China's AI content labeling measures: explicit and implicit labels since September 1, 2025
China's Measures for Labeling AI-Generated and Synthetic Content took effect on September 1, 2025. The text separates explicit labels, which people can see, from implicit labels, which sit in the file's data. Article 5 requires providers to add implicit labels in file metadata. Our checker has no reader for those fields.
What are the Measures, and when did they start?
The notice on the Measures for Labeling AI-Generated and Synthetic Content is on the Cyberspace Administration of China's site. The page names four issuing bodies: the Cyberspace Administration, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. It shows a signing date of March 7, 2025 and a publication date of March 14, 2025.
The text has 14 articles. Article 14 sets the start date: September 1, 2025.
What is the difference between explicit and implicit labels?
Article 3 defines the two kinds. In our own words:
| Kind | What Article 3 says | Where it sits |
|---|---|---|
| Explicit label | A marking in text, sound or graphics that users can readily notice. | In the content or in the interface where it is shown |
| Implicit label | A marking added by technical means that users typically cannot easily notice. | In the data of the content file |
The text covers text, images, audio, video and virtual scenes.
What do the Measures require for images?
Article 4 covers explicit labels. It applies to generation services in the situations of Article 17, paragraph 1 of China's deep synthesis provisions; we did not read that provision, so we do not list which services it reaches. For images, Article 4 calls for a clear marker in a suitable place. It also says that when a provider offers download, copy or export, the file must contain a compliant explicit label.
Article 9 adds a limit. A provider may supply content without an explicit label when a user asks for it, once the user agreement sets out the user's labeling duty and responsibility. The provider must then keep related logs for at least six months.
Article 5 covers implicit labels. It says providers must add them in the metadata of the files they generate. It says the label carries production details, with an attribute showing the content was generated, the provider's name or code and a content number as examples. It also encourages digital watermarks. In the same article, file metadata means descriptive information embedded in the file header in a set encoding format.
What do platforms and users have to do?
Article 6 covers platforms that distribute content. As we read it, a platform checks the metadata for an implicit label. If the metadata shows the content is AI-generated, the platform adds a prominent notice. If there is no such label but the user declares the content as AI-generated, the notice says it may be. If neither applies but the platform finds an explicit label or other traces, the notice says it is suspected.
Article 10 asks users who publish AI-generated content to declare it. It also says no person or organization may maliciously delete, alter, forge or hide the labels the Measures require. Article 13 says violations are handled by the cyberspace, telecommunications, public security and radio and television authorities under existing laws and regulations. The Measures name no fines, and we do not describe enforcement in practice.
What is GB 45438-2025?
Article 11 says labeling must also follow other laws and mandatory national standards. A mandatory national standard on the same subject is GB 45438-2025. China's national standards platform lists it as a mandatory national standard, in force, issued on February 28, 2025 with an implementation date of September 1, 2025. In our English, its title is "Cybersecurity technology: method for labeling AI-generated and synthetic content".
We could not open the standard's full text. The field names below come from a catalogue page for the standard, which is a secondary source, so we write them as reported. It is reported that Appendix E specifies an extension field whose name includes "AIGC", holding a string with these keys:
| Key | Reported meaning |
|---|---|
| Label | Whether the content is, may be, or is suspected to be AI-generated (values 1, 2 or 3) |
| ContentProducer | Name or code of the generation service provider |
| ProduceID | The provider's unique number for the content |
| ReservedCode1 | Reserved for the provider's own security measures |
| ContentPropagator | Name or code of the service that spread the content |
| PropagateID | That service's unique number for the content |
| ReservedCode2 | Listed in the format; we read no description |
The same page is reported to say that only one metadata implicit label should be kept in a file. We did not verify the namespace or the layout inside an image file.
What did we measure?
We could not open the primary text of the standard, so we tested only what we can say about our own checker. On October 30, 2026 we made one JPEG whose unsigned XMP carries those seven field names, with invented values, under a namespace we made up. Our checker reads it as "Metadata found; source not established", with the line "The file carries metadata, but none of it says how the image was made." It shows no finding, because it has no reader for these fields.
We also hold one real file filed in the Wikimedia Commons category "Images generated by Seedream", uploaded by a third party. We cannot know how it was handled before upload. It reads "No supported provenance signals found"; all three checks ran on it.
These are results about our checker, not about the rules, the standard or any product. Our accuracy page explains how we test, including the real and made files.
What this means when you check an image
Open the checker and choose the file. Our test file with only XMP of these field names read "Metadata found; source not established". The headline says the file holds metadata, not how the image was made, because we do not read these fields. We tested one such file, so we cannot say how a file from a real service would read.
If you see "No supported provenance signals found", we found no signed record, no source label in the metadata and no Stable Diffusion watermark pattern. That does not mean the image is real or human-made, or that a label was never added; a label can be lost when a file is shared or saved again. The checker's limits line applies: "These checks cannot determine whether an image is AI-generated."
Guides on this topic
- Visible AI watermarks is the main page on this topic, and it mentions the reported explicit-label requirement.
- EU AI Act Article 50 covers the European rule on marking and disclosure.
- California's AI Transparency Act covers latent disclosures and the dates.
- Where AI declarations live in a file maps XMP, EXIF, IPTC, PNG text and C2PA.
- Platform labels compared shows how platforms decide on a label.
Frequently asked questions
- Do the Measures require a visible label on images?
- For services in the situations Article 4 names, it calls for a clear explicit marker on images. Article 9 allows content without one when a user asks and accepts the duty. We give our own English, so check the original for exact wording.
- Does the metadata label prove an image is AI-generated?
- No. Metadata is data in a file and can be missing, changed or copied. Article 10 itself bars anyone from maliciously deleting, altering, forging or hiding the labels.
- Does our checker read the China implicit label?
- No. It has no reader for those fields, and it shows no finding for them. Our test file that carries them read as other metadata.