A user in Chengdu scrolled past a polished "news" clip last week and noticed a small "AI生成" tag in the corner. A year ago that tag did not exist; now it is required by law. China's rules for marking AI-produced (生成式内容) media moved from guideline to enforceable standard on September 1, 2025.
Key takeaways
- The Measures for Labeling AI-Generated/Synthetic Content (《人工智能生成合成内容标识办法》), issued by the CAC, MIIT, Ministry of Public Security, and NRTA, took effect September 1, 2025.
- A companion mandatory national standard, GB 45438-2025, sets technical rules for both visible (显式) and hidden (隐式) labels.
- Visible text labels must include both "人工智能/AI" and "生成/合成"; video tags must stay on screen at least 2 seconds and sit in a corner at no less than 5% of the shortest edge.
- Hidden metadata must carry an "AIGC" field with producer identity and a content ID for tracing.
- CNNC reported China's generative AI users reached 515 million by June 2025, making labeling a mass-scale task.
Two kinds of labels
The rules draw a hard line between what people see and what machines record. An explicit (显式) label is the on-screen marker — a corner tag on a video, a line at the top of an article, a beep pattern in audio. An implicit (隐式) label is embedded in file metadata: a fingerprint that survives downloads, file copies, and re-uploads, carrying the producer's code, a unique content number, and a hash for verification.
What providers must do
Service providers — apps, mini-programs, websites, and API services that offer generation to the public in China — carry the main load. They must:
- Add visible labels to generated text, audio, image, video, and virtual scenes.
- Bake hidden labels into file metadata on export, copy, and download.
- State labeling rules clearly in user agreements and prompt users to comply.
- App stores must check, at upload, whether an app offers generative services and verify its labeling materials.
Users are not off the hook. Anyone publishing AI-produced (生成式内容) media through a platform must use the provider's labeling tools and declare the content. Deleting, forging, or hiding labels is prohibited, and the rules authorize penalties ranging from warnings to business suspension.
The technical standard behind it
GB 45438-2025, published by the State Administration for Market Regulation, is the enforceable backbone. It specifies, for example, that text labels must combine "AI/人工智能" with "生成/合成," that image labels sit at an edge with a minimum size, and that video labels remain visible for at least two seconds at normal playback. The hidden layer must include an "AIGC" metadata field. A six-month transition period let companies upgrade products; services not compliant by the deadline are barred from serving the public in China.
A first-of-its-kind combo
China's approach pairs a regulatory measure with a mandatory national standard — the first such "办法 + 强标" combination for AI content, according to the China Electronics Standardization Institute. The measure sets who-must-do-what; the standard sets how. Together they create a closed loop of generation, transmission, distribution, and accountability that earlier rules on deep synthesis and generative services only sketched. For an industry moving as fast as generative media, locking the technical spec alongside the legal one is the part that makes compliance checkable rather than aspirational.
Why it matters for global builders
The standard is domestic, but its ripple is international. Any product that touches users in China — including overseas text-to-video (文生视频) and digital human (数字人) tools with a China-facing app or API — must emit both label types or risk removal. Because the hidden label is metadata-based, it travels with the file; a clip made abroad and re-uploaded in China can still be inspected. For multinational platforms, labeling becomes a build-time requirement, not a China-only patch.
Honest limitations
This article summarizes publicly released Chinese government documents (CAC, SAMR, gov.cn) and state-media explainers; it is not legal advice. We did not audit how individual platforms implemented the standard, and enforcement intensity varies by provider and region. The "515 million users" figure comes from a CNNC generative-AI report cited by state media, not from a regulatory filing. Translations of label wording ("AI生成," "人工智能生成合成") follow the standard's Chinese text; exact UI phrasing may differ by platform. We did not test whether every major app's hidden metadata currently passes the GB 45438-2025 spec.
What readers can do now
- If you publish AI-produced media in China, turn on your platform's built-in label and never strip it before reposting.
- Builders shipping to China should add both visible tags and an "AIGC" metadata field at export, not as an afterthought.
- Audiences should treat an "AI生成" marker as a cue to verify, not as a guarantee of accuracy or safety.