Information date: 14 September 2026 — Generative AI services offered in China sit inside a framework of algorithmic filing, service registration and content duties, and a company can hold a filing record while the product customers actually use has moved to a different model, a wider capability or a different content flow. The record and the service drift apart quietly as releases accelerate, and the gap is invisible from the customer-facing product. Knowing that statement is not enough for an operating, research or compliance decision. The team must first establish who and what it applies to, how the effect reaches the real process, and which evidence would justify action.
Verified facts and scope
Generative AI services offered in China sit inside a framework of algorithmic filing, service registration and content duties, and a company can hold a filing record while the product customers actually use has moved to a different model, a wider capability or a different content flow. The record and the service drift apart quietly as releases accelerate, and the gap is invisible from the customer-facing product.
The case file should list the model and version actually serving traffic, the capability set as described at filing, the user-facing disclaimers and labelling, the content moderation route and its escalation owner, the complaint and takedown path, the training and data sources documented internally, and the change log that connects product releases to filing updates. Where a release changed capability without a documented review, the file should say so rather than describe the intended state.
How the effect reaches operations
An AI filing is a snapshot of a described service, while an AI product is a moving target. Each release that changes capability, data flow or content handling can move the service outside what was described. Where the change log is not tied to the filing record, nobody can tell whether an update needed a fresh assessment, and the answer is usually found during a customer audit rather than at release time when it would have been cheap to fix.
The typical failure is a capability added for a demonstration and left enabled in production, or a model swap made to cut cost, either of which can change what the service is. Labelling and disclaimer text is then out of date and the moderation route may not cover the new output type. The company also loses the ability to show which version was live on a given date, which is the first question an auditor asks.
For “China AI Service Case: A Filing That Looked Complete and Was Not”, official rules or published findings, direct evidence from the relevant product or process, and assumptions that remain untested should be recorded separately. A broad source defines the external boundary; it does not replace batch records, protocols, contracts, labels or direct observations.
Decision
Link every release to the filing record and give one named owner the authority to block a release whose capability no longer matches the described service. Re-verify the description, labelling and moderation route before major releases rather than after, and keep the version history as the primary evidence rather than the marketing description.
Implementation checklist
- Record the model, version and capability set serving users and diff it against the filing description.
- Tie labelling, disclaimers and the moderation route to a named owner and a release checklist.
- Block releases that change capability, data flow or output type until the filing position is confirmed.
- Assign one decision owner, one implementation owner and a dated review point for “China AI Service Case: A Filing That Looked Complete and Was Not”.
- For “China AI Service Case: A Filing That Looked Complete and Was Not”, archive the source page, access date, applicable population or entity, and internal evidence both supporting and opposing the current decision.
- When a rule, formulation, supplier, protocol or observed result changes, reopen only the affected question in “China AI Service Case: A Filing That Looked Complete and Was Not”.
Evidence and review
For “China AI Service Case: A Filing That Looked Complete and Was Not”, start with one real case rather than an abstract checklist. Record the input version, responsible owner, start time, observed result and stop condition. If the team cannot complete “Record the model, version and capability set serving users and diff it against the filing description.” with current evidence, it should not expand the process to more products, patients, suppliers or markets. The first review should focus only on facts capable of changing the decision.
The second control follows “Tie labelling, disclaimers and the moderation route to a named owner and a release checklist.”. Keep the source date, applicable population or entity, deadline, cost effect and owner in the same evidence file. A wording preference does not justify a new version. A repeated discrepancy, an unsupported health claim or a regulatory mismatch does: correct that point and hold release until the evidence is available.
After “Block releases that change capability, data flow or output type until the filing position is confirmed.”, compare the intended outcome with what actually happened. Apply the same success criteria to each later expansion. If only one number, date or responsibility changes, update that field and the affected conclusion instead of recreating evidence that remains valid. This keeps the decision traceable without turning review into an open-ended rewrite cycle.
Limits of the conclusion
Filing requirements, content duties and enforcement practice depend on the service, the provider and current rules. This case method does not determine whether any particular service needs a filing, and it is not a compliance opinion.
