A video produced by a model now carries a fingerprint whether the viewer sees it or not. In China, that is no longer optional. A set of rules that took effect at the start of September 2025 turned labeling from a courtesy into a legal requirement for any service that generates synthetic text, images, sound, or footage.
The change matters well beyond China's border because the country is now one of the largest producers of generative media on earth. When the plumbing for tagging is mandated at this scale, the habits and the tooling spread with the content.
Two kinds of label
The rules, formally the Measures for the Labeling of AI-Generated and Synthesized Content (人工智能生成合成内容标识办法), were issued jointly by four bodies: the Cyberspace Administration of China (国家网信办), the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration.
They define two distinct marks:
- Explicit label (显式标识): a tag a user can actually perceive — text, sound, or a graphic placed on the content or in the interface. On a clip, that might be a spoken notice or an on-screen icon; on text, a line at the top or bottom saying it was machine-made.
- Implicit label (隐式标识): data buried in the file's metadata that a person normally cannot see — recording the content's nature, the service provider's name or code, and a content identifier for tracing.
Both are required in the cases the rules cover. A provider offering deep synthesis (深度合成) services under the earlier framework must add the explicit tag where the law spells it out, and must embed the implicit tag in the file whenever someone downloads, copies, or exports.
Who has to do what
The obligations are spread across the chain, not dumped on one party:
- Service providers that generate synthetic content must attach the explicit label and write the implicit one into metadata, and must explain the labeling method in their user agreements.
- Distribution platforms — the social apps and sites where content spreads — must use technical measures to manage how labeled material travels.
- App stores must ask, at review time, whether an app offers generative services, and must check the provider's labeling materials before letting it list.
- Users who post synthetic content are told to declare it and use the provider's labeling tools, and everyone is barred from maliciously deleting, tampering with, forging, or hiding a required mark.
That last clause is the enforcement teeth. Stripping a watermark to make a clip look human-made is explicitly forbidden, as is building a tool that does it for others.
How platforms moved first
The platforms did not wait for the deadline. Xinhua reported that before 1 September 2025, Tencent, Douyin, Kuaishou, and Bilibili each published detailed rules of their own. Douyin, in particular, switched on two features: a content-labeling tool that helps creators tag their posts, and a metadata read-write function that can recognize and write the hidden marker for traceability.
For a creator, the practical effect is that the "stamp" is increasingly built into the publish button rather than left to conscience. The hope is that provenance survives every re-share instead of evaporating the moment a file is saved and re-uploaded.
The mandatory standard behind it
The measures did not arrive alone. On the same day, a mandatory national standard — Cybersecurity Technology: Identification Methods for AI-Generated Content — came into force, pairing management rules with a technical spec. Officials framed the pairing as a new phase: standards and regulation advancing together rather than one lagging the other.
The scale of what is being governed is large. A senior engineer at the National Computer Network Emergency Response Technical Team and Coordination Center was quoted saying more than 300 generative products and 3,000 services had been registered in China. And a CNNIC report cited by Xinhua put the country's AI industry above 700 billion yuan (about US$99 billion, HK$770 billion) in scale for 2024, growing more than 20 percent a year. Those figures are context, not the legal text, but they show why a uniform tag became hard to avoid.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), labeling is becoming table stakes for any platform that wants to ship generative features, and the teams that treat it as a product feature rather than a compliance tax will adapt faster.
The reasoning is that provenance, once standardized, stops being a burden and starts being a trust signal — a clip that can prove its origin is easier to license, safer to advertise against, and simpler to defend in a dispute. The platforms that build the tag into the creative flow, instead of bolting it on at upload, are the ones creators will tolerate.
Honest limitations
- The measures are a Chinese domestic regulation; they bind services operating in China and Chinese platforms, but they are not a global standard, and enforcement reach abroad is limited.
- The "300 products, 3,000 services" figure is a quoted expert estimate from a government-adjacent engineer, not a precise published registry count we could independently verify.
- The 700 billion yuan industry scale is drawn from a CNNIC report cited by Xinhua for 2024; it is an aggregate estimate, and our US dollar and Hong Kong dollar conversions use fixed rates of 7.1 and 1.1 respectively, not live exchange.
- We did not test platform labeling tools ourselves; the Douyin feature descriptions come from Xinhua's reporting, not a hands-on audit.
- The rules set obligations but leave room in how providers implement explicit tags, so user-visible marks will differ from app to app.
What readers can do now
- Label at the source: if you generate clips or images with a China-based tool, use its built-in tagging before downloading, because the implicit metadata is your easiest proof of origin later.
- Keep the metadata intact: avoid re-encoding or screenshotting in ways that strip file metadata; export the original file so the hidden marker survives the share.
- Build a provenance habit: add a short visible "AI-generated" note on anything synthetic you publish, even where the law does not yet require it, so your work stays licensable and defensible.