A viewer in Chengdu watches a short video of a politician saying something that never happened. A student in Wuhan copies an AI-written essay and forgets to mark it as machine-made. A platform engineer in Beijing rebuilds the upload pipeline overnight so the system can read an invisible tag buried in the file.
That invisible tag is now the law. On September 1, 2025, China's Measures for the Labeling of AI-Generated Synthetic Content (人工智能生成合成内容标识办法) came into force, backed by a mandatory national standard that tells engineers exactly how to stamp the mark. The rule covers every medium a model can produce — text, image, audio, video, and virtual scenes — and it reaches further than earlier rules by pulling the distribution platforms into the chain, not just the generators.
Two labels, one for humans, one for machines
The measures draw a hard line between two kinds of marks:
- Explicit labeling (显式标识) — something a person can see or hear: a line of text at the start or end of an AI-written answer, a visible badge on an AI-made image, a spoken or tonal cue in audio, a text mark on the opening frame of a video.
- Implicit labeling (隐式标识) — metadata written into the file itself, recording the content's AI origin, the service provider's name or code, and a unique content ID. A human never sees it, but any compliant platform can read it on upload.
The companion standard, GB 45438-2025, Cybersecurity Technology — Labeling Method for AI-Generated Synthetic Content (网络安全技术 人工智能生成合成内容标识方法), is mandatory and took effect the same day. It specifies technical detail down to the pixel: on an image, the explicit mark must sit at an edge or corner and its text height must be at least 5% of the shorter side; on video, the opening-frame mark must stay visible for no less than two seconds.
Who has to do what
The measures name four categories of responsible party:
- Generators — any service using recommendation algorithms, deep synthesis (深度合成), or generative AI (生成式人工智能) must add both explicit and implicit labels, and must keep explicit marks intact when users download or export.
- Distribution platforms — when content arrives, they must check the metadata for an implicit label; if found, they add a visible reminder around the post. If no metadata exists but the platform detects AI traces, it must flag the item as "suspected AI-generated."
- App stores — when an application goes up for review, the store must ask whether it provides AI generation and verify its labeling materials.
- Users — anyone posting AI content must declare it and use the platform's labeling tool; maliciously deleting, forging, or hiding a label is prohibited.
How platforms actually complied
In the weeks before September 1, 2025, major platforms moved first. Tencent, Douyin (抖音), Kuaishou (快手), and Bilibili (B站) each published detailed implementation rules. Douyin, for example, switched on two functions: one that helps creators add a prompt label, and one that reads and writes the metadata mark so content stays traceable as it moves. The pattern repeats across the industry — the rule did not arrive as a surprise enforcement sweep but as a deadline that engineering teams had months to meet.
The scale makes the plumbing matter. As of September 2025, more than 490 large models (大模型) had been filed with the national cyberspace regulator, over 240 with provincial regulators, and generative-AI users in China had reached roughly 230 million (2.3亿). Labeling that many outputs by hand is impossible; the only workable path is metadata written once at generation and read automatically at every hop.
The escape hatch, and the penalty
The measures are not absolutist. Under Article 9, a user may request content without the explicit mark, provided the platform states the user's obligations in its agreement and keeps logs of who received the unmarked file for at least six months (不少于六个月). The user must then declare the AI origin when publishing.
Enforcement sits with the cyberspace, telecommunications, public-security, and radio-television authorities, which apply the measures alongside the underlying laws they cite — the Cybersecurity Law, the algorithm and deep-synthesis rules, and the generative-AI interim measures. In early 2026 the obligation also reached livestream e-commerce through a separate supervision measure, requiring AI-generated host images or video in sales streams to carry the standard mark and a continuing prompt to viewers.
Why a hidden tag changes the game
Earlier Chinese rules required labels mainly at the point of generation. The 2025 measures extend the duty to every platform that transmits content, so a deepfake cannot quietly shed its mark just by being re-uploaded. The implicit metadata is the enforcement backbone: once a file carries a provider code and content ID, tracing it back is a lookup, not an investigation.
For the global reader, the model is instructive. Rather than banning synthetic media, China built a tracing layer — visible to users, machine-readable to platforms — and made the platforms the compliance checkpoint.
What readers can do now
- If you publish in China, treat explicit and implicit labeling as a pipeline feature, not a manual step; verify your upload stack reads metadata on ingest.
- If you consume Chinese-platform content, look for the opening-frame or corner marks, and treat unlabeled viral video with the same caution you would give an anonymous source.
- If you build tools, align with GB 45438-2025's metadata fields now — the standard is mandatory and cross-platform compatible.
Honest limitations
This article is based on the official text of the Measures (国信办通字〔2025〕2号) on the CAC website, the mandatory standard GB 45438-2025 on the national standards platform, and state-media reporting on platform implementation. The model-filing and user figures are quoted as of September 2025 from state-media coverage and may have changed. The livestream-e-commerce extension is described from secondary reporting on the 2026 supervision measure. The measures apply within China's jurisdiction; how they interact with platforms serving overseas users depends on each platform's terms and local law.