AI Content Rules: Deep Synthesis and Labeling
Two regulations, one mandate: every piece of AI-generated content in China must carry both a visible label and embedded metadata. Here is how to comply.
The Deep Synthesis Provisions and AI Content Labeling Measures primarily target services that generate or provide AI-generated content to the public. Internal enterprise tools require careful scoping:
- Public-facing services (chatbots, content generators, customer-facing AI tools) are clearly within scope and require full labeling, watermarking, and metadata tagging.
- Internal-only tools (manufacturing quality control, internal knowledge management, predictive maintenance) that are not accessible to external users may have lighter obligations, but the boundary is not always clear.
- Gray area: Tools used by factory workers, field staff, or external contractors may be considered “public-facing” if the output reaches external parties or the public. If an internal AI tool produces reports shared with customers, the output labeling requirements likely apply to those reports.
- When in doubt, apply labeling requirements to any AI-generated content that could be viewed by someone outside your organization.
Deep Synthesis Provisions vs. AI Content Labeling Measures
The Deep Synthesis Provisions set the foundation. The Labeling Measures add specific technical requirements. Both apply simultaneously.
Deep Synthesis Provisions
- 25 articles across 5 chapters covering all AI-generated media and deepfakes
- Label all AI-generated content with visible disclaimers
- Real-name identity verification for all users of deep synthesis services
- Content management systems for reviewing generated output
- Safety assessments required before deploying biometric editing features
- Biometric editing (faces, voices) requires notifying affected individuals and obtaining consent
AI Content Labeling Measures
- Explicit labels: visible “AI-generated” disclaimers on all output
- Implicit labels: metadata tags embedded in file headers
- Both explicit and implicit labels required at the same time
- Text label height must be at least 5% of the image’s shortest side
- 6-month log retention for all labeled content
- GB 45438-2025: mandatory national standard specifying the labeling method
Label Placement and Sizing Rules
GB 45438-2025 specifies exactly how explicit and implicit labels must appear. These are the minimum requirements.
Explicit Label (Visible)
Label text height must be at least 5% of the image’s shortest side. Placed at a visible location that does not obscure the main content.
Implicit Label (Metadata)
Metadata tags embedded in file headers per GB 45438-2025. Must survive file format conversion and basic editing operations.
Required Metadata Fields
Per GB 45438-2025 implicit labeling standard.
Labeling Requirements per Output Format
Each content type has distinct labeling obligations. Scroll horizontally on mobile.
| Requirement | Text | Image | Audio | Video |
|---|---|---|---|---|
| Explicit label required | ✓ | ✓ | ✓ | ✓ |
| Implicit metadata required | ✓ | ✓ | ✓ | ✓ |
| 5% shortest-side rule | – | ✓ | – | ✓ |
| Visible text disclaimer | ✓ | ✓ | – | ✓ |
| Audio announcement | – | – | ✓ | ✓ |
| Biometric consent flow | – | ✓ | ✓ | ✓ |
| 6-month log retention | ✓ | ✓ | ✓ | ✓ |
Consent Flow for Face and Voice Editing
The Deep Synthesis Provisions require a three-step consent process before any AI system edits a person’s face or voice. This applies to face-swap, voice cloning, and facial reenactment features.
User Initiates Biometric Edit
User uploads or selects content containing another person’s face or voice for AI editing (face-swap, voice clone, facial reenactment).
Notify the Affected Individual
The platform must prompt the user to inform the person whose biometric data is being edited. The notification must describe the intended use and editing scope.
Obtain Individual Consent
The affected individual must provide separate consent for the biometric editing. This is distinct from the platform’s general terms of service.
Enforcement Is Not Hypothetical
The CAC’s Qinglang campaigns target unlabeled and falsely labeled AI content. These numbers are from the February 2026 enforcement wave.
Content Rules Compliance Checklist
Check each item as you complete it. Your progress is tracked automatically.
Need Help with Content Labeling Compliance?
TJS advisors help teams implement AI content labeling, watermarking, and metadata tagging to meet the September 2025 requirements.
Talk to a TJS Advisor →