China's AI content labelling measures: visible and metadata labels

Who does it apply to?

Article 2 applies the Measures to "network information service providers" that fall under three earlier regulations: the algorithm recommendation provisions, the deep synthesis provisions and the interim measures on generative AI services. Those regulations cover services offered to the public in mainland China (from memory; the three texts were not re-read for this page). A company that serves no users in China is outside them. A company that does is the "service provider" whether it trained the model or calls someone else's.

Three other groups have duties: platforms that distribute content (Article 6), app stores, which must ask at review whether an app offers generation and check its labelling materials (Article 7), and users, who must declare generated content when they publish it and use the labelling function the platform provides (Article 10).

What must a provider that generates content do?

Add an explicit label where the service is one of those listed in Article 17(1) of the deep synthesis provisions. We did not re-read that article; Annex B of the national standard, which is informative, lists the typical cases as services that "may cause public confusion or misidentification": chat and writing that imitate a person, synthetic or cloned voice, face generation and swapping, immersive scenes, and text-to-image, music and text-to-video or image-to-video generation. Article 4 sets the placement by medium: for text, a text or symbol notice at the start, the end or a suitable point in the middle, or a prominent notice in the interface; for audio, a spoken or rhythmic cue at the start, end or middle; for images, a prominent label at a suitable position; for video, a prominent label on the opening frame and around the player. When the service lets users download, copy or export the content, the provider must "ensure the file contains an explicit label that meets the requirements".

Add an implicit label to the file metadata. Article 5 requires it to contain "production element information such as the attribute of the generated content, the service provider's name or code, and a content number". The same article says providers are "encouraged" to add implicit labels in the form of digital watermarks. Metadata is defined as descriptive information embedded in the file header in a specific encoding format.

Say how in the user agreement. Article 8 requires the agreement to state the method and style of the labels and to prompt users to read the labelling requirements.

Keep logs if the visible label is removed on request. Under Article 9 a user may ask for content without the explicit label. The provider may supply it after the user agreement has made the user's labelling obligations and responsibility clear, and must keep the relevant logs, including who received the content, for at least six months.

What must a platform that distributes content do?

Article 6 sets three cases, each ending in a prominent notice next to the published content:

1. The metadata carries an implicit label marking the content as generated: the platform tells the public it is generated content.
2. No implicit label is found, but the user declares the content generated: the notice says it may be generated.
3. No implicit label and no declaration, but the platform detects an explicit label or other traces of generation: it is treated as suspected generated content and labelled so.

In all three the platform adds its own "distribution element information" to the metadata: the content's attribute, the platform's name or code, and a content number. It must also offer users a labelling function and remind them to declare.

What does the national standard add?

The form of both labels. GB 45438-2025 was issued on 28 February 2025 by the State Administration for Market Regulation and the Standardization Administration, took effect on 1 September 2025 with the Measures, and is mandatory. We read the full Chinese text on 13 October 2026; clause numbers are given so each point can be checked.

It applies to "generation service providers", defined as those offering generation to the public "including by providing programmable interfaces", and to content distribution providers (clauses 1 and 3.7). An API is expressly within it.

The visible label must contain two elements: "人工智能" or "AI", and "生成" (generated) or "合成" (synthesised), or both.

Medium	Form and placement	Minimum
Text (5.1)	Text or a superscript "AI" at the start, end or a suitable midpoint	Clearly legible type and colour
Image (5.2)	Text at an edge or corner	Text height at least 5% of the shortest side
Audio (5.3)	Spoken notice at normal speed, or a "short-long, short-short" rhythm, which is "AI" in Morse code; at the start, end or midpoint	Clearly distinguishable
Video (5.4)	Text at an edge or corner of the opening frame; optionally at the end or midpoint	5% of the shortest side; at least 2 seconds
Interface (5.6)	Text shown continuously near the content, or at the top, bottom or background of the interface	Clearly legible

The metadata label is set by Annex E, which is normative. The file's metadata gets an extension field whose "name or keyword shall contain AIGC", and its value is a string in this format:

{"AIGC": {"Label": "1", "ContentProducer": "…", "ProduceID": "…", "ReservedCode1": "…",
          "ContentPropagator": "…", "PropagateID": "…", "ReservedCode2": "…"}}

• Label is "1" where the content is generated, "2" where it may be, and "3" where it is suspected to be. Those are the three cases of Article 6 of the Measures.
• ContentProducer is the generating provider's "name or code"; ProduceID is that provider's unique number for the content.
• ContentPropagator and PropagateID are the same pair for the distributing platform. When the generator writes the label first, it puts its own values in both pairs (Annex E, note 1).
• ReservedCode1 and ReservedCode2 may hold information each party uses to protect the integrity of the content and the label. Annex F.4 gives the example of a digital signature over the metadata stored in ReservedCode1.
• A file keeps "only one" metadata label (clause 6.1(c)), so a platform rewrites the field in place.

