China’s AI-content labels span creation, export and distribution

Rules introduced in 2025 extend provenance responsibilities beyond the moment an image or text is generated.

An AI-generated image may move through a creator, an editing application, an export service and a publishing platform before a reader sees it. A disclosure attached only at creation can vanish along the way. China’s 2025 synthetic-content measures address this chain by discussing visible and file-level identification, distribution responsibilities and user declarations. The governance question becomes how information survives a workflow.

The [2023 generative-AI service measures](https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm) took effect on 15 August 2023 and address services supplied to the public within China, with stated scope distinctions. The [2025 labelling measures](https://www.nrta.gov.cn/art/2025/3/14/art_113_70340.html) took effect on 1 September 2025. A [regulator explanation](https://www.cac.gov.cn/2025-03/14/c_1743654685896173.htm) describes how labelling responsibilities extend across creation and dissemination. The English descriptions here are editorial paraphrases of primary Chinese texts, not certified translations or legal advice.

The measures distinguish labels that users can perceive from identification carried in file data. The practical implication is that a product team should examine both the visible experience and what happens during export. This does not mean that any label proves the truth of the content. A disclosed synthetic scene can still mislead about an event, and an authentic photograph can be presented with a false caption.

An original illustrative publishing tool lets a small business generate an image for a clearly fictional seasonal advertisement. Staff can see that it is synthetic while editing. When the image is downloaded, the team checks which identification remains in the file and which disclosure appears to viewers. They also record the provider and version used. This proposed workflow illustrates a provenance task; it does not establish compliance for a particular service.

Now suppose the image passes through an editor that changes its format. Metadata may be lost during ordinary processing. The team needs a documented way to detect that loss and preserve the appropriate disclosure in the final publication. The solution should not encourage removing required labels. It should help authorised users understand why an export differs and how to complete the workflow correctly.

A platform receiving uploaded media has different information. It may inspect supplied metadata, ask the uploader to declare synthetic content or identify material that needs further review. A product specification should distinguish these sources of evidence. A user declaration, an embedded record and a probabilistic detector are not equally conclusive, and their disagreement needs an understandable handling process.

Testing can use clearly prepared sample files. Include an exported image, a resized copy, a short audio clip and a document containing generated text. Observe what each authorised processing stage preserves. Record the exact tool versions and file transformations. Such tests examine the product’s handling of provenance; they do not establish that a classifier can reliably determine the origin of every file on the internet.

Human-facing language matters. A warning should state what the platform knows, what remains uncertain and what the uploader can do. An unexplained refusal can cause users to try repeated uploads without understanding the problem. A clear review option is particularly useful when authentic material is incorrectly suspected of being synthetic. Keep sensitive file details out of public error messages.

The broader service measures contain responsibilities beyond labelling, including treatment of data and public-facing services. A company should therefore avoid treating a successful metadata test as approval to operate. The relevant scope, licences, content rules and data obligations require current specialist review. The official measures are the starting point for that analysis, not a substitute for it.

The potential benefit is a more persistent account of how content was produced and handled. The limitation is that technical provenance can be incomplete and public understanding can still be poor. Teams adopting AI publishing tools should test the entire authorised path from generation to publication and explain labels accurately. The dated Chinese instruments make that path a concrete policy topic, while observed product behaviour remains a separate evidence requirement.

An internal support guide can show users the difference between a failed export and a provenance review. Staff should preserve the draft, identify the stage that needs attention and provide a safe next step. That reduces repeated attempts and keeps the content’s history available for authorised examination.