A user scrolling short videos in China now stops at a clip of a dancing artifact — and notices a small line in the corner: "AI generated." Across the country, a fraud investigator opens a doctored audio file and, instead of guessing, reads a hidden fingerprint buried in the file's metadata that points straight back to the service that made it. These are not features a platform chose voluntarily. They are now the law.
The change came from a tightly paired package: the 《人工智能生成合成内容标识办法》 (Measures for Labeling AI-Generated Synthetic Content), issued jointly by the Cyberspace Administration of China (国家网信办), 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). Announced in mid-March 2025, it took effect on September 1, 2025. Running alongside it is a mandatory national standard, GB 45438-2025 《网络安全技术 人工智能生成合成内容标识方法》 (Cybersecurity technology — Labeling method for content generated by artificial intelligence), released 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 synchronized to the same September 1 start date.
Why the rule exists
Generative AI (生成式人工智能) stopped being a novelty and became infrastructure. According to a CNNIC report cited in Chinese state media, generative-AI users in China surpassed 500 million by mid-2025. That scale cut both ways: the same tools that helped small studios produce ads in minutes also fed a rise in "AI face-swap" scams, fabricated news clips, and copyright disputes that left platforms and courts guessing who made what.
Public pressure built for a clear rule rather than a patchwork of platform policies. The four agencies framed the 标识办法 (labeling measures) as a response to that demand — a way to keep innovation moving while giving every piece of synthetic media a traceable identity.
Two labels, two jobs
The core idea is deceptively simple: a label can be seen, or it can be hidden, and both matter.
- Explicit labeling (显式标识) is the visible kind. A text post must carry an "AI generated" note at its start, end, or a key point. An image gets a visible caption in its lower-right corner. An audio clip opens with a spoken or tonal prompt. A video shows the same prompt on its first frame. The point is that an ordinary viewer can tell, at a glance, that a machine — not a person — produced the content.
- Implicit labeling (隐式标识) is the invisible kind. It lives inside the file's metadata: a marker that must include the "AIGC" symbol, plus the service provider's identity, a unique content number, and a digital signature or hash. You cannot crop it out the way you might blur a corner watermark. It is the part built for auditors, police, and courts.
The two work as a pair. The explicit tag protects the viewer in the moment; the implicit tag protects the truth later, when someone needs to prove origin.
Who actually has to comply
The rule does not only point at the company that pressed "generate." It spreads responsibility across the chain:
- Service providers that run the generation tools must attach both explicit and implicit labels, and must preserve those labels when a user downloads, copies, or exports the file.
- Content distribution platforms — the apps where people actually watch and share — must apply technical measures so labeled content stays labeled as it travels.
- Users and third parties are explicitly barred from maliciously deleting, tampering with, forging, or hiding the required labels, and from offering tools that do so for others.
That last point is the quiet hammer. A sticker you can peel off is not much of a sticker; the law treats deliberate removal as its own offense.
How the standard turns intent into mechanics
The GB 45438-2025 standard is where the "how" lives. It specifies, for example, the exact metadata format for hidden labels and typical placement for visible ones across text, images, audio, video, and virtual scenes. A companion practice guide (《网络安全标准实践指南——人工智能生成合成内容标识 服务提供者编码规则》, document 网安秘字〔2025〕29号), released the same day as the measures, hands each provider a platform code so the hidden fingerprint can name its maker unambiguously.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the "measures-plus-mandatory-standard" structure is what makes this effort distinct: a regulation sets the duty, and a binding technical standard tells engineers precisely what to build, leaving less room for creative non-compliance.
A six-month runway, not a cliff
The agencies did not flip the switch the day they announced it. The gap between the March announcement and the September effective date was a deliberate six-month transition period, meant to let platforms upgrade products and adopt low-cost labeling schemes rather than face instant penalties. That "firm rule, soft landing" pacing shows up repeatedly in Chinese tech regulation and is worth reading as a signal: the state wants compliance, not a wave of shutdowns.
Where this fits in the bigger picture
The labeling package does not stand alone. It leans on earlier building blocks — the 《互联网信息服务深度合成管理规定》 (Provisions on the Administration of Deep Synthesis (深度合成) of Internet Information Services) and the 《生成式人工智能服务管理暂行办法》 (Interim Measures for the Management of Generative AI Services) — and is itself part of a broader push to govern content at the point of creation rather than only at the point of distribution. For an industry that spent years arguing about who is responsible when a model lies, the answer here is blunt: the maker, the carrier, and the tamperer each carry a defined duty.
What readers can do now
- If you publish AI content in China, audit your export pipeline today — confirm both a visible tag and a metadata watermark survive download and re-upload.
- If you build tools, read GB 45438-2025 and the provider-coding practice guide; bake the "AIGC" metadata field into your file formats before September deadlines become enforcement memories.
- If you are a consumer, start treating the "AI generated" corner note as routine — and report content where the label has clearly been stripped.
Honest limitations
This article focuses on the labeling measures and the GB 45438-2025 standard as adopted; it does not track every local pilot, platform-specific rule, or the separate recommended standards still being drafted for individual file formats. My reading of enforcement intensity is inferred from the stated transition period and prior regulatory patterns, not from published penalty statistics, which were not available at writing. The user-count figure is drawn from a CNNIC report relayed by state media and should be treated as an official estimate rather than an independently audited number. I have not assessed how the rules interact with non-Chinese jurisdictions or with end-to-end encrypted services, where hidden metadata may not survive.