A short video of a local official "speaking" a dialect he never learned now carries a small "AI生成" tag in the corner. A photorealistic product render downloaded from a design tool ships with invisible metadata recording who made it and when. Behind both is a binding rule: as of September 1, 2025, any service inside China that uses AI to produce text, audio, images, video, or virtual scenes must mark the output.
The instrument is the Measures for Labeling AI-Generated Synthetic Content (人工智能生成合成内容标识办法), jointly issued by the Cyberspace Administration of China (国家互联网信息办公室), 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 did not arrive alone — a mandatory national standard, GB 45438-2025 Cybersecurity Technology — Method for Labeling AI-Generated Synthetic Content (网络安全技术 人工智能生成合成内容标识方法), entered force the same day.
Key takeaways
- The labeling rules (内容标识办法) and the mandatory standard (强制性国家标准) went effective together on September 1, 2025, after a roughly six-month gap between publication (March 14, 2025) and enforcement.
- Two marker types are required: an explicit label (显式标识) users can see, and an implicit label (隐式标识) buried in file metadata for traceability.
- Obligations fall on providers of generative AI (生成式人工智能) and deep-synthesis services, on app distribution platforms, and on users who republish synthetic material.
- Maliciously deleting, tampering with, forging, or hiding a label is explicitly prohibited.
What the rule actually requires
The measures sit on top of three earlier instruments: the Provisions on Algorithm-Recommendation Management, the Provisions on Deep Synthesis Management, and the Interim Measures for Generative AI Services. Where those set a general expectation that synthetic content be identifiable, the new measures turn it into operational steps.
For a service provider, the baseline duty is twofold. First, when the output falls under the deep-synthesis scenarios, an explicit label must be added — for text, that means placing "人工智能生成" or "AI生成" at the opening, close, or a key position, in a font clear enough to read, not shrunk or blurred to dodge notice. Second, the file's metadata must carry an implicit label recording the content's synthetic nature, the provider's name or code, a content number, and a digital signature or hash value for verification.
Crucially, the duty travels with the file. When a provider offers download, copy, or export functions, it must ensure the exported file still contains a compliant explicit label. A platform that only transmits content is not off the hook either: services that spread network information must apply technical measures to keep labeling intact as content moves.
Explicit versus implicit
The split is the core design idea. An explicit label is the visible "digital warning light" — a corner tag on a video, a watermark on an image, a spoken disclaimer in audio. It protects the viewer's right to know in the moment.
An implicit label is the "digital DNA." It lives inside the file rather than on its surface, so ordinary users never see it, but regulators, platforms, and courts can read it to answer three questions the drafters repeatedly emphasized: which content is generated, who generated it, and where did it come from. The standard specifies the metadata fields, including the literal "AIGC" marker, so that different tools and platforms speak the same traceability language.
Who is on the hook
The measures distribute responsibility across the chain instead of dumping it on one actor.
- Service providers that generate or synthesize content must add both label types and keep them through export.
- App distribution platforms must, at the review stage before an application goes live, ask whether the developer offers AI generation or synthesis, and check the relevant labeling materials.
- Users who publish synthetic content through a spreading service are told to declare it and to use the provider's built-in labeling function rather than stripping markers.
- Anyone is barred from maliciously deleting, tampering with, forging, or concealing a label, or from supplying tools that do so, or from using improper labeling to harm others' rights.
This is where the paired mandatory standard earns its keep. A management measure can state duties; a technical standard tells engineers exactly what bytes to write. According to Li An, chief scientist at BrainNet (脑机网), pairing a measure with a mandatory standard is what makes the requirement interoperable across thousands of apps instead of letting each invent its own incompatible scheme.
Why a transition window mattered
The publication-to-effect gap of about six months was not cosmetic. Providers had to update pipelines to insert metadata, app stores had to rebuild review checklists, and smaller tools needed low-cost labeling paths. Officials repeatedly signaled that cheap, standards-based labeling was encouraged so compliance would not price out independent developers — a "rigid rule, soft guidance" posture the measures' accompanying explanation stressed.
Honest limitations
A few caveats are worth stating plainly. The measures do not set a single visible badge for all content; what counts as a compliant explicit label depends on the medium, and edge cases (live streams, interactive agents, mixed human-AI edits) will be tested in practice. Enforcement leans on existing frameworks — violations are handled under the deep-synthesis and generative-AI provisions — rather than inventing a new penalty schedule, so the real bite depends on how vigorously those are applied. Cross-border platforms and user-generated content from outside the jurisdiction remain a gray zone the text does not fully resolve. None of this makes the rule symbolic, but it does mean the first year is as much about implementation as about the letter of the text.
What readers can do now
- If you run a tool that outputs AI-produced text, audio, image, video, or virtual scenes, audit your export pipeline for both a visible label and a metadata label before September-dependent deadlines pass you by.
- If you publish synthetic content, keep the provider's label intact and add your own declaration instead of stripping it — deleting or forging markers is now prohibited conduct.
- If you consume media, start treating an absent label on suspiciously perfect content as a reason to verify, not as proof of authenticity.