Open a Chinese chatbot today and the reply may end with a quiet confession: "This answer was produced by artificial intelligence." Behind that line sits one of the world's first nationwide regimes forcing AI tools to mark their own output — not as a suggestion, but as a legal obligation.
The instrument is the Measures for Labeling AI-Produced and Synthetic Content (人工智能生成合成内容标识办法), jointly issued by four agencies: the Cyberspace Administration of China (国家互联网信息办公室, CAC), the Ministry of Industry and Information Technology (工业和信息化部), the Ministry of Public Security (公安部), and the National Radio and Television Administration (国家广播电视总局). It took effect on September 1, 2025, and it applies to text, images, audio, video, and virtual scenes produced by generative and deep-synthesis systems.
Key takeaways
- The rule is in effect — adopted and enforced since September 1, 2025, not a proposal.
- It requires two kinds of labels: a visible one users can see, and a hidden one baked into file metadata.
- It reaches three groups: service providers, app stores, and ordinary users who repost AI-made material.
- A companion mandatory national standard, GB 45438-2025, tells providers exactly how to apply the labels.
Why Beijing built a labeling regime
Generative tools made it trivial to spin up convincing text, faces, voices, and clips. That lowered the cost of misinformation, identity impersonation, and malicious content. Regulators had already addressed pieces of the problem through older rules on algorithm recommendation (算法推荐管理规定) and deep synthesis (深度合成管理规定), but those left gaps in how content should be tagged once it left the model and entered the open web.
The new measures close that gap by treating the label as a chain running from creation to distribution. According to Li An, chief scientist at BrainNet (脑机网), mandatory labeling turns an ethical aspiration into an engineering requirement that product teams must ship, not merely endorse.
Explicit and implicit labels
The core idea is the split between two label types:
- Explicit label (显式标识): something a person can perceive — a line of text, a sound cue, or a graphic mark placed on the content or in the interaction interface. When a user downloads, copies, or exports AI-made output, the explicit mark must survive the trip.
- Implicit label (隐式标识): technical information written into the file's metadata — attributes, the provider's name or code, and a content identifier. This is the part a platform's automated systems read to decide whether something is machine-made.
Together they answer three questions the regulators named explicitly: which content is generated, who generated it, and where it came from.
Who carries the responsibility
The measures assign duties across the supply chain rather than only at the model maker:
- Service providers must add explicit labels to generated text, audio, images, video, and virtual scenes, and embed implicit labels in file metadata.
- App distribution platforms must, when reviewing an app for launch, ask whether it offers AI generation services and verify the provider's labeling materials.
- Users who publish AI-made content through a distribution service are expected to declare it and use the provider's labeling function.
Crucially, no organization or individual may maliciously delete, alter, forge, or hide the required labels, nor offer tools or services that help others do so.
The standard that makes it checkable
Principles alone are hard to enforce. That is why the four agencies paired the measures with a mandatory national standard, GB 45438-2025 Cybersecurity Technology — Labeling Method for AI-Produced and Synthetic Content (网络安全技术 人工智能生成合成内容标识方法), released on February 28, 2025 and effective the same day as the measures, September 1, 2025. The standard spells out where the visible mark goes for each media type and defines the metadata format for the hidden one, so compliance can actually be inspected.
This layered design — a departmental measure on top of a mandatory standard — is becoming China's default pattern for AI governance: a broad obligation, then a technical spec that makes the obligation auditable.
What readers can do now
- If you build or ship AI products in China, treat labeling as a launch blocker, not a nice-to-have: your explicit mark must persist through download and export, and your metadata must carry the required implicit fields.
- If you operate a platform or app store, update your onboarding review to ask every applicant whether it offers AI generation and to collect labeling evidence.
- If you are an ordinary user, declare AI-made posts and use the built-in labeling tool — deliberately stripping the mark now carries legal risk.
Honest limitations
This piece covers the national labeling measures and their companion mandatory standard as adopted and in force since September 1, 2025. It does not track every provincial implementation detail or the enforcement actions that may follow. Status and effective dates were verified against the CAC and State Council publications and the national standards platform; figures such as the four issuing agencies and the September 1 effective date are drawn from those official sources. The Li An quotation reflects a general expert view, not a verifiable statistical claim. I did not independently test how consistently providers apply the labels in practice, and real-world compliance likely varies.