On 14 September 2026, the Cyberspace Administration of China (CAC) released its filing information for generative AI services covering July and August. According to Digital Policy Alert's log of the release, it lists 124 newly registered services, 7 of them mobile-side. On 23 September the CAC added 3 more mobile-side services. The release looks routine, but it shows how China's algorithm registry operates: as a recurring, bureaucratic gate on who may offer AI and recommender services to the public.
The strongest case for the registry
The case for a filing system is real. Recommender systems shape what hundreds of millions of people read, buy and believe. A regulator that has no list of who runs them cannot audit anyone. The 2021 Provisions on the Administration of Algorithm Recommendation of Internet Information Services (CAC Order No. 9, in force 1 March 2022) also contain user protections that many jurisdictions have not adopted. Article 17 of the Provisions requires providers to offer an option that does not target a user's personal characteristics, or a convenient way to switch recommendations off, and to stop immediately once the user does. A registry is also a cheap way to find out who is operating in a market.
That is the best version of the argument, and it should be taken seriously. The question is whether the instrument China built delivers it.
What the filing actually is
The filing duty sits in Article 24 of the same Provisions. Providers of recommendation services "with public-opinion attributes or social mobilization capability" must file through the algorithm filing system within ten working days of starting service. They must submit their name, service form, application domains, algorithm type, an algorithm self-assessment report, and the content they propose to publicize. Changes must be filed within ten working days, and cancellation within twenty working days of ending service.
The 2023 Interim Measures for the Administration of Generative AI Services (Order No. 15, in force 15 August 2023) extended this to generative AI. Article 17 says that providers of generative AI services with public-opinion attributes or social mobilization capability must carry out security assessments and complete algorithm filing, change and cancellation procedures under the algorithm Provisions.
The trigger is the phrase "public-opinion attributes or social mobilization capability." It is not defined by risk to users, such as discrimination, safety failures or fraud. It is defined by a service's potential to move public opinion. A registry built on that trigger is a tool for managing information flows, and consumer protection is secondary to that purpose.
The scale is the story
The cadence matters more than any single release. The CAC's May–June 2026 batch covered 120 new services. The July–August batch covered 124. In December 2025, Xinhua reported that more than 700 generative AI large-model products had completed filing, according to the CAC. Throughput that steady means filing is a permanent step in the product lifecycle for any company that wants to serve the Chinese public. It is not a one-time registration drive.
Two features of this design are worth noting for anyone building AI products.
- Approval before launch in practice, disclosure after launch on paper. The Provisions speak of filing within ten working days of starting service, but the published lists show which services have been cleared, and unlisted services are exposed to enforcement. Whatever the legal text says about timing, the practical effect is a licence.
- The disclosure runs one way. The regulator receives the algorithm type, the self-assessment and the content to be publicized. The public sees a list of names. Users learn that a service is registered, and they learn nothing about how it ranks or generates content.
Why this is the wrong model for proportionate regulation
A pro-innovation approach would tie obligations to demonstrable harm and keep the compliance cost low enough that small developers can enter. China's registry does the opposite. It places the burden on every provider with an opinion-shaping service, whatever its actual risk. A medical-imaging model and a chatbot face the same question about their potential influence on public discourse, not about their error rates.
That has two costs. First, it favours incumbents with compliance teams, since the self-assessment report and security assessment need staff and legal support. Second, it chills speech at the design stage: a developer who knows the regulator reviews what the service will publicize has every reason to under-build. Counting registered services as an achievement mistakes permission for safety.
There is also a lesson for democracies tempted to copy parts of the model. The user-facing rights in Article 17, such as opting out of personalized recommendation, are a transferable idea. They can be enforced through ordinary consumer and competition law, with disclosure aimed at users and independent auditors rather than at a licensing authority. The registry is the part that does not transfer well, because its logic depends on a regulator deciding which speech-shaping services may exist.
What to watch
Three things will show where this goes. First, whether the 7 plus 3 mobile-side filings signal a push to make app-level filing a distinct step for AI features inside phones and apps. Second, whether the published lists ever include the content of the self-assessment, which would turn a licensing list into a transparency tool. Third, whether the pace of 120-odd services per two-month cycle holds as the market consolidates. A falling count would suggest either saturation or tighter gatekeeping, and the published data alone cannot tell the two apart.
For now, the 14 September release is best read for what it shows about the architecture. The registry is steady, large and permanent, and it governs by approval. Democracies should borrow the user's right to switch off personalization and leave the licensing gate where it is.