Measures for Labelling AI-Generated and Synthetic Content

For content within Art. 17 §1 of the Deep Synthesis Provisions, add the explicit label by modality under Art. 4: for text, a textual or common-symbol prompt at the start, end or an appropriate midpoint, or a conspicuous prompt in the interactive interface or around the text; for audio, a spoken or rhythmic prompt at the start, end or an appropriate midpoint, or a conspicuous prompt in the interactive interface; for images, a conspicuous prompt at an appropriate position; for video, a conspicuous prompt on the opening frame and around the player, optionally also at the end and at appropriate midpoints; for virtual scenes, a conspicuous prompt on the opening frame, optionally also at appropriate points during the session; and for other scenarios, a conspicuous prompt suited to the application. Where download, copy or export functions are offered, the file must still contain a conforming explicit label.

Applies when

Providers of intelligent dialogue and writing, voice synthesis and cloning, face generation and swapping, pose manipulation, and immersive realistic scenes. The requirement that exported files still carry the explicit label rules out implementations that only watermark the front-end interface.

Applies from: 2025-09-01

Evidence to retain

• · Implementation and placement note for the explicit label in each modality: text, audio, image, video and virtual scenes
• · Sample labelled files per modality, including samples after download, copy and export
• · Mapping table comparing product functions against the five categories in Art. 17 §1 of the Deep Synthesis Provisions
• · Technical verification that the label survives the export path

Estimated first-time effort: 15–40 person-days (estimate, not a commitment)

As required by Art. 16 of the Deep Synthesis Provisions, add an implicit label to the metadata of generated and synthetic files, carrying the generated-content attribute, the provider’s name or code and a content identifier among other production elements. Art. 5 §3 defines file metadata as descriptive information embedded in the file header in a specific encoding format, recording the file’s source, attributes and use. The Measures also encourage adding implicit labels in the form of digital watermarks. The specific fields and encoding must conform to GB 45438-2025.

Applies when

Every provider of generative or synthetic services, including API-only model platforms — a downstream customer cannot supply the provider’s own name or code, so the upstream must write it at output time.

Applies from: 2025-09-01

Evidence to retain

• · Field design document for the implicit label: attribute, provider name or code, and content identifier
• · Implementation note for metadata writing in each file format, with exported metadata from sample files
• · Field and encoding conformity check against GB 45438-2025
• · Digital watermark design and robustness test results, where used
• · API documentation showing metadata carried in outputs

Estimated first-time effort: 15–40 person-days (estimate, not a commitment)

A provider of network content dissemination services must: verify whether the file metadata carries an implicit label, and where the metadata confirms the content is generated or synthetic, add a conspicuous prompt around the published content telling the public it is generated content; where no implicit label is found but the user declares the content as generated, add a conspicuous prompt telling the public the content may be generated; where no implicit label is found and the user has not declared, but an explicit label or other traces of generation are detected, treat it as suspected generated content and add a conspicuous prompt; and provide the necessary labelling function while prompting users to declare whether their post contains generated content. In the first three situations the platform must also write dissemination elements into the file metadata — the generated-content attribute, the platform’s name or code and a content identifier.

Applies when

Products with content distribution, community, feed or sharing features; AI communities and UGC platforms first, though enterprise knowledge-base products that share content externally may also fall in.

Applies from: 2025-09-01

Evidence to retain

• · Metadata verification workflow, plus the copy and display samples for each of the three prompt tiers
• · Description of the detection capability for traces of generation, with detection-rate statistics
• · Implementation note and samples for writing dissemination elements back into the metadata
• · Screenshot of the user declaration entry point and declaration-rate statistics
• · Note on how this connects to Art. 9 §2 of the Algorithmic Recommendation Provisions, which requires labelling before transmission continues

Estimated first-time effort: 15–35 person-days (estimate, not a commitment)

The user service agreement must set out the method and style of content labelling and remind users to read and understand the labelling requirements. Where a user asks for generated content without an explicit label, the provider may supply it after the user agreement has made the user’s own labelling duty and usage responsibility clear, and must retain logs including details of the recipient for no less than six months.

Applies when

Model platforms and creation tools offering watermark-free export or white-label output. This is the only lawful route to supplying output without an explicit label, and it requires both the agreement terms and the six-month logs — neither alone suffices.

Deadline: Recipient and related logs retained for no less than six months.

Applies from: 2025-09-01

Evidence to retain

• · Service-agreement clauses describing the labelling method and style, with evidence of user acknowledgement
• · Documented application and approval process for supplying output without an explicit label
• · Recipient logs (account, entity, timestamp, content identifier) and screenshots of the six-month retention configuration
• · Usage statistics and anomaly monitoring for unlabelled output

Estimated first-time effort: 8–20 person-days (estimate, not a commitment)

When completing algorithm filing, security assessment and similar procedures, the provider must submit the labelling materials required by these Measures, and must strengthen the sharing of labelling information to support the prevention and investigation of related offences. Labelling must also conform to the applicable laws, administrative regulations, departmental rules and mandatory national standards — namely GB 45438-2025 “Cybersecurity technology — Labeling method for content generated by artificial intelligence” (issued 28 Feb 2025, effective 1 Sep 2025, currently in force).

Applies when

Every generative or synthetic service provider subject to algorithm filing or security assessment. A mismatch between the filed materials and the shipped implementation is the easiest gap for an inspection to find.

Applies from: 2025-09-01

Evidence to retain

• · Labelling description filed with the algorithm filing and security assessment materials
• · Self-assessment checklist and gap list against GB 45438-2025
• · Reconciliation between the labelling implementation described in the filing and the live product
• · Description of the mechanism for sharing labelling information and cooperating with enforcement

Estimated first-time effort: 8–20 person-days (estimate, not a commitment)