U.S. and Chinese security experts proposed joint principles to prevent artificial intelligence from independently deciding on attacks on nuclear command systems or the use of nuclear weapons. A Brookings Institution and Fudan University report urged expanding the human-control principle used for nuclear weapons to broader military AI. It called for humans alone to decide on cyberattacks targeting NC3 or critical infrastructure and warned that fast, unpredictable AI could escalate crises. The proposal also includes a military AI hotline and a shared definition of human control.

The U.S.-China AI rivalry is moving toward using AI to build stronger AI, with companies focusing on coding, experiment design and training methods to boost next-generation models. The long-term goal is recursive self-improvement, in which AI produces improved successor models in a feedback loop. OpenAI, Anthropic and Google DeepMind cite progress, while analysts say the United States leads in AI-based independent research. Researchers and officials warn safety risks are rising, but neither country has publicly demonstrated fully autonomous recursive self-improvement.