The law layer · China · Cyberspace Administration of China
China: labelling AI-generated synthetic content
Measures for Labelling of AI-Generated Synthetic Content · Cyberspace Administration of China · in force 1 September 2025, with a mandatory national standard from the same date
Page verified 5 August 2026. Recorded from the published Measures and their English translations, and from contemporaneous analysis. Where a detail is from analysis rather than the text, we say so.
The most prescriptive content-provenance law in the world, and the only one that names what has to be inside the file. Not “mark it somehow”, which is what the EU AI Act says. Provider name. Content reference number. Watermark where feasible. In the metadata.
The two labels
- ExplicitA visible or audible marker added to content generated or synthesised using AI, covering text, images, audio, video and virtual scenes. Examples given include an “AI-Generated” notice placed at the start, at the end, or at another appropriate position in the content.
- ImplicitA machine-readable marker written into the file metadata, including the service provider’s name, a content reference number, and watermarks where technically feasible.
The second one is the interesting one, and it is the reason this instrument belongs at the top of any provenance reading list. A visible caption is a claim on the surface of a file, and it does not survive a screenshot. A metadata marker travels with the file and can be checked by anyone, without asking the sender anything.
Who it binds
- ProvidersInternet information service providers and providers of online content distribution services that generate the content.
- UsersPeople publishing AI-generated content through online transmission services must proactively declare it and use the labelling function the platform supplies. The duty does not stop at the platform.
- StandardA mandatory national standard took effect on the same day, so the technical form of the label is specified rather than left to each provider.
Why it matters outside China
Three reasons, and none of them require operating in China.
- It sets the ceilingWhen regulators elsewhere ask what “machine-readable marking” should actually contain, this is the only existing answer at national scale. The EU AI Act’s Article 50 requires the marking; China specifies the fields.
- It validates the shapeProvider identity plus a content reference plus a watermark is, independently arrived at, the same triple that content-provenance work elsewhere has converged on, including C2PA Content Credentials.
- It reaches supply chainsAny organisation publishing into Chinese platforms, or receiving content produced by a Chinese provider, will encounter labelled files and has to be able to read them.
What the law does not do
It says what must be attached to a file. It does not say who is accountable for the human judgment behind the file, and it does not create a record of what a person checked. A label saying “this was made by AI” and a record saying “a named person reviewed this and here is what they verified” are different documents doing different jobs. China has mandated the first. No jurisdiction has yet mandated the second, though New South Wales comes closest by requiring expert witnesses to annex their prompts.
What to do about it
This is an estate problem, not a document problem.
A label says what made a file. It does not say who stands behind it or what they checked, and where that second question is asked, by a regulator, a court or a customer, a tag is not an answer. Regulayer records the decision as it is taken, outside the system being governed, content-free, and verifiable by anyone holding the file with public mathematics, without a call to us.
Regulayer for enterpriseHumanMark, the provenance surface →
For an individual practitioner, the same engine is $349 a month with a free week: start sealing. See also the law layer and every jurisdiction we track.
Sources
- Measures for Labelling of AI-Generated Synthetic Content, English translation, China Law Translate
- Cyberspace Administration of China, issued March 2025, effective 1 September 2025, with an accompanying mandatory national standard.
- We have not published a section-by-section block for this instrument because we have worked from translations rather than the Chinese original, and say so rather than cite article numbers we cannot verify.
Related
Checked 5 August 2026. Information, not legal advice, and not advice on the law of the People’s Republic of China.
Common questions
What does China's AI labelling law require?
Two kinds of label on content generated or synthesised using AI. Explicit labels are visible or audible markers, such as an 'AI-Generated' notice placed at the start, end or another appropriate position in the content. Implicit labels are machine-readable markers written into the file's metadata, including the service provider's name, a content reference number, and watermarks where feasible.
When did the Chinese AI labelling measures take effect?
1 September 2025, alongside a mandatory national standard that took effect the same day.
Who has to comply with China's AI labelling rules?
Internet information service providers and providers of online content distribution services that create AI-generated content. Separately, users who publish AI-generated content through online transmission services must proactively declare it and use the labelling function the platform provides.
How is China's approach different from the EU AI Act?
The EU AI Act's Article 50 requires that synthetic content be marked in a machine-readable format and that deepfakes be disclosed, but leaves the technical means open. China specifies the content of the implicit label: the provider name, a content reference number and, where feasible, watermarks. It is the more prescriptive of the two.
What is an implicit label?
A marker that a person cannot see but a machine can read, carried in the file's own metadata. It survives being copied or re-shared in a way a visible caption does not, and it lets anyone downstream check where a file came from without trusting the person who sent it.
Information, not legal advice, and not advice on the law of any jurisdiction. Every entry is verified against the issuing body’s own document; where a source is reporting rather than the document, we say so.