On 14 September 2024 the Cyberspace Administration of China opened public consultation on draft measures for labelling AI-generated and synthetic content. The draft sets out two categories of requirement, explicit and implicit labels, covers service providers, distribution platforms and users, and connects to the rules on deep synthesis. It is not yet in force, but the technical path it describes is specific enough to affect product design and platform terms.
Explicit and implicit labels solve different problems
An explicit label speaks to people: it tells a user, at the point of viewing, that the content was generated or synthesised by artificial intelligence. An implicit label speaks to machines and processes: metadata and similar carriers record the origin of generation so that the content remains identifiable after transmission, reposting and editing. Neither substitutes for the other. With only an explicit label, a screenshot or re-upload strips the notice; with only an implicit label, an ordinary user cannot perceive anything.
Responsibility is distributed across three parties
Service providers add labels at the generation stage and bear corresponding explanation duties for outputs. Distribution platforms verify and display labels during distribution, reposting and recommendation, and preserve the label information. Users must not deliberately remove, alter or conceal labels when using or further sharing content. The three form a chain; a break at any point defeats the regime.
The practical impact on companies lies in interfaces and agreements. Platforms will require upstream providers to supply label fields and metadata specifications, while service providers must state in API documentation, SDKs and user terms where the label exists and how it is presented. Once written into contracts, these details cost more to change than to get right at the design stage.
Interaction with the deep synthesis rules
The labelling measures refine rather than replace the deep synthesis rules. The latter already require labels on generated content; the draft specifies the form of the label, its placement and metadata requirements. When assembling compliance materials, companies should address both sets of requirements in one technical plan, so that a product does not satisfy the deep synthesis label while failing the implicit-label requirement.
Additional considerations for cross-border products
Products offered in overseas markets are often subject to labelling or transparency requirements in several jurisdictions at once. Rules differ on label placement, language and retention periods. A workable approach is one internal labelling specification mapped by market, rather than a separate system for each market. The mapping should record the source, purpose and permissible disclosure scope of each field.
Three tasks to start now
Inventory existing generation capabilities, distinguishing text, image, audio, video and digital humans. Decide the label placement and metadata fields for each content type, and write them into product requirements. Confirm with major distribution platforms how label fields are received and verified. The consultation stage is the cheapest window for adjustment.
For label design and platform verification arrangements, see our AIGC labelling service; for the interaction between deep synthesis filing and algorithm filing, see algorithm filing; for personal information in metadata, see data compliance.
This article is based on the draft measures on labelling AI-generated and synthetic content published for public consultation by the Cyberspace Administration of China on 14 September 2024, and on public information. It is for reference only and does not constitute legal advice.