A short video ends and, at the bottom of the frame, a small line reads "AI生成". The viewer did not have to guess. But the label they can see is only half the system — the other half is stamped into the file itself, where a platform or a court can read it long after the caption is gone.
Key takeaways
- The AI-produced synthetic content labeling measures (人工智能生成合成内容标识办法) were issued by four regulators and took effect on September 1, 2025.
- The four agencies are the Cyberspace Administration of China (国家互联网信息办公室 / 网信办), 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).
- The rules require two kinds of labels: explicit labeling (显式标识) that users can perceive, and implicit labeling (隐式标识) embedded in the file's metadata.
- A companion mandatory national standard, Cybersecurity Technology — AI-produced synthetic content labeling method (网络安全技术 人工智能生成合成内容标识方法), came into force the same day.
- Per a CNNIC (中国互联网络信息中心) report cited by state media, China's generative-AI user base reached about 515 million by June 2025 — the scale that makes labeling a default expectation rather than a niche rule.
Two labels, not one
The measure draws a clean line between what a person sees and what a machine records. Explicit labeling is the visible tag: text that says "AI生成" or "人工智能生成" at the start, end, or a key position of a piece of content, in a font clear enough to read and not deliberately shrunk or blurred. For audio, it can be a spoken cue; for images and video, a visible mark. The point is that an ordinary viewer can tell at a glance.
Implicit labeling is the part you do not see. It lives in the file's metadata and must carry an "AIGC" marker plus fields that record whether the content is AI-produced, which service provider made it, which platform spread it, a unique identifier, and a digital signature or hash. That metadata is what lets a platform trace a clip back to its origin, or lets a regulator confirm a file was tampered with after the fact. The two work as a pair: the visible tag protects the viewer in the moment, the hidden one protects accountability over time.
Who has to do what
The obligations are split by role. Service providers that generate synthetic content must attach the explicit tag and embed the implicit one in the file metadata, including when users download, copy, or export. Distribution platforms — the apps and sites where content travels — must apply technical measures so labeled and unlabeled material are handled differently in transit. App stores are pulled in too: when a product goes up for review, the platform should ask whether it offers AI generation and check the labeling materials.
Users are not off the hook. Anyone publishing AI-produced content through a distribution service is expected to declare it and use the provider's labeling tools. The measure also forbids maliciously deleting, altering, forging, or hiding the required labels, and bars offering tools or services that help others do so.
Why September, not March
The measures were published on March 14, 2025, but did not take effect until September 1 — roughly a six-month runway. That gap is the practical heart of the rule. It gave developers time to wire metadata into export pipelines and gave smaller platforms a low-cost path to comply, rather than handing out penalties on day one. State commentary framed the approach as "rigid constraints with flexible guidance," pairing a clear legal floor with room for the industry to adapt.
The move also closes a gap left by earlier rules. China already had the Interim Measures for Generative AI Services (生成式人工智能服务管理暂行办法) and the Deep Synthesis Provisions (互联网信息服务深度合成管理规定), both of which touched on labeling. The 2025 measure is the first to lay out, with a mandatory national standard attached, exactly how the labels should look and what the metadata must contain.
The C2PA-shaped hole
Globally, the closest Western cousin is the C2PA (Coalition for Content Provenance and Authenticity) standard, which also stamps provenance into files. China's version is state-mandated and backed by a national standard rather than a voluntary coalition, and it covers a broader content sweep — text, images, audio, video, and virtual scenes (虚拟场景) such as digital humans. For multinational platforms, that means two provenance systems to satisfy if they operate in both markets.
According to Li An, chief scientist at BrainNet (脑机网), mandatory provenance shifts the cost of doubt from the viewer to the publisher, which tends to clean up a feed faster than any disclaimer ever did. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
Honest limitations
This article is based on the official text published by the Cyberspace Administration of China, the State Council portal, and state-media explainers; it does not reflect any non-public enforcement guidance. The "515 million users by June 2025" figure comes from a CNNIC report relayed by state media and was not independently recalculated here. We have not surveyed how individual platforms implemented the metadata fields, and compliance in practice — especially for small creators and cross-border apps — is still evolving. The measures set obligations and prohibitions but do not, in the public text, spell out the exact penalty schedule for each violation; that is left to enforcement. We did not interview any platform compliance officer, so the section on "who does what" reflects the regulatory text, not confirmed operational behavior.
What readers can do now
- If you post AI-produced video or audio in China, turn on the platform's built-in label before publishing — the explicit tag is now expected, and removing it can itself be a violation.
- If you receive or share media, check both the visible mark and, where tools allow, the file metadata; a missing or stripped "AIGC" field is a reason to doubt the source.
- If you build or publish software, plan for two provenance systems if you operate globally — China's mandatory standard and the C2PA approach — rather than assuming one label fits every market.