Watermarks are permitted and not required: clause 6.2 says a content-level implicit label "may take the form of a digital watermark", and Annex A says such labels are "not required in this document".

Three things follow for anyone who also signs with C2PA. First, the standard does not mention C2PA, and a signed manifest is not the Annex E field, so a file for the Chinese market needs the AIGC field whatever else it carries. Second, the field has to be written before the manifest is signed, because the signature covers the file's other bytes. Third, when a Chinese platform rewrites the field to add its own name and number, as the Measures require, those bytes change and the manifest's binding to the file will no longer validate. The third point is our analysis of how the two formats interact, not something we have tested: UNVERIFIED.

What we have still not read are the TC260 practice guides: the provider coding rules of 14 March 2025, which say how a provider code is formed, and the six guides put out for comment on 23 June 2025, which say where the field goes in each text, image, audio and video format. The standard itself accepts a name in place of a code.

What happens if you don't?

The Measures carry no penalty schedule. Article 13 says violations are handled by the cyberspace, telecommunications, public security and broadcasting authorities "according to their duties, under the relevant laws, administrative regulations and departmental rules".

Two actions show what that means in practice:

• 28 April 2026. The CAC said the Jianying and Maoxiang apps and the Jimeng AI website, all ByteDance products, had "failed to effectively implement the labelling requirements for AI-generated and synthetic content", in breach of the Cybersecurity Law, the interim measures on generative AI services and the labelling Measures. Local cyberspace offices were directed to summon the operators, order corrections, issue warnings and deal strictly with the people responsible. The notice gives no fine (CAC notice, read 13 October 2026; verbatim wording from IT Home's reprint).
• 12 February 2026. The CAC reported that platforms had dealt with 13,421 accounts and removed more than 543,000 items in a sweep against false information spread without AI labels, and that shops selling tutorials and software for removing labels were taken down (CAC notice, read 13 October 2026).

How does it compare with Article 50?

	China's labelling Measures	EU AI Act Article 50(2)
Since	1 September 2025	2 August 2026; 2 December 2026 for systems on the market before then
Visible label	Required for listed services; on images and video, text at least 5% of the shortest side	Not required of providers; deployers disclose deepfakes under 50(4)
Machine-readable mark	An "AIGC" metadata field with seven keys, set by a mandatory standard	Machine-readable and detectable; technique left to the provider and the Code of Practice
Watermark	Encouraged; "not required" by the standard	The Code of Practice pairs signed metadata with an imperceptible watermark
Platforms	Must verify metadata on upload and label three ways	No equivalent duty in Article 50
Removing marks	Prohibited for everyone, including tool suppliers	The Code asks signatories to preserve marks

The largest difference is the direction of the metadata. China specifies its own field and expects platforms to read it; the EU leaves the format open and the industry has settled on C2PA. A product serving both markets will write both.

What this means if you ship AI generation

If your product is offered to users in mainland China, both labels have been due since 1 September 2025, and the metadata label has to be the AIGC field of GB 45438-2025, which a C2PA manifest does not replace. If you export files, the visible label goes in the file unless the user has asked otherwise under your agreement and you keep the logs. If you serve no one in China, the Measures do not reach you, though content your users post to Chinese platforms will be checked for the metadata label and marked as suspected generated content if it shows other signs.

SignUnder signs C2PA manifests and adds a watermark; it does not write the GB 45438-2025 metadata field today, and this page will say so until it does.

Changes

• 2026-10-13 — Read GB 45438-2025 in full: added the visible-label rules by medium, the Annex E metadata format and how it sits with C2PA; the field layout is no longer marked unverified.
• 2026-10-13 — First published, from the CAC's text of the Measures, its announcement of the standard, and its notices of 12 February and 28 April 2026.

SignUnder is not a law firm and this is not legal advice. The Chinese text is authoritative; translations are ours, and anything not read from a primary source is marked.