China's Algorithm Registration Requirement, Explained

China’s algorithm registry predates the generative AI rules by more than a year and binds a much wider set of services. A recommendation feed, a search ranking and a scheduling system are all in scope, and none of them involves a language model.

Where the obligation comes from

The Provisions on the Administration of Algorithmic Recommendation of Internet Information Services (互联网信息服务算法推荐管理规定) were issued at the end of 2021 by the Cyberspace Administration of China with three other departments and took effect on 1 March 2022. The official text is published by the CAC: the Algorithmic Recommendation Provisions. The filing duty is in Articles 24 to 27; the rest of the instrument governs how recommendation services must behave.

Five algorithm types

Article 2 defines algorithmic recommendation technology by enumerating five categories, and the breadth of the list is the point. It covers algorithms for:

• Generation or synthesis (生成合成类) — producing or synthesising content. This is why generative systems were already within the registry before any generative-specific rule existed.
• Personalised recommendation (个性化推送类) — the feed, the “for you” surface, personalised product suggestions.
• Ranking and selection (排序精选类) — trending lists, hot topics, editorial ranking done by algorithm.
• Retrieval and filtering (检索过滤类) — search result ordering and filtering.
• Scheduling and decision-making (调度决策类) — dispatch and allocation systems, the classic example being ride-hailing and delivery assignment.

A single product often runs several. The registry expects them identified, not collapsed into “our algorithm”.

It is worth noting that the filing is only part of what the Provisions impose, because a company that treats them as a registration exercise misses the behavioural half. Providers must regularly review, evaluate and verify the mechanisms, models, data and outcomes of their algorithms, must not set up models that induce addiction or excessive consumption, and must give users a convenient way to turn personalised recommendation off and to delete or alter the tags used to target them. Separate articles add protections for specific groups: services must not use algorithms to harm the physical or mental health of minors, must accommodate older users and guard against telecoms fraud aimed at them, and must protect the legitimate interests of platform workers where algorithms allocate work and set schedules and pay.

There is also a rule against algorithmic price discrimination: providers selling goods or services must not use algorithms to impose unreasonable differential treatment on consumers based on their preferences or transaction habits. That provision predates the more recent enforcement campaigns against “big data killing familiar customers” (大数据杀熟), and it is the article most likely to be cited in a consumer complaint. None of these duties depends on whether the service has public opinion attributes; only the filing does.

What triggers a filing

Article 24 requires a provider of algorithmic recommendation services with public opinion attributes or social mobilisation capability to file within ten working days of beginning to provide the service. The qualifier does substantial work: not every ranking algorithm in China requires a filing, only those on services with the capacity to shape public opinion or mobilise people, which in practice covers most consumer-facing platforms with user-visible content and few purely internal industrial systems.

Article 27 extends the duty to change: where the filed information changes, the provider must file the change, and where the service ends, the provider must file for cancellation. These are short deadlines counted in working days, and they are the part most likely to be missed by a company that treated the initial filing as a one-off project. An algorithm registry entry describing a system that has since been replaced is not a compliant entry.

The filing, the number and the register

The filing is made through the CAC’s internet information service algorithm filing system, an online portal, and Article 24 lists what goes in: the provider’s name, the form of service, the field of application, the algorithm type, an algorithm self-assessment report, and the content intended for publication. The self-assessment report is the substantive document—it is where the provider explains what the algorithm does, what risks it presents and what controls exist around it.

Article 25 gives the regulator a review window measured in working days, after which a completed filing results in a filing number being issued. Article 26 requires providers to publicise the filing number in a conspicuous position on the service; the CAC also publishes batches of filed algorithms, which is why the register is externally observable and why researchers can count entries by company and by category. That publicity is deliberate. The registry is partly a transparency instrument, and the published entries are the only part of it most outsiders ever see.

What the register is not is an approval of the algorithm’s quality or fairness. A filing number records that a description was submitted and accepted as complete. It says nothing about how the system performs, in the same way that an AI management system certificate says nothing about model outputs. Reading a filing number as a regulatory endorsement is the reverse of the mistake foreign readers usually make about Chinese instruments, and it is still a mistake.

Consequences of not filing

The Provisions carry an escalating administrative response rather than a headline penalty: a warning and an order to rectify within a period, with a fine in the range of tens of thousands of renminbi where rectification is refused, and—more consequentially in practice—suspension of information updates or of the service itself. For a consumer service in China, an order to stop updating is a far heavier sanction than the monetary figure suggests.

The other consequence is downstream. Where a generative AI service needs to satisfy the launch precondition described in the generative AI Measures, the algorithm filing is one of the two steps required. An unfiled algorithm is therefore not only a breach of the Provisions but a blocked launch under a different instrument.

Two further Chinese instruments interlock with this one and are worth keeping distinct in your head: the deep synthesis provisions, which add their own filing and labelling duties for synthetic content, and the 2025 content-labelling measures, which set out how the labels themselves must be implemented. The registry is the common substrate—several Chinese AI rules discharge part of their machinery through it rather than building a new one, which is why understanding it once pays for itself across the rest of the Chinese AI regulatory stack.