China — Provisions on the Administration of Deep Synthesis Internet Information Services

The Provisions on the Administration of Deep Synthesis Internet Information Services (互联网信息服务深度合成管理规定) are a Chinese regulation governing the use of "deep synthesis" technology — generative and synthetic media — in internet information services. They were issued by the Cyberspace Administration of China (CAC) together with the Ministry of Industry and Information Technology (MIIT) and the Ministry of Public Security (MPS). The provisions combine content-control duties (prohibited uses, content review, real-name verification) with synthetic-media transparency duties (watermarking and conspicuous labeling). They sit between China — Internet Information Service Algorithmic Recommendation Management Provisions and China — Interim Measures for the Management of Generative AI Services in the sequence of Chinese AI rules, and reuse the filing and security-assessment machinery established by the algorithmic recommendation provisions.

Status and timeline

• Promulgated: 2022-11-25
• Effective: 2023-01-10

The provisions were issued jointly by the CAC, MIIT, and MPS, and are grounded in the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law (PIPL). Combined with the China — Internet Information Service Algorithmic Recommendation Management Provisions (2022) and the China — Interim Measures for the Management of Generative AI Services (2023), they form a sequence of Chinese AI rules running from general algorithmic governance to synthetic-media specifics to generative AI specifics. Each layer reuses the filing and security-assessment mechanisms of the one before it; the labeling and real-name verification obligations established here are carried forward by the generative AI interim measures, which cross-reference the labeling regime in their Article 12.

Scope and definitions

The provisions govern the application of "deep synthesis technology" to provide internet information services. Deep synthesis is defined broadly as technologies using deep learning and virtual reality with generative-sequencing algorithms to create text, images, audio, video, virtual scenes, or other information. Article 23 enumerates six categories:

1. Text generation/editing (chapter generation, style conversion, Q&A dialogue)
2. Voice content (TTS, voice conversion, voice attribute editing)
3. Non-voice audio (music generation, scene sound editing)
4. Image/video biometric features (face generation, face swap, personal-attribute editing, gesture manipulation)
5. Image/video non-biometric (image generation, enhancement, restoration)
6. Digital characters / virtual scenes (3D reconstruction, digital simulations, immersive scenes)

Key provisions

General (Arts. 1–5)

Based on the Cybersecurity Law, Data Security Law, and PIPL. The CAC has overall coordination; MIIT and MPS have domain-specific responsibilities. Industry self-discipline is encouraged.

Ordinary provisions (Arts. 6–13)

• Art. 6: Prohibited uses — subversion, national security, economic order, and "fake news information." Reprinting news from deep-synthesis outputs must follow lawful source-unit rules.
• Art. 7: Places primary information-security responsibility on providers, requiring systems for user registration, algorithm review, ethics review, publication review, data security, personal information protection, fraud prevention, and emergency response.
• Art. 8: Disclosure of management rules and platform covenants.
• Art. 9: Requires real-name verification via mobile number, ID number, uniform social credit code, or the state public identity verification service. No information may be published for unverified users.
• Art. 10: Content review (technical and manual); a pool of characteristics to identify illegal or negative content; mandatory network log storage; and reporting of illegal content to authorities.
• Art. 11: Requires a rumor-dispelling mechanism.
• Art. 12: Complaint and reporting portals.
• Art. 13: App-store gatekeeping — internet app stores must check security assessments and filings, and take action against non-compliant deep synthesis apps.

Data and technical management (Arts. 14–18)

• Art. 14: Training-data management. Consent is required from individuals whose biometric information (faces, voices) is being edited, to be obtained by the user and prompted by the service.
• Art. 15: Security assessments are required (self or third-party) for tools that generate or edit biometric information (faces, voices), or generate or edit non-biometric information that might involve national security, image, interests, or the public interest.
• Art. 16: Providers must attach symbols (watermarks or log labels) to generated or edited content that do not impact usage, and must store logs.
• Art. 17: Conspicuous public-facing labels are required for deep-synthesis services that might confuse or mislead the public:
  1. Smart dialogue / smart writing (text-simulating natural persons)
  2. Voice synthesis / imitation / significant voice identity changes
  3. Face generation / face swap / face or gesture manipulation
  4. Realistic immersive scenes
  5. Other services that generate or significantly alter information content

For services not in the enumerated list, providers must offer labeling functionality and alert users.

• Art. 18: Deleting, altering, or concealing labels is prohibited, by any party.

Oversight and legal responsibility (Arts. 19–22)

• Art. 19: Providers with public-opinion or social-mobilization capacity must file via the algorithmic recommendation filing regime. Technical supporters must file similarly. The filing number must be displayed publicly.
• Art. 20: New products or functions with public-opinion or social-mobilization capacity require a security assessment before launch.
• Art. 21: Requires cooperation with oversight inspection; authorities may require suspension of updates or registration.
• Art. 22: Graduated penalties, enhanced for serious consequences, with public-security and criminal referrals available.

Obligations by actor

The provisions distinguish two main obligated parties. Service providers bear the primary information-security responsibility under Article 7, including the registration, review, labeling, watermarking, log-storage, filing, and security-assessment duties above. The provisions also reach upstream: "technical supporters" — those who supply the underlying technology rather than operate the public-facing service — must file under Article 19 in the same way as providers with public-opinion or social-mobilization capacity. App stores bear a distinct gatekeeping duty under Article 13 to verify security assessments and filings and to act against non-compliant apps. Users are brought in through Article 14, under which the consent of an individual whose biometric information is edited is obtained by the user at the service's prompting.

Enforcement and penalties

Article 21 requires providers to cooperate with oversight inspections and allows authorities to require suspension of information updates or of user registration. Article 22 provides for graduated penalties, enhanced where serious consequences result, and allows referral to public-security authorities and for criminal liability. The app-store gatekeeping duty (Art. 13) and the prohibition on tampering with labels (Art. 18) extend enforcement beyond the primary service provider.

Comparison with other approaches

China's deep-synthesis labeling regime predates the EU AI Act's transparency obligations and operates in two tiers: a technical watermark or log label under Article 16, and a conspicuous public-facing label under Article 17. The labeling duties are embedded within a regime that also requires real-name identity verification (Art. 9) and adherence to prohibited-use and content-review rules (Arts. 6, 10). By extending obligations upstream to technical supporters rather than only to deployers, the provisions establish a liability model comparable in principle to the EU AI Act's general-purpose AI provider obligations, though tied to content and filing rather than to safety and transparency. Article 9's real-name verification requirement structurally prevents anonymous use of deep-synthesis tools, a baseline incompatible with US First Amendment jurisprudence (see AI and the First Amendment). Article 18's prohibition on deleting, altering, or concealing labels binds any party rather than only providers.

Relationships

• depends-on: China — Internet Information Service Algorithmic Recommendation Management Provisions (filing regime reused under Art. 19)
• related: China — Interim Measures for the Management of Generative AI Services (reuses labeling regime via cross-reference; Art. 12)
• related: EU AI Act (Regulation 2024/1689) (transparency labeling comparison)
• related: AI and Content Moderation
• related: AI and Authoritarianism
• related: AI and the First Amendment
• related: AI Sovereignty