The obligation follows the service, not the shareholder
Foreign founders often assume that a Chinese entity changes the regulatory position. It does not. Under the Algorithm Recommendation Provisions and the Generative AI Measures, the obligation attaches to the provider of a service offered to the Chinese public. A wholly foreign-owned enterprise, a Sino-foreign joint venture and a domestic company are treated alike once the service is live in China.
What foreign ownership does change is the path to being able to offer the service at all. Value-added telecommunications licensing carries foreign ownership limits for certain categories, and the ICP filing or licence attached to the domain is a precondition for filing. Entity structure therefore tends to determine the timetable before any regulatory question is reached.
Two tracks, not one
The filing landscape is commonly described as a single obligation. In practice there are two tracks, and a third process that sits alongside them.
| Track | Applies to | Statutory basis |
|---|---|---|
| Algorithm filing | Services with public opinion attributes or social mobilisation capacity deploying recommendation, ranking, synthesis, search or scheduling algorithms | Algorithm Recommendation Provisions |
| Generative AI filing | Providers offering generative AI services to the public, where the model is self-developed or substantively modified | Generative AI Measures, Art. 17 |
| Security assessment | The same category of generative AI services, before launch | Generative AI Measures, Art. 17 |
An application built on a third-party model that has already been filed usually takes the registration route instead of the generative AI filing route. Registration is not a simplified filing; the materials and the review focus differ. Treating a substantively modified model as an application, or the reverse, is the most common category error and usually results in the materials being returned.
What a foreign-owned entity must assemble
The materials themselves are ordinary. The difficulty is that several of them depend on decisions taken much earlier.
Entity and domain. A domestic entity, a domain registered in its name, and the corresponding ICP filing or licence. Where the licence is constrained by foreign ownership limits, the structure must be settled first.
Model provenance. Whether the model is self-developed, fine-tuned, or called through an API determines the track. Continuation pre-training, large-scale instruction tuning on proprietary data, or a change of base model all count as substantive modification; application-layer packaging around a filed model ordinarily does not.
Corpus authorisation. Traceability of training data, including licence terms, procurement records and the exclusion of sources that cannot be authorised. A filter log alone does not establish provenance.
Content safety. Mechanisms for input and output moderation, complaint handling and user appeal, documented rather than asserted.
Labelling. Generated content must carry a detectable identifier in accordance with the labelling rules, and the identifier must not interfere with normal use.
Security assessment report, where required. Prepared before launch, not after.
A realistic sequence
For a newly established entity, the work tends to run in this order:
- Entity and capital structure, including whether the intended service can be licensed at all.
- Product scoping: which functions are offered to the public, and which algorithms are involved.
- Model route decision: self-developed, fine-tuned, or registered on a third-party model.
- Materials: corpus records, moderation design, labelling implementation, security assessment.
- Filing or registration submission, followed by iteration on reviewer comments.
- Launch, then ongoing obligations.
Four to six months from incorporation to approval is a realistic estimate where the product is stable. Where the security assessment is required, it sits on the critical path: it cannot be accelerated by adding resources, and starting it late moves the launch date rather than the workload.
Where filings fail
Six points account for most delays.
Category misjudgement. Filing a modified model as an application, or registering a self-developed model, forces a restart.
Entity not ready. Filing before the domain and ICP position is settled produces a submission that cannot be completed.
Corpus records absent. Procurement, licence and exclusion records are typically not kept during development, and cannot be reconstructed afterwards.
Moderation described but not built. A policy document without an operating mechanism is treated as an assertion.
Labelling unimplemented. The obligation is technical, and reviewers look for the mechanism rather than the statement.
Changes after approval not filed. Information changes must be handled within ten working days, and service termination requires deregistration within twenty. Products taken offline without deregistration tend to surface during annual verification.
What cannot be delegated
Drafting can be delegated. The decisions behind the draft cannot. Which model route is taken, what the corpus contains, how moderation actually operates, and whether the entity structure permits the intended service are judgements the company has to make and be able to explain. A filing prepared without those decisions produces documents that survive review only until the first follow-up question.
Related reading: Algorithm Filing: The Complete Process, Algorithm Filing and LLM Filing: Which Comes First, Data Export: Three Routes and How to Budget.
Frequently Asked Questions↑ Back to article
Does a foreign-owned company need to complete algorithm filing in China?
Yes, if it provides a service to the Chinese public that falls within the statutory categories. The obligation attaches to the service and its provider, not to the nationality of the shareholder. A WFOE, a joint venture and a domestic company are treated alike once the service is offered in China.
Can filing be started before the Chinese entity is established?
No. The filing is made by the entity that provides the service, and the filing system requires a domestic entity, a domain and the associated ICP filing or licence. Entity establishment should therefore be the first step, and it usually determines the rest of the timetable.
If we call an already-filed Chinese model through an API, do we still need to file?
Not necessarily. Where the model is used without substantive modification, the application-layer service ordinarily follows the registration route rather than the filing route. The distinction turns on whether you continue pre-training, fine-tune on your own data at scale, or change the base model.
How long does the process take in practice?
For a newly established entity with no prior filings, four to six months is a realistic estimate from incorporation to approval, assuming the product is stable and materials are complete. Security assessment, where required, sits on the critical path and cannot be compressed.
What if the service is offered only outside China?
A service not offered to the Chinese public generally falls outside the filing obligation, even if developed in China. The analysis turns on the target of the offering — language, payment, marketing and user base — rather than on where the servers sit.
Are filings made public?
Algorithm filings are published in the CAC filing system with the entity name, algorithm name and application scenario. Generative AI filings are announced in batches. Competitors and customers can therefore see that a filing exists, which makes accuracy in the published scope worth attention.