China · CAC labelling Measures
China's AI labelling rules: two labels, and four parties who must apply them
In force since 1 September 2025, and enforced. One label a person can see, one written into the file. The same provider owes both, on different triggers, and only the file label is unconditional; platforms, app stores and users carry their own duties.
01 / Who owes what
The chain: four parties, four duties
The EU regulates the producer. China regulates the distribution chain — four parties, each with its own duty, and the file carries provenance from one to the next.
| Who | Must do | Where |
|---|---|---|
| Generation service provider | Apply the explicit label where the trigger is met; always write the metadata label; preserve the label through download, copy and export; state the labelling method in the user agreement; supply labelling materials at algorithm filing and security assessment. | Arts. 4, 5, 8, 11, 12 |
| Dissemination platform | Verify the metadata on what users publish, classify it, display the matching notice, and write its own provenance fields back into the file. Provide a declaration function. | Art. 6 |
| App distribution platform | At listing or go-live review, require the developer to state whether AI generation services are offered, and if so verify the labelling materials. | Art. 7 |
| Publishing user | Proactively declare generated content and use the platform's labelling function. | Art. 10 |
Article 10 also prohibits anyone — not just regulated providers — from maliciously deleting, tampering with, forging or concealing a label, or supplying tools or services for others to do so. False labelling is caught too.
02 / Find your duty
Quick decision
The Measures do not define their own class of regulated entity. Article 2 borrows the scope of three upstream instruments.
Any one yes, and the labelling duties attach.
Write the metadata label into every file. It carries no condition.
Apply the explicit label a person can perceive.
The exported file must still carry the explicit label.
Verify metadata, display the matching notice, and write your own provenance fields back into the file.
At listing review, ask whether AI generation is offered and verify the labelling materials.
Fix the user's own labelling obligation in the agreement first, then keep logs identifying who received it for at least six months.
Probably not you: if none of the three upstream instruments reaches you, neither do the Measures — the scope is derivative, not free-standing. Whether the regime reaches a provider outside China is not addressed in CAC guidance; the Generative AI Measures apply to services provided to the public in China, wherever the provider sits.
03 / Dates
The rule and the specification arrived together
The technical answer was published at the same time as the duty, with roughly six months of transition.
The management duties and the technical method take effect on the same day.
Issued jointly by the Cyberspace Administration of China with MIIT, the Ministry of Public Security and the NRTA. Fourteen articles.
GB 45438-2025, issued by SAMR and SAC. "GB", not "GB/T": mandatory, not recommended.
No change to the Measures or the standard has been announced. CAC's four-month campaign from 30 April 2026 made labelling an enforcement priority.
04 / The duties
Two labels, not one
The regime splits every obligation along one axis: can a person perceive it? The two labels have different legal triggers.
The trigger is Deep Synthesis Provisions Art. 17(1): services that may cause public confusion or misidentification. Its list ends in a catch-all for any service that generates or significantly alters content, so the real limit is the confusion test, not the list. CAC's own Q&A describes explicit labels for text, audio, images, video and virtual scenes. The sentence that has drawn enforcement: an exported file must still carry the label. A badge that lives only in your web player does not satisfy it.
▸ Read the provision
Text, sound or graphics presented in the content or the interface, which the user can plainly perceive. Placement is prescribed per modality: at the start, end or an appropriate middle position for text and audio; a prominent mark on images; the start frame and the playback surround for video, with end and middle positions permitted but not required.
No equivalent condition: if the Measures reach you, the metadata label applies. Digital watermarking is encouraged, not required. The mandatory layer is also the fragile one — a screenshot, a re-encode or a CMS that normalises metadata destroys it.
▸ Read the provision
Metadata embedded in the content file carrying the generation attribute, the service provider's name or code, and a content number. GB 45438-2025 fixes the encoding: a JSON object under an AIGC key, written into the format's native metadata — EXIF or XMP for images, XMP for audio and video, the document information dictionary for PDF, the docProps entries for Office formats.
Metadata present and marked as generated: it is generated content. No metadata, but the user declared it: it may be. No metadata, no declaration, but traces detected: it is suspected. In all three cases the platform writes dissemination elements back into the file — the step platforms skip, cited in CAC's November 2025 sweep.
▸ Read the provision
In all three cases the platform must then write dissemination elements into the file metadata — the generation attribute, its own platform name or code, and a content number.
Not a general log-retention rule. It binds the generation provider only, only in the unlabelled-supply case, only after the agreement is in place — and the metadata label still applies. It is a duty on the record, not only on the act.
