China has proposed tighter export controls on 23 artificial intelligence technologies, including large models, AI agents and computing clusters, while introducing an immediate registration requirement for affected technology transfers abroad.
The plan, unveiled on Oct. 16, covers 40 technologies in total, with five biomedical and 12 photovoltaic technologies also included. The proposed changes would expand government oversight of intellectual property and technical capabilities in industries central to China’s competition with the United States.
The draft update to China’s catalogue of prohibited and restricted technology exports was released by the commerce and science ministries. While the catalogue revision remains a proposal, the temporary registration system takes effect immediately. It also covers transactions that have begun but are not yet complete.
Exporters must register covered transactions involving transfers of technology ownership or licensing arrangements that allow foreign businesses to use Chinese technology. The distinction between transferring underlying technology and simply providing access to a service is central to the proposed AI framework.
Overseas users would still be able to access Chinese AI services through applications, websites, application programming interfaces and software-as-a-service platforms. Deploying restricted AI technology on servers abroad, however, could require an export licence. Chinese companies using their own overseas cloud infrastructure may qualify for streamlined procedures.
The proposed AI controls reach beyond finished products. They encompass technologies associated with model weights, checkpoints, source code, model architecture, and training or inference methods. Model weights are the numerical values learned during training; distributing them can allow another organisation to operate or adapt a model without relying on its original developer’s service.
The draft identifies several triggers for controls on core large-model technologies, including training computation of at least 10 to the power of 26 floating-point operations, a model size of at least two trillion parameters, or a place among the global top 10. It also addresses models trained on information classified as important data under China’s data-security framework.
Cybersecurity capabilities receive separate treatment. AI tools capable of independently creating or modifying malware to attack critical infrastructure, military targets or government agencies would face an export ban. Some vulnerability-discovery technologies would instead be subject to restrictions rather than an outright prohibition.
Technology already made public through open-source releases or academic papers is exempt. The framework also provides streamlined licensing routes for certain future open-source releases and international cooperation activities. Those provisions distinguish material already in the public domain from restricted technology that a developer intends to release or transfer later.
The proposal comes against a backdrop of U.S. restrictions on China’s access to advanced chips and semiconductor-manufacturing equipment. China’s Ministry of Commerce described its revision as routine technology-trade management and said it was not directed at any particular country, region or organisation.
Beyond AI, the proposed additions include cell and gene therapies and technologies used to manufacture polysilicon, silicon wafers and solar cells. Companies handling covered exports must use the interim registration process while consultation and the formal catalogue revision proceed.
Sources: Reuters, NTD News, The Decoder, The420.in