New export controls proposed

Chinese ministries of Commerce and Science and Technology have proposed changes to the list of technologies that cannot be exported or require special permission. The proposal includes adding forty new technologies, with twenty-three related to artificial intelligence. The public can comment on the proposal until November 15, but some restrictions are already being applied.

Targeted technologies specified

The proposal focuses on technologies related to large AI models, including their weights and code sources, as well as intelligent agents, computing cluster designs, and chip manufacturing methods using 14 nanometers or smaller. Fourteen of these technologies are completely banned, such as those involved in creating harmful software, while others require approval before export.

Focus on open access research

The changes specifically address AI research that is published openly by Chinese laboratories, including projects like DeepSeek, Qwen, GLM, and Kimi. Existing models already in use are not affected, but any new models might need approval or could even be banned. The proposal also includes plans to support the sharing of open-source AI code.

Supporting open-source development

The ministries have introduced measures to help open-source AI development, such as automatic approvals for certain technologies. These efforts are paired with plans for Chinese companies to build data centers overseas, produce chips outside of China, collaborate internationally on research, and contribute to global AI standards.

Rationale and political stance

The Chinese government says the changes are meant to protect national security, as some AI technologies could be used to attack key infrastructure or bypass security systems. The government emphasizes that the changes are not directed at any specific country or company and that its political stance remains unchanged.

Expert perspective on the proposal

Researcher Kristy Loke suggests the proposal is more about managing risks related to AI rather than being a tool for negotiation. The proposal also includes restrictions on AI that can find security flaws or generate harmful code, though the challenge is that such capabilities are tied to the general intelligence of the AI models.