China's Cyberspace Administration (CAC) is rewriting the legal foundation of its internet regulation for the first time in a generation. On July 3, 2026, the CAC — working with the Ministry of Industry and Information Technology and the Ministry of Public Security — reopened public comment on a revised draft of the Internet Information Services Management Measures, the framework statute that has governed Chinese websites since 2000. The comment period closed August 2, 2026. The draft expands the current 27-article measure to six chapters and 94 articles, nearly double the 54-article version China floated and abandoned in 2021.
The headline provisions concern algorithmic recommendation systems and, for the first time, autonomous AI agents. Article 61 requires providers of algorithm-driven "intelligent information services" to "establish human intervention and user self-selection mechanisms" and to give users a way to receive a feed not based on personal characteristics, or a convenient means to refuse personalized recommendations outright. Article 63 extends the regime to AI agents (智能体服务) for the first time, requiring operators to build "security management systems and technical control capabilities" and barring agents from being used in ways that endanger national security or public order. Providers must also file their traffic-allocation rules and recommendation algorithms with authorities and publicly disclose the underlying logic — formalizing an obligation that, in practice, has existed informally since 2021.
The Steelman: This Isn't Coming From Nowhere
Critics of Chinese tech regulation often treat it as a monolith of control, but the substantive case for mandatory opt-outs and algorithm disclosure is one plenty of democracies have made too. Engagement-optimized feeds have well-documented effects on compulsive use, especially among minors, and users in most jurisdictions currently have no practical way to turn personalization off short of quitting a platform. A recommendation system that can quietly determine what a billion people see has no real analogue in pre-digital media law, and treating it as unregulatable-by-default is its own kind of ideological choice. The EU's Digital Services Act (Article 27) already requires large platforms to offer a non-profiling feed option, and the EU AI Act's transparency obligations for high-risk systems rest on a similar theory: opacity itself is the harm. On the AI agent provisions specifically, autonomous software making purchases, filing complaints, or interacting with other systems on a user's behalf genuinely raises new liability and security questions no existing statute anticipated. A regulator moving first here is not obviously wrong to do so.
Where the Draft Diverges From Comparable Regimes
What distinguishes China's approach is less the opt-out mandate itself — which mirrors the DSA — than the compliance architecture wrapped around it. The 2021 Algorithm Recommendation Provisions already required registration of algorithms with "public opinion properties" in a CAC-run registry; the first disclosed batch, in August 2022, held just 30 entries from Tencent, Alibaba, ByteDance, and Weibo, filed at a level of abstraction the Carnegie Endowment for International Peace described as "devoid of meaningful detail." That registry has since scaled dramatically: Trivium China's analysis of CAC filing data found 3,739 generative algorithmic tools registered by roughly 2,353 companies as of April 2025, growing 250-300 entries a month, concentrated overwhelmingly in Beijing, Guangdong, Shanghai, Zhejiang, and Jiangsu. The new draft's Article 74-89 liability chapter backs this with fines reportedly ranging from RMB 50,000 to RMB 10 million, plus suspension or account closure — a scale of exposure that makes the filing regime a genuine market-access gate, not a disclosure formality.
That's the real distinction from the EU model: Brussels requires disclosure of the opt-out; Beijing requires pre-clearance through a state registry before a recommendation system or AI agent can operate at all, with the same authority that approves the filing empowered to revoke it. A platform that gets crosswise with Beijing on an unrelated matter — content moderation, foreign investment screening, an executive's public statements — now has its algorithm's regulatory standing as fresh leverage against it. The opt-out right is genuinely pro-consumer; the registry underneath it is a licensing chokepoint dressed in consumer-protection language, and the two shouldn't be conflated when evaluating the draft.
Why the Timing Matters
The draft lands as China's AI sector is trying to project openness abroad. Moonshot AI released its Kimi K3 model as open-weight on July 27, 2026 — letting any government or company run and retrain it locally, a pitch aimed squarely at states wary of dependence on U.S. cloud providers, per Rest of World's reporting. A domestic compliance regime built on mandatory registries and content controls sits awkwardly next to that pitch: governments evaluating Chinese open models for sovereignty reasons are also watching how Beijing regulates the same technology at home, and a system where every commercially significant AI service needs pre-clearance from a security-focused regulator is a data point for those calculations, not a footnote.
What to Watch
The comment period closed August 2, 2026; a final text and effective date haven't been announced. The details worth tracking are which categories of algorithm trigger mandatory filing versus disclosure-only obligations, how narrowly "endangering national security or public interest" is defined for AI agents, and whether the human-intervention mandate becomes a genuine default-off option or a buried toggle — the difference between real user control and compliance theater. Proportionate transparency rules for opaque recommendation systems are a legitimate regulatory goal that other jurisdictions are pursuing too. Whether this draft achieves it, or simply extends state licensing power over a market Beijing is simultaneously trying to sell as open to the rest of the world, will depend on implementation details that remain deliberately vague in the version published for comment.