China’s labelling rules are the most prescriptive synthetic-content regime currently in force anywhere. They require two labels rather than one — a visible mark for people and a metadata mark for machines — and they distribute the obligation across generators, distribution platforms, app stores and users, so no single party can discharge it alone.
The instrument and its date
The Measures for Labelling of AI-Generated Synthetic Content were issued on 14 March 2025 by the Cyberspace Administration of China together with the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration, and took effect on 1 September 2025. They are published by the Cyberspace Administration of China. A mandatory national standard, GB 45438-2025, was issued alongside them and carries the technical detail: which fields go where, in which file formats, and how they are encoded.
The Measures do not stand alone. They sit on top of three earlier instruments that already applied to the same services: the Provisions on the Administration of Deep Synthesis Internet Information Services, effective January 2023; the Interim Measures for the Management of Generative AI Services, effective August 2023; and the algorithm recommendation provisions of 2022. The labelling Measures give teeth and specificity to labelling duties those instruments had stated more generally. See the deep synthesis provisions and the generative AI service filing requirement.
These are Chinese administrative regulations enforced by Chinese regulators against services provided to the public within mainland China. Whether your service is within scope, and what a filing obligation would mean for you commercially, are questions to put to Chinese counsel. This page is not legal advice, and the authoritative text is the Chinese original rather than any English rendering of it.
Explicit labels: what users see
An explicit label is a human-perceptible mark presented at or around the content. The Measures require it where the synthetic nature of the content could otherwise cause confusion, and the standard specifies placement by modality:
- Text. A mark at the start or end of the text, or somewhere in the surrounding interface, indicating it is AI-generated.
- Images. A visible mark at an appropriate position on the image itself.
- Video. A mark at the start, and prominent marks at appropriate points during playback for longer material.
- Audio. An audible cue or an equivalent prompt, since a visual mark is useless in an audio-only channel.
- Virtual scenes and interactive content. A mark on entry, and persistent or repeated marking where the interaction continues.
The design intent is that a person encountering the content in its normal channel cannot miss the mark. That is a different design goal from the EU’s Article 50(2), which is aimed at machines, and it is why a product built for one regime does not automatically satisfy the other.
Implicit labels: what goes in the file
Implicit labelling is the mandatory metadata requirement, and it is the part with real engineering consequence. The service provider must embed identifying information in the file metadata, and the national standard prescribes the fields — in substance, an indication that the content is AI-generated, the name or code of the service provider that generated it, and a content identification number allowing the item to be traced.
Three practical points follow. The metadata is prescribed rather than free-form, so an existing provenance manifest in some other schema does not discharge the duty by itself; the fields have to be present in the specified form. The provider identity is part of the record, which means the mark reveals which service produced the content — a materially different disclosure from a bare “synthetic” flag. And because the identifier is meant to support tracing, retention of whatever maps the identifier back to a generation event is implicitly part of the arrangement.
The Measures also contemplate encouraging, rather than requiring in all cases, more robust techniques such as watermarks embedded in the content itself. Metadata is strippable, and the drafters plainly knew it: the platform obligations described below are the answer to stripping.
The four-party duty chain
The distinctive structural feature is that the obligation does not stop at the generator.
- Service providers that generate content must apply both the explicit and the implicit labels at the point of generation, and must state their labelling method in their user-facing terms.
- Content distribution platforms must check uploaded material for implicit metadata; where it is present, label the content accordingly to users; where it is absent but the platform detects or the user declares that content is synthetic, apply a label anyway; and record relevant information in their logs. This is the provision that makes metadata stripping a platform problem rather than a loophole.
- App stores and distribution channels must verify that an application offering generative services has a labelling mechanism before distributing it.
- Users must declare synthetic content when publishing it and must not maliciously remove, alter, forge or conceal labels; nor may anyone provide tools whose purpose is to do so.
Enforcement runs through the existing supervisory machinery of the cyberspace and telecommunications authorities rather than through a new body, and it interacts with the filing and security-assessment requirements that generative services in China already carry.
How this differs from the EU approach
Both regimes want synthetic content to be identifiable, and they reach it from opposite ends. The EU’s Article 50(2), applying from 2 August 2026, imposes a machine-readable marking duty on the provider and leaves the technique open, with the human-facing disclosure sitting separately on the deployer under Article 50(4). China requires both marks from the outset, specifies the metadata fields in a mandatory standard, and pushes verification duties onto platforms and app stores that the EU instrument does not address at all.
The consequence for a product shipped in both markets is that compliance is not a single feature. You need a machine-readable mark whose format satisfies the Chinese standard where that regime applies, a visible mark placed per modality, a mechanism for platforms to read your metadata, and — separately — whatever you decide satisfies the EU’s effectiveness and interoperability test. Building one and assuming it covers the other is the predictable mistake. For the EU side see the Article 50(2) marking duty, and for the wider Chinese framework these Measures sit inside, China AI regulation.
Top comments (0)