A draft circulating in Beijing this month has rattled parts of the AI industry. On October 16, 2026, the Ministry of Commerce (商务部) and the Ministry of Science and Technology (科技部) published Announcement No. 48 of 2026, opening a public comment period on a revision to the Catalog of Technologies Prohibited or Restricted from Export (《中国禁止出口限制出口技术目录》). The comment window runs from October 16 to November 15, 2026. Crucially, this is a draft for comment — not a law in force. Nothing in it binds exporters yet.
The proposal would add 40 technologies across three sectors — artificial intelligence, biomedicine (生物医药), and photovoltaics (光伏). Within AI, the draft lists 23 technologies spanning large models (大模型), intelligent agents (智能体), computing-power clusters (算力集群), and information processing. The other 17 fall in biomedicine (5) and photovoltaics (12).
Key takeaways
- This is a draft for comment (征求意见稿), open until November 15, 2026; it is not yet effective and could still change before adoption.
- It would add 23 AI-related technologies to the export-control catalog, covering large models, intelligent agents, compute clusters, and information processing.
- The aim is framed as "safe and controllable" AI: the listed items include AI-driven malicious-code generation and AI-driven vulnerability discovery, not ordinary commercial model APIs.
- Exemptions and conveniences matter as much as the restrictions: open-sourced or academically published AI tech is excluded, and several outbound scenarios get fast-track "automatic licensing."
What is actually being listed
The draft sorts items into two buckets. A set proposed for export prohibition includes AI-driven malicious-code generation and mutation technology (人工智能驱动的恶意代码生成与变异技术) and automated evasion-variant batch generation. A set proposed for export restriction includes AI-driven vulnerability discovery (人工智能辅助漏洞挖掘技术) and distributed-training fault tolerance (分布式训练容错技术). Officials have described the AI portion as a mix of proposed bans and proposed restrictions; the precise per-item split sits in the annex we could not fully verify line-by-line, so we present it only as a mix.
The ministries are explicit that this is routine, not a pivot. China has adjusted the catalog repeatedly since 2001, and the spokesperson said the revision "does not target any specific country, region, or entity." Officials note that major economies have imposed varying export controls as AI, biotech, and PV technologies spread into sensitive uses.
What is NOT being controlled
The fine print is where the draft gets interesting — and where panic overstates reality. The commerce ministry clarified that providing model API calls, app or web interaction, and SaaS services to overseas users is not treated as technology export. Technology that has entered the public domain through open source or published academic papers is excluded from management. And several outbound scenarios — Chinese firms building and operating their own compute centers overseas, taping out chips abroad, joint R&D using overseas resources, and joining international standard-setting — qualify for automatic-licensing conveniences that are effectively "file and go" (即报即出).
A temporary measure runs alongside the comment period: from the announcement date until the catalog is formally revised and implemented, exporters of the listed technologies must register beforehand on the ministry's trade-management platform. This pre-export registration (出口前登记) is a holding step, not the final rule.
How this fits China's broader toolkit
The catalog sits alongside other controls rather than replacing them. Technology export in China is already governed by the Foreign Trade Law (对外贸易法) and the Regulations on Technology Import and Export (技术进出口管理条例); the catalog is the list that triggers licensing. Sensitive AI models and algorithms are also caught by the existing generative-AI service filing system and the algorithm-reporting rules, though those regulate domestic operation rather than outbound transfer. The draft therefore layers a new, more specific boundary on top of a system that already exists — which is why industry lawyers stress reading it together with, not instead of, the older frameworks.
Why it matters
For AI labs and startups with overseas users, the practical question is whether a model "going abroad" — local deployment overseas, or providing model invocation abroad — counts as technology export. The ministry's answer, as reported, is that ordinary commercial service delivery generally does not; the controls target transfers of underlying know-how, ownership, or control to foreign parties, and sensitive dual-use capabilities. The draft therefore reads less like a "China AI ban" and more like a boundary-drawing exercise around the most sensitive cores.
Honest limitations
We verified the announcement number, the comment window, the 40/23/5/12 sector split, and the exemption language against the official MOFCOM notice and Xinhua/China News reporting. The detailed per-item ban/restriction split within the 23 AI technologies comes from secondary summaries of the annex; we did not independently line-count the annex and present it only as "a mix." The draft may be revised before adoption, so any specific item list is provisional. We have not assessed how foreign governments or companies will respond, and we did not quantify trade exposure.
What readers can do now
- If you export or license AI technology from China, treat the current period as "register first" for the listed items and submit comments before November 15, 2026 if the scope affects you.
- If you run overseas model services, document that your delivery is API/SaaS/app interaction — the draft explicitly excludes ordinary commercial service from the technology-export definition.
- If you follow China policy, watch for the formal revision and implementation date after mid-November; only then do restrictions become binding.