Deep synthesis
Deep synthesis rules and generative AI measures: how do both apply at once?
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: It is not either-or. The deep synthesis Provisions govern applying deep synthesis technology to provide internet information services, and the technical definition covers text, images, audio, video, virtual scenes and more; the generative AI rules govern providing generative AI services to the public. Capabilities such as large-model dialogue and AI writing fall inside both sets of rules, so both must be satisfied — and the shared engineering work only has to be done once.
What you may need to do
- Draw an applicability matrix of business lines by regulation and mark the clauses that hit
- Build identity verification, review, labeling and logs into one shared foundation
- Re-check biometric editing and public-opinion-attribute features separately
- Align the two sets of materials so they do not contradict each other
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 2
These Provisions apply to applying deep synthesis technology to provide internet information services (hereinafter deep synthesis services) within the territory of the People’s Republic of China. Where laws and administrative regulations provide otherwise, those provisions shall prevail.
Provisions on the Administration of Deep Synthesis of Internet Information Services Article 23
Deep synthesis technology means technology that uses generative and synthesized algorithms such as deep learning and virtual reality to produce network information such as text, images, audio, video and virtual scenes, including but not limited to:
Provisions on the Administration of Deep Synthesis of Internet Information Services Article 23
(4) technologies that generate or edit biometric features in image or video content, such as face generation, face replacement, character attribute editing, face manipulation and pose manipulation;
One-line conclusion
The two rule sets stack rather than replace each other. The deep synthesis Provisions govern applying deep synthesis technology to provide internet information services; the generative AI rules govern providing generative AI services to the public. The same large-model product usually triggers both, so both must be satisfied — which is why building the shared compliance capabilities as one foundation is the cheapest route.
Basis
Article 2 sets the scope: “These Provisions apply to applying deep synthesis technology to provide internet information services (hereinafter deep synthesis services) within the territory of the People’s Republic of China. Where laws and administrative regulations provide otherwise, those provisions shall prevail.”
Article 23 defines the scope of deep synthesis technology to include: “technologies that generate or edit text content, such as passage generation, text style transfer and question-and-answer dialogue”, “technologies that generate or edit voice content, such as text-to-speech, voice conversion and voice attribute editing”, and “technologies that generate or edit biometric features in image or video content, such as face generation, face replacement, character attribute editing, face manipulation and pose manipulation”, among six categories.
This yields a direct conclusion: what is now called generative AI — dialogue, writing, voice-over and image generation — itself falls inside the enumerated scope of deep synthesis technology. The test has two steps: first, whether the technology is generative or synthesized and is used to provide an internet information service; second, whether the business also amounts to providing generative AI services to the public. If both steps hit, both apply.
Note that Article 24 of these Provisions contains a connecting clause: those engaged in online publishing services, online cultural activities and online audiovisual program services shall also comply with the provisions of the competent press and publication, culture and tourism, and radio and television authorities. In other words, “both apply” happens along several lines.
What you need to do
- Draw an applicability matrix: business lines (dialogue, writing, voice-over, face swap, digital humans, live streaming) on one axis and applicable regulations on the other, marking the clusters of clauses that hit in each cell.
- Build one shared foundation: real identity information verification, input and output content review, prominent and technical labeling, log retention, and appeal and complaint channels — these duties point to the same engineering capabilities on both lines, so doing them once covers both.
- Re-check high-risk items separately: editing of biometric information such as faces and voices; products or features with public opinion attributes or social mobilization capacity going live; algorithm filing. Put each on a list.
- Align the wording: keep external materials, internal systems and product pages consistent, and avoid different teams writing their own versions.
Common pitfalls
- Assuming that once the generative AI compliance work is done, the deep synthesis duties are automatically satisfied. Requirements such as labeling, and notice and separate consent for editing biometric information, have to be checked on their own.
- Narrowing “deep synthesis” to face swap and voice conversion, and missing intelligent dialogue and intelligent writing.
- Preparing only one set of documents, so nobody corresponds to the checklist of the other line.
- Ignoring the link to filing and security assessment: services with public opinion attributes or social mobilization capacity must complete filing procedures under the corresponding provisions.
- Citing the wrong clause. When quoting, follow the original article numbers rather than recollection.
Frequently asked follow-ups
Q: If we only do text-to-text and never touch images or video, do these Provisions still apply? A: Yes. Passage generation, text style transfer, question-and-answer dialogue and similar technologies are explicitly listed in the scope of deep synthesis technology; whether they apply still turns on whether an internet information service is provided within the territory.
Q: Can we just follow one set of rules? A: No. The regulated objects and clause requirements of the two lines do not fully coincide. The overlapping part can reuse the same capabilities, but the duties unique to each still have to be implemented clause by clause.
Frequently asked follow-ups
- Is a product deep synthesis or generative AI?
- Both labels can hold at once. Deep synthesis looks at the technical method and whether an internet information service is provided; generative AI looks at whether generative AI services are provided to the public, and the two overlap on most large-model products.
- Which capabilities make up the shared foundation?
- Real identity information verification, input and output content review, the feature library and log retention, prominent and technical labeling, appeal and complaint channels, and management of security assessment and filing materials.