sides agreed, tacitly and then explicitly, not to do it. The experts are saying AI needs its own version of that understanding before a crisis forces one into existence.

The hotline proposal is the most concrete borrowing. The original US-Soviet hotline was created after the Cuban Missile Crisis revealed how dangerously slow and garbled crisis communication could be. An AI-incident hotline would serve the same function for a faster age: a direct channel for Washington and Beijing to clarify, in real time, whether a strange signal, a cyber intrusion or an autonomous system's behavior was an attack or an accident.

The dialogue that produced these papers has been running quietly since 2019 — through the trade war, the chip-export controls and the balloon crisis — a rare channel where American and Chinese national-security thinkers could speak candidly about emerging technology without negotiating on behalf of their governments. That it has now produced convergent proposals, published independently on both sides of the Pacific, suggests the concern about military AI has moved from the seminar room to something closer to consensus among the experts who advise both capitals.

But the analogy has limits, and the papers' critics will surely note them. Nuclear weapons were countable, testable and attributable — you could verify a treaty by counting warheads and monitoring test sites. AI capabilities are none of those things. A red line against "AI independently deciding to use nuclear weapons" sounds clear until you ask what counts as "deciding," or how either side would verify the other's compliance. The history of cyber norms is instructive here: years of expert dialogue produced elegant frameworks that states mostly ignored.

There is also the asymmetry problem. The United States and China are not racing in identical ways — their military AI programs, their command structures and their doctrines differ substantially. A red line that constrains one side's advantage more than the other's will never be jointly recognised, no matter how sensible it looks on paper.

What happens next

The papers land in a narrow window. The September 24 Trump–Xi meeting in Washington is expected to cover AI alongside trade, and lower-level officials from both countries are planning separate AI-risk discussions. As Treasury Secretary Scott Bessent argued recently, the two countries need to work together on shared AI risks — a position that gives the experts' proposals a potential hearing at the highest level.

But expectations should be modest. Arms-control history suggests that frameworks like these emerge after scares, not before them — the hotline followed Cuba, the test-ban treaties followed atmospheric testing's horrors. The experts are trying to do something harder: build the guardrails before the crisis. Whether Washington and Beijing, each convinced it cannot afford to slow down, will listen to their own scholars remains the open question.

What is not in doubt is the direction of travel. AI has moved, in Chen's words, from supporting player to key battlefield role. The systems that will make the next decade's military decisions are being built now, in both countries, at speed. Sisson and Jiang are asking a simple question: before those systems are finished, can the two powers that build them agree on what they must never be allowed to do? The Cold War's answer, eventually, was yes. This time, the clock is running faster.

Sources: Reuters (Sept. 16, 2026, via syndication); Brookings Institution; Tsinghua University Center for International Security and Strategy. Internal links: Bessent on US–China AI risk talks; Altman, Musk and Amodei head to the White House for Trump–Xi summit.