Deep synthesis and the labeling regime
Lesson 3 of 5 in China: The World’s Most Developed Binding AI Rulebook.
Ten months after the algorithm provisions came the Provisions on the Administration of Deep Synthesis of Internet Information Services, effective 10 January 2023 — the world’s first dedicated deepfake regulation, finalised weeks after ChatGPT launched but drafted well before it. “Deep synthesis” covers technology that uses deep learning to generate or edit text, audio, images, video, and virtual scenes: face swaps, voice cloning, text generation, digital humans.
Three duties define the regime. First, consent: anyone whose face or voice is edited or cloned must give consent — the provision aimed squarely at non-consensual deepfakes. Second, labeling: content that could confuse or mislead the public must carry a conspicuous label, and all synthetic output must carry technical marks. Third, a split of responsibility between deep synthesis service providers (who face users) and technical supporters (who supply the underlying models and tools) — an early answer to the value-chain question the EU later wrestled with for GPAI, asked and answered in Beijing a year earlier.
Two years of enforcement revealed the gap: labels existed, but each platform improvised its own, and stripped or missing marks were rampant. The answer was the Measures for Labeling of AI-Generated Synthetic Content — issued March 2025 by the CAC with MIIT, MPS, and NRTA, effective 1 September 2025, alongside a mandatory national standard (GB 45438-2025) specifying exactly how labels must be implemented. Together they form the world’s most prescriptive synthetic-content labeling regime.
The measures split every label into two kinds. Explicit labels are what a human perceives: text notices, audio announcements, on-screen watermarks. Implicit labels are what a machine reads: metadata embedded in the file carrying the content’s synthetic status, the provider’s name or code, and a content ID. Both are generally required, and the duty chain runs the full distribution path: generators must label, platforms must check metadata and add notices when they detect (or suspect) synthetic content, app stores must verify that AI apps have labeling functionality, and users themselves must not maliciously strip labels.
| Content type | Examples | Explicit label (human-perceptible) | Implicit label (machine-readable) |
|---|---|---|---|
Text | Chatbot replies, generated articles | Text or symbol notice at the start, end, or an appropriate point in the text (or prominent notice in the interface) | Metadata in the file: synthetic status, provider name/code, content ID |
Audio | Cloned voices, generated music | Voice announcement or audio cue (e.g., a rhythm marker) at start, end, or an appropriate point | Embedded metadata; watermarking encouraged |
Image | Generated or edited images | Prominent visual mark on the image itself | Embedded metadata identifying synthesis and provider |
Video | Generated video, face swaps | Prominent mark at the start and at appropriate points during playback | Embedded metadata; must survive re-encoding where feasible |
Virtual scenes | Immersive/VR environments | Prominent mark on the start screen and, as needed, during the experience | Embedded metadata |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.