Experts urge US, China to bar AI from controlling nukes

A coalition of national security analysts, former military officers and technologists has urged the United States and China to pledge never to allow artificial intelligence to control nuclear weapons. The group argues that delegating launch authority to autonomous systems would increase the risk of accidental or inadvertent nuclear use.

The experts point to two main hazards. First, AI can amplify false or ambiguous information. Second, AI systems can fail, be manipulated by adversaries, or be exposed to cyber intrusions. They say faster data processing and pressure to shorten decision timelines could push planners toward automation of target selection, timing or firing sequences.

The group recommends a bilateral pledge between Washington and Beijing that keeps humans in the decision loop for any action involving nuclear release. Technical safeguards they propose include enforced human-in-the-loop controls, physical separation of launch hardware from networks, and strict limits on the use of machine-learning algorithms in command-and-control pathways. The experts also call for transparency measures such as exchanges of doctrine and notifications about deployments and tests that involve automation.

On verification, the experts acknowledge practical difficulties but offer steps to reduce uncertainty. They suggest inspections focused on software and hardware interfaces, agreed definitions distinguishing “control” from “advisory” AI, and shared test procedures that can demonstrate human authorization is required for any launch. They add that reciprocal political commitments would be needed to sustain trust.

The appeal arrives as both countries modernize nuclear forces and develop AI tools that can analyze sensor data and produce recommendations in seconds. Analysts reference past incidents in which human judgment prevented mistaken retaliation after false warning signals when explaining the rationale for keeping humans central to final launch decisions.

Some defense analysts raise objections to a formal pledge. They question how to draw clear lines between advisory systems and control systems, how to verify compliance, and how to set limits without constraining battlefield decision-making. These critics caution that unilateral declarations without verification mechanisms could leave gaps and that ambiguous definitions might be exploited.

The group calls on Washington and Beijing to begin talks to draft precise pledge language, agree verification options and set timelines before any operational AI systems are integrated into nuclear command-and-control architectures.