Deep synthesis
Does AI-generated content need a label? Can the label be removed?
Basis: Provisions on the Administration of Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定)
Conclusion: Yes. For the five categories of deep synthesis service that may cause public confusion or misidentification, a prominent label must be placed in a reasonable position and area of the generated or edited content. Services outside those categories must still offer a prominent labeling function and tell users they can use it. In addition, a provider must add a technical marking to generated or edited content that does not affect users’ use. No organization or individual may use technical means to delete, tamper with or conceal these labels.
What you may need to do
- Tick your product features against the five categories in Article 17, one by one
- Place a prominent label in a reasonable position and area, and keep a screenshot of the styling
- Offer a labeling function for services outside those categories
- Forbid the silent loss or cropping of labels during transcoding and compression
Provisions relied on
Provisions on the Administration of Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定) Article 17
Where a deep synthesis service provider provides the following deep synthesis services, which may cause public confusion or misidentification, it shall place a prominent label in a reasonable position and area of the generated or edited information content to inform the public that deep synthesis is involved:
Provisions on the Administration of Deep Synthesis of Internet Information Services Article 16
For information content generated or edited using its service, a deep synthesis service provider shall take technical measures to add a label that does not affect users’ use of the service, and shall preserve log information in accordance with laws, administrative regulations and relevant state provisions.
Provisions on the Administration of Deep Synthesis of Internet Information Services Article 18
No organization or individual may use technical means to delete, tamper with or conceal the deep synthesis labels provided for in Articles 16 and 17 of these Provisions.
One-line conclusion
Deep synthesis labeling has two layers: a technical marking that machines read and that does not affect users’ use, and a prominent label that the public reads — wherever a service may cause public confusion or misidentification, the content must say in a reasonable position and area that it is deep synthesis content. Once the label is in place, no organization or individual may use technical means to delete, tamper with or conceal it.
Basis
Article 17 first lists the five categories of service requiring a prominent label, then sets the catch-all and promotion duty:
- (1) services that simulate a natural person generating or editing text, such as intelligent dialogue and intelligent writing;
- (2) voice generation or editing services such as synthesized human voice and voice imitation, or editing that significantly changes an individual’s identity characteristics;
- (3) services that generate or edit images and video of people, such as face generation, face replacement, face manipulation and pose manipulation, or editing that significantly changes an individual’s identity characteristics;
- (4) generation or editing services such as immersive realistic scenes;
- (5) other services with the function of generating or significantly changing information content.
Article 17, paragraph 2: “Where a deep synthesis service provider provides deep synthesis services other than those in the preceding paragraph, it shall offer a prominent labeling function and prompt deep synthesis service users that they may apply a prominent label.”
Article 16 is the technical-layer requirement: “For information content generated or edited using its service, a deep synthesis service provider shall take technical measures to add a label that does not affect users’ use of the service, and shall preserve log information in accordance with laws, administrative regulations and relevant state provisions.”
Article 18 is the prohibition: “No organization or individual may use technical means to delete, tamper with or conceal the deep synthesis labels provided for in Articles 16 and 17 of these Provisions.”
What you need to do
- Classify first: tick your product features against the five items of Article 17. Face replacement, synthesized human voice, intelligent dialogue and intelligent writing are typical hits.
- Then design: put the prominent label in a reasonable position and area that a user sees during normal browsing, with wording that clearly states this is deep synthesis content rather than a vague “AI-assisted”.
- Keep an opening: for services not listed among the five categories but capable of generating or significantly changing information content, offer a prominent labeling function and tell users they can label content themselves.
- Control the downstream: state in the product specification that one-click label removal is prohibited, and ensure transcoding, compression and screenshot tool chains cannot silently lose the label.
- Leave a trail: archive the label styling, launch date and coverage as materials for regulatory inspections.
Common pitfalls
- Writing “this platform may use AI technology” in the user agreement while the interface carries no label at all.
- Hiding the label behind a click-to-expand or hover interaction, invisible during normal browsing.
- Second-round editing, mash-ups or sticker overlays covering the label — equally inside the Article 18 prohibition.
- Assuming avatar generators and virtual character tools are not image-of-people services. Face generation, face replacement and pose manipulation are all listed.
- Reading “prominent” as “big enough” and overlooking that it must also sit in a reasonable area and state that deep synthesis is involved.
Frequently asked follow-ups
Q: How prominent does the label have to be? A: The standard the Provisions set is “a prominent label in a reasonable position and area of the generated or edited information content, to inform the public that deep synthesis is involved.” A workable self-test is: an ordinary user sees it without looking for it, and understands that the content is synthesized. The local cyberspace administration office sets the operative reading on specific styling.
Q: If a user crops the label off, is the provider responsible? A: Article 18 binds “any organization or individual”, which covers both users and providers. A provider should constrain cropping through technology and rules, and keep its handling measures on record.
Frequently asked follow-ups
- Does text AI output also need a prominent label?
- Intelligent dialogue, intelligent writing and similar services that simulate a natural person generating or editing text are listed in Article 17(1), so they fall within the scope requiring a prominent label.
- Are technical marks and prominent labels the same thing?
- No. Article 16 requires a technical marking that does not affect users’ use; Article 17 requires a prominent label that informs the public that deep synthesis is involved. Both are required.