A chatbot answers a medical question. A newsfeed quietly decides what you see. An agent (智能体) books a meeting on your behalf. None of these fits neatly inside China's oldest internet rulebook, written in 2000, when "the internet" mostly meant a static website and a dial-up connection.
That gap is now the subject of a concrete legal draft. On July 3, 2026, the Cyberspace Administration of China (CAC) reopened the Internet Information Service Management Measures (Revised Draft for Comment) for public consultation. The headline change is a brand-new "smart information service" (智能信息服务) chapter that, for the first time at the administrative-regulation level, would define a "smart information service provider" and attach hard obligations to anyone using artificial intelligence to publish information.
What the draft actually says
The draft was prepared by the CAC together with the Ministry of Industry and Information Technology and the Ministry of Public Security. It does not replace the existing generative-AI and deep-synthesis rules; instead it pulls them up into the baseline statute that governs every internet information service.
- Disclosure. A smart information service provider must publicly explain the basic principles, purpose, main operating mechanism and training-data sources of its technology, "in accordance with relevant state provisions."
- Safety assessment and filing. Services with opinion-shaping or social-mobilization capability must pass a security assessment and complete algorithm filing, plus change and cancellation procedures.
- Training data. Providers must use lawfully sourced data and base models, respect others' intellectual property, and — for news services — draw corpus only from news published by authorized units.
- Labeling. Generated or synthesized content must carry labels per the mandatory national standard on AI content labeling (生成合成内容标识).
- No forced use. Algorithm-recommendation providers cannot force users onto smart services and must offer an easy way to opt out.
- Labor and agents. Separate clauses protect gig workers from opaque scheduling algorithms, and require agent (智能体) services to use compliant agents with security-management systems.
Why a 26-year-old law matters
The original Measures took effect in 2000 and have shaped everything from licensing to content takedowns. Lifting AI governance into that foundational layer changes the legal footing: instead of sector-by-sector notices, AI obligations would sit in the same statute that applies to every online publisher.
According to Li An, chief scientist at BrainNet (脑机网), tying AI duties to the baseline internet statute gives companies a single compliance reference point rather than a patchwork of separate notices.
How this differs from today's rules
China already regulates AI through layered documents: the Generative AI Interim Measures, the Deep Synthesis Provisions, the AI content-labeling Measures (生成合成内容标识办法) and various algorithm-filing rules. Those are mostly departmental rules and notices. The draft would lift core duties into an administrative regulation — a higher tier in China's legal hierarchy, sitting just below laws passed by the legislature.
That matters for enforcement. A duty written into the Measures carries the weight of the statute that licensing, penalties and platform liability already hang on. It also means AI obligations would travel with the platform, not live in a separate AI-specific silo. For a platform, that means a takedown or penalty today justified under a departmental rule could, after adoption, rest on the same Measures that govern its license — simplifying enforcement but also raising the stakes.
Who falls inside the net
The draft is not only about AI labs. Because it amends the baseline statute, its reach extends to any internet information service provider — news portals, social platforms, e-commerce sites — that layers AI onto what it publishes. A separate "platform information service" (平台信息服务) chapter also tightens duties for influential public-account operators, including network-violence detection and management of multi-channel distribution accounts. In practice, the AI chapter is the headline, but the surrounding revisions are what make compliance a company-wide task rather than a model-team task.
The stage — read this carefully
This is a revised draft for comment (修订草案征求意见稿). The consultation window opened July 3, 2026 and closed on August 2, 2026. The text has not been adopted by the State Council, and it is not in force. Nothing in it is binding today. Any report that calls it "effective" or "enacted" would be wrong.
The CAC said the revision followed the amended Cybersecurity Law and public feedback from an earlier round of comments. So this is a second, refined draft — not a first ask, and not a final rule.
Honest limitations
We verified the draft's existence, issuing body, comment dates and chapter structure against the CAC notice and corroborating legal-media reporting. We have not seen a finalized adoption timetable; Chinese media describe the next step only as "pending further legislative procedure." Exact penalty amounts and final article numbering may shift before any adoption. All figures above describe the draft text, not enacted law, and the consultation has already closed.
What readers can do now
- If you run an AI product in China: map your current obligations — generative-AI filing, deep-synthesis labeling, algorithm registration — against the draft's "smart information service" chapter now. The duties largely echo existing rules, so early alignment is low-cost.
- If you shape policy feedback: the formal window has closed, but provincial and industry associations still collect consolidated positions; route concerns through them before the text advances.
- If you build compliance tooling: the draft's disclosure and labeling clauses are a ready checklist for a single "AI service compliance" dashboard that serves both the old and the proposed regimes.