On 2 April 2024 the Cyberspace Administration of China published a notice on generative AI services that have completed filing, with an annex listing filings for April 2024. The notice is not a new rule. Its significance is that filing has moved from an internal submission to a public list. For companies launching or already operating generative AI services, the framing of the notice matters more than its contents.
Three things the notice shows
Filing is continuous, not a one-off licence. The list is updated in batches, which means launch, a significant model update, or a change of entity or service information can each trigger a further submission.
Filing runs through local channels. Generative AI services with public-opinion attributes or social mobilisation capacity file through the cyberspace authority where they operate. Where the registered address and the operating location differ, confirm which authority receives the file before drafting.
Filing becomes public. Entity name, service name and filing number enter a public list that customers and partners use to verify claims. What is written in the submission is, in effect, a public commitment.
Who needs to file
The test is whether a service has public-opinion attributes or social mobilisation capacity, not company size or parameter count. Services offered to the public that generate content and can spread it usually fall within scope; purely internal systems are assessed differently. Where the boundary is unclear, talking to the local authority is safer than inference.
Materials and security assessment
A filing typically includes entity information, service information, security assessment material and internal policy documents. Reviewers look less at volume than at consistency: whether the service description matches actual functions, whether security measures are operational, and whether model source and training data are explained.
The security assessment is the core. An assessment written from a template cannot support later inspections or incident handling. Align conclusions with actual controls and keep records that can be verified.
Changes, cancellation and timing
Service shutdown, change of entity, renaming, or a substantive change in model capability should each be reflected in a timely filing update. The common failure is careful filing at launch and no updates afterwards, leaving the public record out of step with the live service.
The filing period depends on the quality of the materials and local scheduling; it should not be planned backwards from a fixed number of days. Filing preparation usually requires legal, technical and business input together, and any missing party causes rework. Treat filing as a precondition for launch, not a post-launch formality.
Division of labour with deep synthesis algorithm filing
Generative AI service filing and deep synthesis algorithm filing are separate procedures; neither substitutes for the other. The first addresses the act of providing a service, the second the algorithm mechanism. A service that also offers deep synthesis functions often needs both. The underlying information can be shared, but the framing differs and cannot simply be copied.
For filing preparation, see our algorithm filing service; for training data and provenance, see data compliance. Definitions are in the glossary.
This article is based on the notice on generative AI service filing published by the Cyberspace Administration of China on 2 April 2024, and on public information. It is for reference only and does not constitute legal advice.