U.S. and Chinese security experts have proposed joint principles to prevent artificial intelligence (AI) from independently deciding on attacks on nuclear command systems or the use of nuclear weapons.

On Sept. 17 local time, blockchain outlet Decrypt reported that a Brookings Institution and Fudan University report urged the United States and China to expand the human-control principle applied to nuclear weapons to military AI overall.

The report centres on blocking AI autonomy in decisions related to nuclear command, control and communications systems, known as NC3. The researchers said U.S. President Donald Trump and Chinese President Xi Jinping could agree that only humans, not AI, decide whether to launch cyberattacks targeting each other's NC3 or critical infrastructure. They wrote: "Decisions to launch cyberattacks on each other's nuclear command, control and communications systems or critical infrastructure must be made only by humans, not AI."

NC3 is a system responsible for threat detection, deciding on the use of nuclear weapons and transmitting launch orders. If any step in this process fails, it can lead to communications disruptions, misidentification of attacks, incentives for pre-emptive strikes, or accidental or unauthorised launches. The report said risks should be reduced together, including chance, human error, hard-to-predict AI behaviour and unauthorised actors triggering attacks with strategic consequences.

The researchers said extending principles used in the nuclear weapons field to other military AI areas does not presuppose a broad U.S.-China agreement on cyber security principles. They said: "Expanding the principle that attacks with strategic effects must reflect human intent and be under human responsibility does not mean Washington and Beijing must also share other cyber security principles."

The proposal also includes establishing a military AI hotline between the two countries and creating a common definition of human control. It is driven by concerns that AI makes decisions too quickly and its actions are difficult to predict. The report warned leaders may have little time to stop escalation. In the military sphere, concerns are growing that automated millisecond-level decisions could turn false alarms or cyberattacks into a nuclear crisis.

That concern also surfaced in a war-game experiment conducted in February. Models from OpenAI, Google and Anthropic chose to use nuclear weapons in 95 percent of the test scenarios. Experts warned, however, that the simulation design itself may have encouraged escalation.

In the same month, Anthropic rejected a request by the U.S. Defense Department to lift restrictions related to large-scale surveillance and fully autonomous weapons for Claude. The Trump administration later ordered federal agencies to stop using Anthropic technology, but a federal court on Aug. 28 found the administration had illegally retaliated against Anthropic.

Calls to slow AI development and maintain human control are spreading across the industry. Earlier this month, Anthropic Chief Executive Dario Amodei (다리오 아모데이) urged developers to slow the pace of the performance race, and Sam Altman and Elon Musk expressed support. OpenAI has asked Congress whether companies can coordinate development speed without violating antitrust law.

The report narrowed its scope to minimum rules that can be applied first in the absence of an agreement on comprehensive limits on development speed. It said only humans should be entrusted with approval of attacks on nuclear command systems and decisions on the use of nuclear weapons. The researchers said: "Governance of military AI is not a negotiation that can wait, but a race against time."