China's Deep Synthesis Provisions: Labelling Synthetic Content

China required synthetic media to be labelled before most jurisdictions had begun consulting on the question. The rule is more specific than “label AI content”: it imposes two different obligations that are satisfied in two different places, and only one of them is visible to a viewer.

The instrument and what it covers

The 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) were issued in November 2022 by the Cyberspace Administration of China with the Ministry of Industry and Information Technology and the Ministry of Public Security, and took effect on 10 January 2023. The official text is published by the CAC: the Deep Synthesis Provisions.

“Deep synthesis” (深度合成) is defined broadly: technology using deep learning or virtual reality to generate or edit text, images, audio, video, virtual scenes and other network information. The enumerated examples cover text generation, text-to-speech and voice conversion, face generation and face replacement, gesture and posture manipulation, and immersive realistic scene generation. It is a wider category than “deepfake” in ordinary usage, and it predates and overlaps the generative AI Measures rather than being replaced by them.

The Provisions bind two roles: deep synthesis service providers, who offer the service, and deep synthesis service technical supporters, who supply the underlying technical capability to a provider. If you supply a model or a capability to a Chinese product rather than running the product, the second role is the one to read.

Two labelling duties, not one

This is the distinction that determines what you build, and it is collapsed in most summaries.

• The traceability label, Article 16. Providers must add a label to content generated or edited by the service that does not affect the user’s use of it, and must keep records enabling the content to be traced. This is the always-on obligation: it applies to synthetic content generally, and because it must not interfere with use, it is satisfied by an embedded or metadata-borne marker rather than by something painted across the image.
• The conspicuous label, Article 17. Where the service provides one of the enumerated functions that may cause the public to be confused or misled—intelligent dialogue, synthetic voice, face generation or replacement, immersive realistic scenes and similar—the provider must add a conspicuous label in a reasonable position or area of the content itself. This one is meant to be seen.

The engineering consequence is that satisfying Article 17 with a visible watermark does not satisfy Article 16, and satisfying Article 16 with metadata does not satisfy Article 17. A pipeline that produces synthetic video for a Chinese consumer service needs both a durable machine-readable marker and a human-visible one, and it needs the records that make the content traceable back to the account that produced it.

Removing a label is itself prohibited

Article 18 prohibits any organisation or individual from using technical means to delete, tamper with or conceal the labels required by the Provisions. This is a separate prohibition addressed to the world at large rather than to providers, and it is the provision that makes the labelling scheme more than advisory: a label that anybody downstream may strip without consequence protects nobody.

It has a consequence for product design that is easy to miss. A tool that re-encodes, crops or otherwise processes synthetic media and drops the embedded marker in the process is doing, by machine, the thing the Article prohibits. Anyone building an editing or transcoding step into a pipeline that handles Chinese synthetic content has to treat marker preservation as a functional requirement rather than a nice-to-have. The same structural problem appears in every provenance scheme, which is why content credentials and their durability is a live engineering question rather than a solved one.

Consent for a real person’s face or voice

Article 14 requires that where a deep synthesis service provides functions that edit biometric information such as a face or a voice, the provider must prompt the user of the service to inform the individual whose information is being edited and obtain that individual’s separate consent. This sits on top of the Personal Information Protection Law’s own rules on sensitive personal information, which require separate consent in their own right.

The obligation is structured as a duty to prompt, which reflects the reality that the provider is usually not in contact with the person whose face is being used. That does not make it decorative: a service that offers face replacement with no consent flow at all has not prompted anybody, and the record of having prompted is the kind of thing an inspection asks to see.

Article 15 adds a security assessment obligation for new products, applications or functions with public opinion attributes or social mobilisation capability, and Article 19 requires providers in that category to complete algorithm filing under the algorithm registration requirement. The filing machinery is shared across instruments rather than duplicated by each.

How it sits with the other Chinese rules

Three instruments now overlap and the order they arrived in explains the shape of the overlap. The Deep Synthesis Provisions (in force January 2023) govern synthetic media as a category. The Interim Measures for generative AI services (in force August 2023) govern generative services offered to the public and expressly route their labelling obligation back to the Deep Synthesis Provisions. The 2025 labelling measures then supplied the implementation detail both of them had left to be specified, with a mandatory national standard alongside.

The practical consequence is that a Chinese generative product typically has obligations under all three at once, and that the answer to “how exactly must we label” is in the newest instrument while the answer to “must we label at all” is in the oldest. Reading only one of the three produces a confident and incomplete answer, which is the usual failure mode when the Chinese regulatory stack is summarised in a paragraph.