How to Understand China’s Measures for Labeling AI-Generated and Synthetic Content

If someone uses AI to create an illustration, alter a voice, or produce a video, must the result be identified as AI-generated before it is posted in China? Is a short notice in the corner enough? Can a file without a visible watermark be treated as non-AI content?

These questions cannot be answered with a universal yes or no. China’s framework assigns different duties to generation-service providers, content-distribution platforms, app-distribution platforms, and users who publish content. It also combines labels people can perceive with labels stored inside files.

The Measures for Labeling AI-Generated and Synthetic Content were published in March 2025. The supporting mandatory national standard, GB 45438—2025, Cybersecurity Technology—Labeling Method for Content Generated by Artificial Intelligence, was published on February 28, 2025. Both took effect on September 1, 2025.

The Measures primarily establish management duties: who adds labels during generation, who verifies and displays notices during distribution, and what a user must do when publishing. GB 45438—2025 translates those duties into technical methods for text, images, audio, video, and virtual scenes.

The official Q&A describes the relationship directly: the Measures define the responsibilities of parties involved in production and distribution, while the mandatory national standard sets out the implementation and operating methods for the mandatory elements. The two documents should therefore be read together.

Under the Measures, AI-generated and synthetic content includes text, images, audio, video, virtual scenes, and other information generated or synthesized using AI technology. The definition is not limited to images created entirely from a prompt; it also covers synthetic content.

This does not mean every computational-photography adjustment, automatic correction, or simple filter should automatically be treated as content requiring a prominent AI label. Article 2 defines the covered service providers, and the explicit-label requirement in Article 4 is tied to Article 17(1) of the Provisions on the Administration of Deep Synthesis in Internet-based Information Services. A particular case depends on the service, the degree of generation or synthesis, the method of publication, and any other applicable rules. This article is general information, not legal advice for a specific business.

An explicit label is a notice that a person can directly perceive. Text may carry a notice or a common symbol at an appropriate place, or a prominent notice may appear in the interface or around the text. Images use a prominent notice in an appropriate position. Audio may use spoken or rhythmic prompts. Video uses a prominent notice in the opening frame and around the playback area.

An implicit label is stored in the file data and normally is not directly visible. A generation-service provider must add metadata describing the generated or synthetic nature of the content, the provider’s name or code, a content identifier, and other production elements.

The Measures encourage digital watermarks and similar forms of implicit labeling. That encouragement does not make the separate requirements for explicit labels and file metadata optional. The visible notice communicates with people; file metadata helps platforms and tools read source signals. They serve different purposes.

For services within the scope of the rules, source-level responsibilities include:

1. adding the required explicit label for the relevant content type;
2. preserving the required explicit label when content is downloaded, copied, or exported;
3. adding the required implicit label to file metadata;
4. explaining the labeling method, style, and user responsibilities in the user agreement; and
5. submitting labeling materials when completing algorithm filing or security-assessment procedures.

A provider is not absolutely prohibited from supplying content without a visible label in every circumstance. Article 9 creates a conditional route: after the user agreement defines the user’s labeling duties and responsibilities, the provider may supply content without an explicit label and must retain information about the recipient and other relevant logs for at least six months.

If the user later distributes that content to the public, the user must still make a declaration and use the distribution platform’s labeling function. No visible label at export does not mean no declaration at publication.

Online content-distribution providers have verification and notice duties. The Measures describe three common cases:

• If file metadata clearly identifies generated or synthetic content, the platform should place a prominent notice around the published content.
• If metadata does not contain such a label but the user declares the content, the platform should notify the public that it may be generated or synthetic.
• If neither metadata nor a user declaration is present, but the platform detects an explicit label or other generation or synthesis traces, it may identify the content as suspected generated or synthetic content and display a corresponding notice.

Platforms must also provide the necessary declaration and labeling functions. For content covered by these cases, they must add distribution elements to metadata, including the generated-content attribute, platform name or code, and a content identifier.

This is not a simple “watermark present or absent” test. Metadata, user declarations, visible labels, and platform detection may all contribute. In August 2025, China’s National Information Security Standardization Technical Committee published six practice guides covering file-metadata labeling and detection for text, images, audio, and video.

Article 10 says that a user publishing generated or synthetic content through an online information-content distribution service must proactively declare it and use the provider’s labeling function.

It also prohibits any organization or individual from maliciously deleting, altering, forging, or concealing the required labels, or supplying tools or services for those acts. Metadata lost through ordinary transcoding, screenshots, or platform compression is not the same as deliberately removing a label to make content appear to be an authentic record. Accidental metadata loss, however, does not automatically remove the user’s declaration duty.

A practical approach for an individual creator or small business is to preserve the original export and creation records, use the platform’s AI-content declaration at publication, and add a clear notice in the content or description when appropriate.

An explicit AIGC declaration in a file is important source information. Reading a metadata field, however, is not the same as validating a digital signature and trust chain. Metadata can be lost or improperly modified, so a report should distinguish among a file declaration, a verifiable provenance credential, and a model inference.

ShanHaiYin checks available provenance credentials and standardized file labels before using model analysis when direct source evidence is missing. That order prevents a probability score from being presented as proof of origin and prevents ordinary metadata from being described as tamper-proof. See our guide to the evidence order used in AI image detection.

Conversely, the absence of a readable national-standard label does not prove that content is not AI-generated. Screenshots, crops, re-exports, messaging-app compression, and platform transcoding may alter the file or remove metadata. Older content and content created with foreign tools may not carry the same format.

No. A detector can identify existing labels and estimate whether content may have been generated, synthesized, or edited. It usually cannot determine a publisher’s identity, the production process, intent, or legal responsibility by itself.

Even when a model finds strong AI-related features, a reviewer should examine the original file, generation records, publication page, platform notices, and distribution time. For disputed material, a detection report is better treated as a technical lead for further verification than as a final finding. Our seven-step AI image verification guide describes the wider process.

The rules have also been enforced. In November 2025, Chinese internet regulators reported action against mobile applications that had failed to implement explicit labels, metadata verification, distribution notices, or user-declaration functions. The published enforcement examples focus most directly on platforms and applications providing generation, distribution, and app-distribution services. That does not eliminate an ordinary user’s duty to declare content.

1. Record which tools were used and which parts were generated or synthesized by AI.
2. Keep the original export rather than retaining only a screenshot.
3. Check whether the content, description, or playback interface already contains a clear explicit notice.
4. Use the platform’s “AI-generated” declaration function when publishing.
5. Do not deliberately remove or cover an existing label.
6. For important commercial content, preserve versions, publication times, and account records.
7. If a platform marks content as suspected AI, review the metadata, original file, and creation trail before deciding whether to appeal.

Before publication, ShanHaiYin AI content detection and provenance verification can help identify source signals that remain in a file. The result can support issue spotting and record organization, but it cannot replace platform review, professional legal advice, or a final responsibility determination. See the AI content detection FAQ for input and report questions.

The central purpose of GB 45438—2025 and the Measures is not to place an identical watermark on every piece of AI content. It is to build a chain of labeling from source generation and file export to platform distribution and user declaration.

For an ordinary user, the most important steps are to declare generated or synthetic content when publishing it publicly, use the platform’s labeling function, and preserve the original file. Generation services and content platforms must build capabilities for explicit labels, metadata, verification notices, and logs.

For a verifier, the boundary is equally important: a label is significant evidence, no label does not mean no AI was used, and a model score cannot establish a violation by itself.