AI and Nuclear: US-China Experts Call for Red Lines

Two specialists, one in Washington and the other in Shanghai, outline rules to keep AI away from nuclear decisions and strategic cyber operations. Their proposals follow a Biden-Xi agreement from November 2024, at a time when no binding mechanisms are in place and the effectiveness of crisis hotlines remains contested.

Binding mechanisms are lacking, doubts persist

To date, there are no binding mechanisms to frame the principles aimed at limiting the use of AI in the nuclear domain. Carla Freeman from Johns Hopkins University questions the effectiveness of hotlines in times of crisis, recalling that in 2023, during the spy balloon incident, China did not respond to calls from the United States. The current recommendations are based on an agreement reached between Biden and Xi in November 2024.

Ban AI autonomy and establish "human control"

Melanie Sisson and Tianjiao Jiang set several clear limits: it should not be possible for an AI system to independently trigger nuclear weapons or attack nuclear command devices. They emphasize that only a human should be able to control AI-driven cyberattacks against strategic infrastructure. The two countries are also urged to agree on a common definition of "human control." These measures aim to prevent decisions regarding the deployment of nuclear weapons from being made without human intervention.

A direct line for AI incidents and a Trump-Xi meeting

Tianjiao Jiang recommends establishing a direct line between the two countries to address incidents related to AI. She specifies that an automatic response from a defensive system to perceived suspicious activity could be interpreted as an offensive action, and an emergency channel would provide a government the opportunity to explain that an incident was unintentional before any escalation occurs. Melanie Sisson and Tianjiao Jiang released these proposals ahead of the scheduled meeting between Trump and Xi on September 24.