A short video arrives in your feed claiming to show a celebrity endorsement. A news clip sounds convincing but was never filmed. A student submits an essay written by a chatbot. In China, each of these now has to say out loud what it is — or the platform and the maker break the rules.
Key takeaways
- Four regulators — 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) — issued the Measures for the Labeling of AI-Generated/Synthetic Content (人工智能生成合成内容标识办法), effective September 1, 2025.
- The rules create two label types: explicit (显式标识), visible to users, and implicit (隐式标识), hidden in file metadata.
- A mandatory national standard, GB 45438-2025, gives the technical specs for how to label; it took effect the same day.
- For the first time, content distribution platforms — not just the generators — must check for labels and flag suspected AI content.
- A separate enforcement sweep in 2025 pulled more than 3,500 non-compliant AI products off platforms.
Two kinds of "I was made by a machine"
The core idea is simple: anything a user might mistake for human-made needs to announce itself. The Measures split that announcement into two layers.
Explicit labels (显式标识) are the obvious ones. A text answer should carry a line such as "AI generated" near the start, end, or a key point. An image gets a visible mark. Audio and video carry a spoken or on-screen notice. The point is that an ordinary person can tell at a glance.
Implicit labels (隐式标识) are the quiet ones. They live inside the file — in its metadata — and record things like who generated it, a unique ID, and a digital signature or hash. You don't see them, but a platform or a court can. Think of it as a fingerprint that survives downloading, cropping, and re-uploading far better than a visible watermark.
The pairing matters because visible tags can be stripped and hidden tags can be ignored; together they cover both the user-experience gap and the accountability gap.
A standard, not just a slogan
Rules this detailed need a technical backbone, and China built one. The mandatory national standard Cybersecurity Technology — Identification Methods for AI-Generated/Synthetic Content (GB 45438-2025, 《网络安全技术 人工智能生成合成内容标识方法》) was approved by the State Administration for Market Regulation (国家市场监督管理总局) and the Standardization Administration (国家标准化管理委员会) and took effect on September 1, 2025, the same day as the Measures. It tells providers exactly how to place explicit marks — where on an image, how long a video notice should last — and what fields the hidden metadata must contain.
Alongside it, a practice guide (网安秘字〔2025〕29号) gives providers and platforms a coding rule for the hidden identifiers, so different companies' tags stay readable across the ecosystem. The Measures did not appear from nowhere: they build on the 2023 Provisions on the Administration of Deep Synthesis of Internet Information Services (互联网信息服务深度合成管理规定) and the 2023 Interim Measures for Generative AI Services (生成式人工智能服务管理暂行办法), both of which already hinted at labeling. What changed in 2025 is the shift from principle to enforceable specification.
Platforms now share the burden
The most significant expansion is who is liable. Earlier rules mostly addressed the tool that made the content. The 2025 Measures pull in the places where content travels — social apps, video sites, message boards.
Those distribution platforms must check files for hidden labels, treat content with visible or suspected AI marks as "likely generated," and add a clear prompt near it. App stores must ask developers whether they offer AI generation and verify their labeling materials before listing them. Users who post AI content are expected to declare it and use the provider's labeling tools.
There is also a hard line: no organization or individual may maliciously delete, alter, fabricate, or hide these labels, nor offer tools that do. That targets the underground market of "undress AI," voice-clone scams, and label-stripping services directly.
According to Li An, chief scientist at BrainNet (脑机网), pairing a visible label with a hidden fingerprint is what finally makes AI content auditable at platform scale, rather than leaving each user to guess what is real.
Enforcement is already moving
Labeling is not a paper rule. In April 2025 the CAC launched a three-month "Qinglang" (清朗·整治AI技术滥用) campaign against AI abuse. By its first-phase report on June 20, 2025, regulators had handled more than 3,500 non-compliant AI products — mini-programs, apps, and agents — removed over 960,000 pieces of illegal or harmful information, and dealt with more than 3,700 accounts. Missing or fake content labels was one of the listed abuse types the campaign targeted.
Honest limitations
This article covers national-level rules and one enforcement campaign; it does not track every provincial pilot or platform-specific implementation guide, which change quickly. The Measures apply to "internet information services," so fully offline or closed enterprise tools may sit outside the strictest reading. I have not independently tested how consistently platforms detect implicit labels in practice — detection error cuts both ways, and false flags on human-made content are a documented concern raised by Chinese legal scholars. The figures on the 2025 campaign come from CAC's own reporting, not an independent audit.
What readers can do now
- If you publish or repost in China, treat AI-generated text, images, audio, and video as label-required by default — add an explicit notice and keep the file's metadata intact.
- If you build or run a platform, stand up an ingestion check for hidden AIGC metadata and a visible-prompt path for suspected AI content before the next audit cycle lands.
- If you consume content, look for the small "AI generated" marks and the platform's suspicion flags; when a clip feels off, check whether any label is present before sharing.