▸ Read the provision
Where a user asks for generated content without an explicit label, the provider may supply it — but only after the user agreement has fixed that user's own labelling obligation and use responsibility — and must then retain logs including information identifying who received it, for not less than six months.
05 / Timeline & enforcement
Enforcement is specific, and it names the failure
GB 45438-2025 is issued. The mandatory national standard for labelling AI-generated content.
The Measures are published. Fourteen articles, issued by CAC with three other departments.
Both instruments commence. Management duties and technical method, the same day.
Enforcement against a batch of mobile apps. Provider side: missing explicit labels, labels not carried through on export, missing metadata. Platform side: no metadata verification, no notice, no dissemination write-back, no user declaration function.
Accounts that posted AI content without labels. 13,421 accounts handled and more than 543,000 items removed; shops selling de-labelling tutorials, software and services delisted.
Three services named. CAC publicly named three services for failing to implement labelling effectively, and directed local offices to apply interviews, rectification orders and warnings.
| Enforcement | What the page states |
|---|---|
| Penalty | Article 13 sets none. The Measures refer enforcement out to the competent departments under existing law. |
| Sanctions so far | Administrative: regulatory interviews, orders to rectify, formal warnings, delisting and account-level sanctions. |
| Cited failures | Missing explicit labels, missing metadata, labels lost on export, and platforms not writing back their provenance fields: questions about what your systems did to a particular file on a particular day. |
06 / Where Planisphere fits
Duty → receipt
This is where Planisphere enters, and not before. Planisphere is an evidence layer. It sits beside your systems, never inside the act.
Writes the AIGC metadata block, applies the explicit label, writes the platform's provenance back.
Hash-chained into a per-tenant series, signed and periodically anchored, so an edit to an earlier record breaks the chain.
A third party can detect an edit without trusting either of us, which helps show what your systems did to a file.
It does not apply labels, verify metadata, classify content into the three tiers or detect synthetic media. It is not an algorithm filing or a determination that a label satisfies GB 45438-2025. A record is evidence a third party can check; it is not a finding that you complied.
QUESTIONS
The questions that come next
Does our C2PA implementation satisfy GB 45438-2025?+
No. The standard specifies a flat JSON object under an AIGC key in the file's native metadata, with named fields for the producer and the propagator. C2PA is a cryptographically signed manifest with assertions and a hard binding to the asset. They can coexist in the same file — and in a multi-jurisdiction pipeline they generally must — but neither satisfies the other.
We only sell an API to businesses. Are we out?+
Unresolved. The Measures contain no B2B or API exemption, and scope is inherited from the upstream instruments — one of which carves out services not provided to the domestic public. An API confined to a closed enterprise pipeline is arguably outside; one whose output reaches public-facing Chinese content is not. This is a question for counsel, not a page.
Our metadata keeps getting stripped by re-encoding. Is that a violation?+
Article 10's prohibition is on malicious removal, so incidental loss in a transcode is not on its face a breach of it. But it can still put you in breach of your own Article 5 or Article 6 duty, and preserving the block through edit, transcode, export and delivery is an engineering programme rather than a configuration change.
Is watermarking mandatory?+
No. Article 5 requires an implicit label in file metadata and separately encourages digital watermarking. The metadata label is mandatory; the watermark is encouraged.
How does this compare with the EU?+
China prescribes the method and regulates the whole distribution chain; the EU states an outcome and binds providers and deployers only. Nothing in Article 50 obliges a platform or an app store to inspect files and re-mark them. <a href="/regulations/eu-ai-act-article-50">The EU page covers that regime.</a>
SOURCES
Go back to the source
All fourteen articles: the derivative scope in Art. 2, the two labels in Arts. 4 and 5, the platform ladder in Art. 6, the six-month log in Art. 9, the prohibitions in Art. 10, and the referral of enforcement in Art. 13.
The legal basis, the relationship to the national standard, the transition period, and the sectoral carve-outs.
Issued 28 February 2025, in force 1 September 2025. The labelling method, and the metadata schema at Annex E.
The itemised provider-side and platform-side failures, including export stripping and the missing dissemination write-back.
A four-month national campaign from 30 April 2026. Labelling that falls short is a phase-one priority.
The standard English rendering of the article text quoted in substance above.
Per-format embedding for text, image, audio and video files, plus a guide on protecting the integrity of the label itself.
RELATED
Related: EU AI Act Article 50 · AI content labeling laws, compared · All five AI transparency regimes · California SB 942 · India IT Rules on synthetic media · Korea AI Framework Act