On 31 October 2026 the Cyberspace Administration of China (CAC) published its 19th batch of domestic deep-synthesis service algorithm filings. The announcement gives no count of services. The full list sits in the CAC filing system at beian.cac.gov.cn. Taken alone it is a routine notice. Taken as the latest entry in a four-year series, it shows how China regulates recommender systems and generative AI: through a standing registry that works like a licensing gate.
What the registry requires
The legal basis is two instruments. The Internet Information Service Algorithmic Recommendation Management Provisions (CAC Order No. 9, signed 31 December 2021, in force 1 March 2022) require providers with "public opinion properties or social mobilization capability" to file within ten working days of launching service. The official text on gov.cn confirms the effective date and the Article 24 filing duty. Stanford DigiChina's translation adds the mechanics. Filing changes are due within ten working days, and cancellation within twenty working days of ceasing service. The CAC is to grant filing within thirty working days. Providers must display their filing number prominently. False filing carries a fine of 10,000 to 100,000 yuan.
The Provisions on the Administration of Deep Synthesis in Internet Information Services extended this to generative AI. They took effect on 10 January 2023. Article 19 of the official CAC text requires deep-synthesis providers with public-opinion or social-mobilisation capability to complete filing, change and cancellation procedures under the recommendation rules. Technical supporters must follow the same procedure.
The strongest case for the registry
Regulators have a serious argument. Recommender systems decide what hundreds of millions of people see, and users have little way of knowing how. A registry gives the state a named responsible entity for each algorithm. It makes filing numbers visible to users. It also creates a lever for enforcement without requiring a new statute for every product category. Several jurisdictions are now debating algorithmic transparency duties. China's registry is the oldest working example of mandatory algorithm disclosure to a regulator.
Where the design falls short
The trouble is that this registry discloses almost nothing to the public. The CAC's running index of filing announcements points readers to beian.cac.gov.cn and publishes lists of algorithm names and filing numbers. When the first batch appeared in August 2022, Linklaters counted 30 algorithms from 24 companies, used across 37 apps, 11 websites and 4 mini programs. The public gets names and numbers. It does not get the substance a researcher, user or competitor would need: how a system ranks content, what it optimises for, or what it was tested against.
The 19th batch continues that pattern. The announcement carries no count, so outside observers cannot measure approvals against applications, rejections, or review times. Our reading of the public record is that the 30-working-day review clock and the filing duty are the two measurable commitments. Neither can be checked against published outcomes.
That matters for three reasons.
- Cadence is a signal, but a weak one. The public tracker Digital Policy Alert logs the 18th batch on 17 July 2026, with earlier batches on 12 March and 6 May 2026. That implies a batch every two to four months. A queue that moves in batches means a product launching just after a cut-off waits for the next one. Firms that cannot wait must either launch without a filing number or delay.
- The trigger is elastic. "Public opinion properties or social mobilization capability" is not defined by a numerical threshold. Providers cannot predict with certainty whether they fall in scope, and the regulator keeps discretion over the line.
- Large firms absorb the fixed cost. Preparing a filing, maintaining a compliance team, and tracking change notices within ten working days are cheap for a large platform. They are a real burden for a start-up shipping weekly model updates, where each material change can mean a fresh filing.
A proportionate alternative
We do not argue for abolishing registration. Knowing who operates a high-reach recommender is a legitimate objective, and many democracies are moving toward some version of it. We argue for a registry that is honest about what it is. Four changes would keep the accountability benefit and cut the costs.
- Publish aggregate statistics with each batch: services filed, approved, returned and withdrawn. A batch announcement that omits its own count invites suspicion that the real numbers are unflattering or simply unmanaged.
- Define the scope trigger with usage thresholds, as the EU does for very large platforms, rather than leaving it to case-by-case judgment.
- Separate filing from approval. If a filing is a notification, it should take effect on submission. If it is a licence, the law should say so and attach appeal rights.
- Publish structured, reusable data. A machine-readable registry would let researchers and journalists test whether the system works.
What to watch
The larger question is whether this model travels. Governments weighing algorithm-transparency rules will look at China's registry as a precedent. They should take the right lesson. A registry can create accountability, but only if its outputs are public and its rules predictable. Without those, it works as a rolling permission gate: it is cheap for incumbents, slow for entrants, and opaque to the people it claims to protect. The 19th batch changes none of that. It simply reminds observers that, four years in, the numbers that would let anyone judge the system are still missing.