Deep synthesis
Using AI face swap, AI voiceover or digital humans in marketing: which compliance steps apply?
Based on: Provisions on the Administration of Deep Synthesis of Internet Information Services
In short: in a marketing setting there are five things to do: obtain notice and separate consent from the individual being edited before using the material; label prominently in a reasonable position and area of the page; provide users with a prominent labeling function and prompt them to fulfil their notification duty; retain review records and network logs; and, for any function involving public-opinion attributes or social-mobilisation capacity, carry out a security assessment in accordance with the relevant national provisions before launch. Keeping the authorisation evidence for a real person’s portrait and voice is especially critical.
What you may need to do
- Obtain notice and separate consent before using a real person’s portrait or voice, and keep the evidence
- Label advertising and live-stream pages prominently in a reasonable position
- Prompt users to fulfil their notification duty and provide a labeling function on the tool side
- Retain material authorisations, review records and logs for inspection
- Carry out a security assessment in accordance with the relevant national provisions before launching any function with public-opinion attributes
Statutory 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》), Article 14
Where a deep synthesis service provider or a technical supporter provides editing functions for biometric information such as human faces and human voices, it shall prompt the user of the deep synthesis service to inform the individual being edited in accordance with law and to obtain that individual’s separate consent.
Provisions on the Administration of Deep Synthesis of Internet Information Services, Article 17
(3) editing services that generate, or significantly alter the personal identity characteristics of, images or videos of persons, such as face generation, face replacement, face manipulation and posture manipulation;
Provisions on the Administration of Deep Synthesis of Internet Information Services, Article 23
(6) technologies for generating or editing digital persons and virtual scenes, such as three-dimensional reconstruction and digital simulation.
One-sentence conclusion
Using AI face swap, AI voiceover or digital humans in marketing is in substance the provision of a deep synthesis service to the public, and the compliance steps come down to five: authorise the material first, label the content, prompt on the tool side, keep records of the process, and assess high-risk functions before launch. The notice plus separate consent requirement for a real person’s portrait and voice is the link most easily overlooked, and the one most likely to cause trouble.
Basis
Article 14, second paragraph: “Where a deep synthesis service provider or a technical supporter provides editing functions for biometric information such as human faces and human voices, it shall prompt the user of the deep synthesis service to inform the individual being edited in accordance with law and to obtain that individual’s separate consent.”
Article 17 brings two high-frequency marketing functions within the scope of prominent labeling: “(2) speech generation or editing services that synthesise human voice or imitate voice, or that significantly alter personal identity characteristics” and “(3) image or video generation or editing services for persons, such as face generation, face replacement, face manipulation and posture manipulation, or that significantly alter personal identity characteristics”, and requires prominent labeling in a reasonable position and area of the generated or edited information content, so as to inform the public of the deep synthesis.
Article 23, item (6) brings digital humans within the technical scope: “technologies for generating or editing digital persons and virtual scenes, such as three-dimensional reconstruction and digital simulation” are deep synthesis technologies.
Where news is being reproduced, Article 6 also applies: deep synthesis services must not be used to produce, copy, publish or disseminate false news information.
What you need to do
- Authorise the material first: before using a real person’s portrait or voice, inform the individual being edited and obtain separate consent, and file the authorisation document, the communication record and the source channel together; material from outside the country likewise needs a verifiable origin.
- Put the label on the page: prominently mark the advertising page, live-stream room or short video as deep synthesis content, in a reasonable position and area — not a single icon tucked into a corner.
- Do the tool-side duty: as provider or technical supporter, prompt users in the product to fulfil their notification and separate consent duties, and provide a prominent labeling function.
- Keep the process documented: material authorisations, content review, logs and handling records form one complete chain, ready for an inspection or a platform check.
- Assess before launch: where a new product, application or function with public-opinion attributes or social-mobilisation capacity is developed and launched, carry out a security assessment in accordance with the relevant national provisions; where a tool involving the editing of biometric information such as human faces and voices is provided, carry out a security assessment yourself or through a professional institution in accordance with law.
- Cooperate with the channels: distribution platforms such as app stores verify the security assessment and filing status of deep synthesis applications, so have the materials complete before submitting for review.
Common pitfalls
- Using an AI voice or face that merely resembles someone as an endorser, on the view that it infringes nothing because it is not the person. Misleading content triggers the labeling duty and the prohibition on disseminating false information at the same time.
- Treating authorisation from the host or model as a reason to relax the labeling duty. Authorisation solves consent; it does not replace prominent labeling.
- Assuming the risk moves to a third party once the material is outsourced — the provider is still the duty-bearing subject.
- Running a digital human live stream for months without putting the label into the template, relying on someone remembering to add it each time.
- Scripts for training, voiceover or livestream selling that use phrases such as tested on a real person or the person appears in person, when the content is in fact deep synthesis.
- Signing only a portrait use licence, with no separate consent for the editing of biometric information.
Follow-up questions
Q: Does a digital human host need to be labeled at every live stream? A: Whenever it is a deep synthesis service that may cause public confusion or misidentification, it must be prominently labeled in a reasonable position and area so as to inform the public of the deep synthesis — not labeled once at the first broadcast.
Q: An employee consents to a digital employee based on their image — is anything else needed? A: Consent is the starting point. Prominent labeling, content review and log retention are also required; and if the content is disseminated to the public, you must assess whether it has public-opinion attributes or social-mobilisation capacity and act in accordance with the relevant national provisions.
Follow-up questions
- Does an AI-generated model image used as a product hero image count as deep synthesis?
- Technologies such as image generation and image restoration fall within the scope of deep synthesis technology. When it is disseminated to the public it should be prominently labeled, and the review requirements for advertising material require the process to be documented.
- The client supplies the material and promises it is authorised — does the provider still have to prompt?
- Yes. Article 14, second paragraph, requires providers and technical supporters to prompt users to inform the individual being edited in accordance with law and obtain separate consent; this prompting duty cannot be waived by a client’s promise.