Deep synthesis
What kind of algorithm counts as deep synthesis, and does your product count?
Based on: the Provisions on the Administration of the Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定)
Conclusion: Three things to check: whether it uses generative algorithms such as deep learning or virtual reality, whether the output is online information such as text, images, audio, video or virtual scenes, and whether it is used within China to provide internet information services. Article 23 of the Provisions lists six categories item by item — text, speech, non-speech audio, biometric features such as the human face, non-biometric images, and digital persons and virtual scenes — and intelligent writing, AI dubbing, AI face swapping and digital humans all fall inside.
What you may need to do
- List every generative, editing and synthesis function in the product
- Tick them off one by one against the six categories of technical method in Article 23
- Confirm whether you are a provider, a technical supporter or a user
- Register the functions that hit the criteria in a compliance register and name an owner for each
Clause basis
Provisions on the Administration of the Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定) Article 23
Deep synthesis technology means technology that uses generative algorithms such as deep learning and virtual reality to produce online information such as text, images, audio, video and virtual scenes, including but not limited to:
Provisions on the Administration of the Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定) Article 23
(6) technology for generating or editing digital persons or virtual scenes, such as three-dimensional reconstruction and digital simulation.
Provisions on the Administration of the Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定) Article 2
These Provisions apply to the use of deep synthesis technology to provide internet information services within the territory of the People's Republic of China (referred to as deep synthesis services). Where laws and administrative regulations provide otherwise, those provisions shall be followed.
One-line answer
To judge whether a function counts as deep synthesis, look at three elements: does it use generative algorithms such as deep learning and virtual reality; is the output online information such as text, images, audio, video and virtual scenes; and is it used within the territory of the People's Republic of China to provide internet information services. Where all three are met, these Provisions apply.
The basis
Article 23 defines it as follows: “Deep synthesis technology means technology that uses generative algorithms such as deep learning and virtual reality to produce online information such as text, images, audio, video and virtual scenes, including but not limited to”, and then lists six categories item by item:
- (1) technology for generating or editing text content, such as article generation, text style transfer and question-and-answer dialogue;
- (2) technology for generating or editing speech content, such as text-to-speech, voice conversion and speech attribute editing;
- (3) technology for generating or editing non-speech content, such as music generation and scene sound editing;
- (4) technology for generating or editing biometric features in image or video content, such as face generation, face replacement, person attribute editing, face manipulation and posture manipulation;
- (5) technology for generating or editing non-biometric features in image or video content, such as image generation, image enhancement and image restoration;
- (6) technology for generating or editing digital persons or virtual scenes, such as three-dimensional reconstruction and digital simulation.
The same article also defines three roles: the deep synthesis service provider (an organisation or individual that provides deep synthesis services), the deep synthesis service technical supporter (an organisation or individual that provides technical support for deep synthesis services) and the deep synthesis service user (an organisation or individual that uses deep synthesis services to produce, copy, publish or disseminate information). The role determines the duties, so identify where you sit first and then read the corresponding articles.
What you need to do
- List every generate, edit and synthesize entry point in the product as a table, check them one by one against the six categories of technical method above, and mark the ones that hit.
- Be clear about your position in the chain. Training your own model and serving the public in your own name usually makes you a provider; delivering a model, template or SDK to others for their use usually makes you a technical supporter; if you do both, both sets of duties need review.
- Typical landing points: intelligent customer service and intelligent writing, AI dubbing and voice cloning, AI face swapping and beautification, digital human streaming, text-to-image and image restoration.
- Write the conclusion into the compliance register with the articles it hits and the team responsible, and reuse it as versions iterate.
Common pitfalls
- Assuming that “we only call a third-party large model API” means you are not a provider. The test is whether you provide internet information services to the public in your own name, not who trained the model.
- Staring only at the final product. Editing actions such as text style transfer, speech attribute editing and image enhancement are on the list too.
- Ignoring the territorial and scenario conditions. Only the use of deep synthesis technology within the territory of the People's Republic of China to provide internet information services falls inside these Provisions; purely internal use does not amount to providing an information service to the public, but once it is published externally it should be reviewed as an external service. For borderline cases, follow the practice of the local cyberspace administration office.
- Narrowing “deep synthesis” to face swapping and voice cloning, and missing text capabilities such as intelligent dialogue and intelligent writing.
Common follow-up questions
Q: Does generating one illustration count as deep synthesis too? A: Image generation is expressly listed in Article 23, item (5) and falls within the scope of deep synthesis technology; whether these Provisions apply depends on whether it is used to provide internet information services to the public.
Q: Do internal enterprise tools need to be managed? A: The premise of these Provisions is the provision of internet information services. Tools used internally, without providing information services to the outside, are outside their reach; the parts that involve external publishing or external services still need a separate assessment.
Common follow-up questions
- Does any use of AI count as a deep synthesis service?
- Not necessarily. Both conditions must be met — “producing online information with generative algorithms” and “using the technology within China to provide internet information services” — and the case must fall within the technical methods listed in Article 23.
- We only provide technical support and do not face users directly. Does that count?
- Yes. Article 23 lists the “deep synthesis service technical supporter” as a role of its own, and Articles 15, 14, 19 and others impose duties on it directly, for example security assessment and filing.