SINGAPORE, July 7 - Over the last month, Chinese authorities convened meetings with leading domestic technology firms to explore measures that would limit overseas access to the country’s most advanced artificial intelligence models, including some that have not yet been released, three people familiar with the discussions said.
The sessions, held amid a broader policy push to keep key AI capabilities onshore, were intended to map out potential controls on the distribution and use of frontier models. Participants at the meetings included Alibaba, ByteDance and startup Z.ai, the sources said. The individuals spoke on condition of anonymity because they were not authorised to discuss the conversations publicly.
Scope of proposed restrictions
Officials leading the meetings were from China’s Ministry of Commerce, which handles export regulation, and the National Development and Reform Commission, the state planning agency, both of which had representatives in attendance, sources said. At the sessions, participants examined options for limiting access to the most capable models, addressing both closed-source systems and versions with more open weights, according to two of the sources.
One line of discussion raised at the meetings was to treat any unauthorized leak or theft of proprietary AI technologies as an offence under China’s national security law, one source said. Officials also discussed potential new rules to constrain who may provide funding to domestic AI startups, the same source added.
The precise contours of any restrictions remain undecided, two people who described the talks said. They indicated the measures under consideration might only apply to models developed in the future and that it was not yet clear if or when any new rules would be implemented.
Companies and models mentioned
Alibaba and ByteDance both offer widely used AI models domestically. Alibaba’s Qwen and ByteDance’s Doubao are among the prominent examples cited in the discussions. Z.ai has drawn attention for its GLM-5.2 model, which some observers say approaches leading U.S. offerings in capability but at substantially lower cost. The emergence last year of DeepSeek’s R1 model was cited in the meetings as part of the recent global inroads made by Chinese AI systems, driven in part by lower prices and improving performance.
All three companies represented at the meetings maintain portfolios that include both closed-source models and open-weight releases that users can download, run and customise.
National security concerns and international parallels
China’s deliberations over access to advanced AI echo actions already taken elsewhere, notably in the United States where national security concerns have informed policy decisions affecting AI product availability. Officials in Beijing are reported to be particularly worried about certain foreign-developed models that could be used to exploit software vulnerabilities, and the possibility that such capabilities could be directed against Chinese interests, two of the sources said.
In the U.S., authorities took steps in June that restricted access by foreign nationals to Anthropic’s most advanced Fable and Mythos models. That move led the company to disable the systems for all users temporarily because it could not verify nationality in real time. Export controls for Fable were later lifted after additional safeguards were implemented, while Mythos remains accessible only to a subset of trusted U.S. organisations, according to the account in the meetings referenced by sources.
Domestic policy actions and investigations
Beijing has already enacted a series of measures this year intended to bolster control over homegrown AI technologies. In April, the state planner ordered Meta to unwind its $2 billion acquisition of the AI startup Manus, which was founded in China. Early in June, authorities introduced sweeping new rules tightening oversight of overseas transactions involving Chinese investors, technology, data and national security considerations.
Separately, Chinese regulators have opened investigations into Manus and other local AI startups that relocated abroad to determine whether export control rules were breached, according to two of the sources and a third person. The outcomes of those inquiries and how they might inform any new restrictions on overseas access to Chinese AI models were not detailed by the sources.
Legal proposals and possible frameworks
While no final policy has been announced, there are indications of potential regulatory frameworks being considered. A May roundtable of Chinese legal experts on open-source AI regulation, summarised in an official Supreme People’s Court journal, proposed a tiered approach. That model would categorise technologies so that basic open-source tools would be subject to a simple filing requirement, more advanced technologies would undergo security reviews, and the most sensitive frontier models would be either barred from public release or restricted to domestic use.
How any new restrictions might be implemented in practice - including technical mechanisms, enforcement measures and coordination with existing export controls - was not specified in the meetings as described by the sources.
Potential market implications
Participants in the meetings and the sources who recounted them noted that limiting foreign access to China’s leading AI models could have ripple effects through AI markets globally. Because some Chinese models have gained traction internationally due to their low costs, tighter controls would likely raise costs for businesses relying on those systems. Beyond cost, the discussions suggest Beijing is treating frontier AI as a sensitive national asset that requires protective regulation similar to other strategic technologies.
At present, the commerce ministry and the National Development and Reform Commission did not respond to requests for comment. The companies reported to have attended the meetings - Alibaba, ByteDance and Z.ai - also did not respond to requests for comment, according to the sources who described the meetings.