Open a Chinese app and the photo you just generated now ships with a small confession baked into its pixels. The line between "made by a person" and "made by a model" is no longer left to guesswork.
As of September 1, 2025, China operates one of the strictest regimes on the planet for marking synthetic media. The rule is not a suggestion — it is backed by four government bodies and applies to anyone generating or distributing AI content at scale.
A digital ID on every synthetic file
The AI-Generated Synthetic Content Labeling Measures (标识办法) were issued on March 7, 2025 by the Cyberspace Administration of China (CAC) together with the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration.
The core idea is blunt: if a machine made it, the file has to say so. And "it" now covers text, images, audio, video, and virtual scenes — the full menu of generative media.
The two-layer system
The measures split the label into two kinds, and both matter:
- Explicit标识 (explicit label). A notice a human can actually see or hear — text at the start or end of an article, a voice prompt in an audio clip, a visible marker on a video's first frame.
- Implicit标识 (implicit label). Hidden data written into the file's metadata: who made it, which service, and a content ID. The measures also encourage a digital watermark (数字水印) embedded in the content itself.
The clever part is the distribution chain. A platform that republishes the content must check the metadata; if it finds a hidden tag, it has to add a visible notice around the post. If it detects an AI signature without metadata, it must mark the content as "suspected" synthetic.
Why it landed now
The measure didn't appear in a vacuum. It sits on top of two earlier rules:
- The deep-synthesis rules (深度合成规定), in force since January 10, 2023, which first required prominent labels on synthetic faces, voices, and scenes.
- The Generative AI Service Interim Measures (生成式人工智能服务管理暂行办法), effective August 15, 2023, which set the broader governance frame.
The 2025 step is notable because it pulls the distribution platform into the responsibility net, not just the generator. That closes a loophole where a labeled file could be stripped and reposted clean.
The numbers behind the rule
The scale of China's generative landscape justifies the heavy hand. As the measures took effect, state media reported that more than 490 large models (大模型) had been registered with the CAC, another 240-plus at the provincial level, and that generative-AI product users had reached about 230 million.
The underlying industry is large, too: CNNIC data cited by state outlets put China's AI industry scale above ¥700 billion (≈ US$99 billion) in 2024, growing more than 20% a year. When that much synthetic content flows through society, a tagging standard becomes infrastructure.
Who has to comply
The obligations cascade across three groups:
- Generators must embed both labels and keep logs.
- Platforms must verify metadata, surface notices, and offer users a declaration tool.
- App stores must check, at upload, whether an app provides AI-generation services and whether it handles labeling.
Users who want content without a visible label can request it, but the provider must log who received it and the user stays responsible for adding a label before public posting.
What it means outside China
China's dual-label model is now a de-facto template. The EU's AI Act and various US state bills are moving in the same direction — visible disclosure plus provenance metadata. A studio that builds labeling into its pipeline for the China market today is, in effect, pre-complying with tomorrow's Western rules.
The bigger shift is cultural: "assume it might be synthetic" is becoming the default posture for any piece of media with no clear provenance.
Honest limitations
This article describes rules that apply within China or to China-facing services; we did not audit how each platform implements them in practice, and enforcement depth varies by service and region. The model-registration and user-count figures are official CAC numbers reported by state media, not independently verified. Hidden labels can be stripped or forged, and the measures address detection but not perfect enforcement. We did not measure public awareness of the labels or their real-world effect on misinformation. Cross-border platforms with layered jurisdiction are only sketched here.
What readers can do now
- If you publish AI-generated content aimed at Chinese audiences, add both a visible notice and metadata tagging now — the rule has been live since September 1, 2025.
- Build detection-friendly habits into your workflow: keep AI watermarks and metadata intact, and never strip them before redistribution.
- For global teams, treat China's explicit-plus-implicit model as a preview of likely EU and US requirements and standardize labeling before it becomes mandatory everywhere.