AI-Generated Content Now Has to Declare Itself

Starting October 1, AI-generated content circulating in China has to wear a label. The Measures for Labeling AI-Generated Synthetic Content, jointly issued by four bodies — the Cyberspace Administration of China, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the National Radio and Television Administration — took effect that day. Text, images, audio, video, virtual scenes: any information generated or synthesized using AI technology must carry a label.

The first changes ordinary users notice come from a handful of familiar apps. WeChat, WeChat Channels, Douyin, Bilibili, Weibo, and Xiaohongshu updated their user agreements around the time the measures landed, requiring publishers to proactively and conspicuously label AI-generated content. The consequences of skipping the label are spelled out bluntly: reduced reach, takedown, and in serious cases a banned account. One short-video platform, when users scroll onto content like “cute kids washing vegetables and tending the fire,” adds a small line of text below the frame noting that the content was made with AI technology and that fictional content should be viewed with caution. Someone screenshotted that line and posted it, saying the elders in their family finally believed those children were fake.

Two Kinds of Labels: One for People, One for Machines

The measures split labeling into two types: explicit and implicit. An explicit label is the kind users can clearly perceive — a corner badge on the image, a notice before playback, a spoken declaration at the start of an audio track. An implicit label is embedded in the file's metadata; invisible to the eye, but readable by platforms and regulators using tools, to answer questions like “who generated this content, and which service did it come from?”

Responsibility is divided by stage. Service providers must keep labels intact at the points of generation, distribution, and download; app distribution platforms must vet whether the apps they list are capable of adding labels; individual users must proactively declare when they publish generated content. This chain moves content review — previously possible only after the fact — forward into product design and app listing.

For marketing accounts that make a living by scraping and rewriting other people's work, implicit labels are a hard constraint. In the past you could change a video's thumbnail and tweak the color and push it out as your own original; now the metadata may still carry source information, and platforms can trace it back to the original service provider to see whether that provider fulfilled its labeling obligations.

Why a Dedicated Regulation for This One Thing

Labeling looks like a small matter, but it addresses the most troublesome part of AI content: identification at scale. Xiao Youdan, a researcher at the Chinese Academy of Sciences' Institutes of Science and Development, noted in a commentary that the main goal of a labeling regime is regulation that is large-scale, highly reliable, and low-cost. Human reviewers can't keep up with content moderation, and model-based judgments tend to produce false positives, whereas a label is a signal that can be checked automatically at low cost.

Another scenario it targets is fraud. The National Anti-Fraud Center app has already launched an “AI Content Identification” feature: upload text, an image, or a video and get a preliminary judgment within seconds. The “National Anti-Fraud AI” app, developed by the Shanghai Public Security Bureau under the guidance of the Criminal Investigation Bureau of the Ministry of Public Security, went live at the same time; enter a suspicious scenario and it breaks down common fraud tactics. Face-swapping and voice cloning have shown up more and more often in telecom and online fraud over the past two years, and labeling AI content adds another layer of friction to that chain.

The labeling measures are not an isolated move. The Supreme People's Court issued its Opinions on the Lawful Adjudication of Cases Involving Artificial Intelligence Disputes, writing AI face-swapping and voice cloning, torts caused by AI hallucinations, and online doxxing into its adjudication approach; the Cyberspace Administration of China launched the “Qinglang · Rectifying Disorder in AI Applications” special campaign, listing seven categories of targets, including failure to file large-model registrations, AI data poisoning, and inadequate implementation of labels on generated and synthesized content; and the AI Safety Governance Framework 3.0, released on September 14, extended the scope of governance from individual technical steps to the entire process of data, algorithms, models, applications, and supply chains. On October 8, the industry's first safety grading standard for generative AI services took effect, and standards work related to AI agents identified 13 major categories and 97 risks. These documents landed close together in time; read as a whole, they trace a governance path from content to conduct, from a single model to the entire chain.

Other Countries Are Doing the Same Thing

Zoom out and you see that legislating on labeling has become a global consensus in recent years. The United States, Singapore, the European Union, and the United Kingdom are all advancing their own generative AI labeling rules. The EU AI Act's transparency provisions take effect on August 2, 2026, requiring disclosure of AI interactions, marking of synthetic content, and labeling of deepfakes, with fines of up to €35 million or 7% of global annual turnover, whichever is higher. To comply, OpenAI announced it would add invisible watermarks to text generated by ChatGPT and Codex within the EU.

The difference lies in where the pressure is applied. The EU uses risk tiering plus hefty fines, putting the squeeze on service providers; China's approach emphasizes whole-chain traceability and platform coordination, splitting the labeling obligation among service providers, distributors, and users. Both paths share the same premise: first make content carry source information, and only then can accountability and governance be discussed.

What It Feels Like for Creators

For ordinary creators, the change isn't technical but habitual. Before, you could post an AI-drawn illustration as an ordinary piece of work; now you have to declare it, and whether traffic gets suppressed by the algorithm after declaring is a question on many minds. There's no public data yet to answer it. The platforms' hint is that content which labels itself voluntarily won't be penalized extra — what actually gets handled is content that stays unlabeled while pretending to be real footage.

What's really worth watching is whether implicit labels can survive re-processing. When a watermarked video is cut into someone else's work, re-encoded a few times, or screenshotted and reposted, the watermark's survival rate drops. This is a weakness OpenAI itself acknowledges: its text watermark's detection rate falls from about 92% to 66% after 10% of words are replaced, and to just 17% after 25% are replaced. Labels raise the cost of faking, but they can't close that road off entirely.

The more practical question is where the line falls. If a real photo has had skin retouched by AI, its background swapped, and passersby removed, does it count as AI-generated? The measures point in a direction, but the specific judgment still depends on platforms' implementation rules. Draw the label too broadly and everyone labels everything, so the label loses its informational value; draw it too narrowly and a large amount of gray-area content slips through.

What the labeling measures change is the default. Before this, determining whether a piece of content was AI-generated was the platform's job, the audience's job — anything but the publisher's job. Now it's the publisher's job first. This step doesn't solve every problem, but it turns “AI-generated” from an option you could hide into a question you have to answer.