The picture looked fine. No corner stamp, no footer, no tell. What was missing sat inside the file, in the fields almost nobody opens, and it was enough to take the whole service down.
China's labelling rules have been in force since 1 September 2025. The argument abroad has been about the visible half: the corner tag, the caption saying a machine made this. The enforcement record says the visible half is not where cases are being won.
Key takeaways
- A Sichuan company's mini-program went offline in September after missing all three label points, including the record buried in the file.
- Platforms removed more than 241,000 short videos for labelling failures in about two months, and handled 21,000 accounts.
- The same period saw 1.275 million videos corrected or re-tagged, which is a compliance workflow, not a policing one.
- Since May, platforms have been required to make that choice a gate before a post can go out.
- A label answers one question only. In September the regulator told platforms that labels cannot be used to dodge content rules.
Two labels, one of them invisible
The 2025 measures (人工智能生成合成内容标识办法) require two things at once. A visible marker where a person will see it, and a hidden record in the file's metadata: what the content is, who provided the service, and a content number.
The hidden half is the interesting one, because it survives the obvious workarounds. A user can crop a corner logo. A repost can strip a caption. Metadata travels with the file until something removes it, and cross-platform recognition is what the rules are actually built on.
The label is less a warning sticker than a serial number.
The dropdown at the publish button
In May, the Cyberspace Administration of China (国家网信办) pushed platforms to move labelling from an afterthought to a gate. Publishers now choose from a fixed list before a short video can go out: dramatised fiction, machine-made content, marketing material, reposted material, personal opinion, or no label needed.
The same push came with a tally: more than 98,000 accounts handled for failing to mark sources or content types. Some of the cases were political and foreign-affairs clips with no source. Others were animal videos made with generative tools and posted as if they were real: a golden retriever holding a baby, a tiger and a ginger cat driving home for the holidays.
The enforcement target was not the model. It was the uploader who left a dropdown alone.
What two months of enforcement looked like
On 10 July, Xinhua reported the first accounting of the labelling campaign. In roughly two months, major platforms had:
- removed more than 241,000 short videos that broke the labelling rules;
- handled more than 21,000 accounts, including closures and posting bans;
- corrected or re-tagged more than 1,275,000 videos.
That third number is the one to sit with. For every video removed, roughly five were quietly fixed. This is not a takedown regime so much as an enormous back-office correction queue, and implies file-level scanning at a scale no human review team could reach.
Smaller platforms that had not implemented the rule were summoned by local regulators. Accounts running "disabled wife pregnant again" and "single father delivering food with his child" storylines were closed or silenced. None of those clips needed a generative model to be harmful. What they lacked was a fiction label.
The case that matters
On 15 September, the national regulator published a batch of enforcement cases. Case nine is the one to read.
A Sichuan technology company ran a WeChat mini-program offering text chat and image generation. According to the notice, it added no visible label, embedded no metadata record with the provider name, code or content number, and added no label when a user exported a file. It had also skipped the required security assessment. The service was ordered offline.
Case ten is its twin: a Jiangsu company relaying other firms' model APIs through two sites, with no security assessment. It got a warning, and so did its responsible individuals.
Both are ordinary compliance failures, resolved with the equivalent of a switch being thrown. Mature enforcement looks like this: boring, fast, and expensive for the company on the receiving end.
A label is not a licence
A second thread runs through the period. In April the regulator opened a four-month campaign against misuse of model tools, with labelling failures named as one of seven target problems, alongside unregistered models, poisoned training data and services that strip labels for a fee.
By 2 September the campaign's second phase had cleared more than 5.61 million items, handled 49,000 accounts and dealt with 2,400 sites and apps. Platforms published 46 separate governance notices. One of the seven published case types was a digital human used to sell products to elderly viewers with false claims.
On 30 September the regulator went further, warning that a fiction or machine-made tag cannot be used to escape content rules. The lesson: the label answers "was this made by a machine". It does not answer "is this true".
Honest limitations
This article rests on regulator notices and state media reports, which describe actions but not base rates; there is no public figure for how many labelled videos exist, so the removal numbers cannot be turned into a failure rate. The identity of the companies in the enforcement cases is anonymised, so I cannot check their side of the story. I also have no visibility into how well cross-platform metadata recognition actually works in practice, which is the weakest link in the whole design. Nothing here assesses whether the regime improves public understanding; that needs audience research nobody has published.
What readers can do now
- If you ship a generator into China, test the export path. The Sichuan case failed on three separate label points, and the export button was one of them.
- Keep the provider name, service code and content number in your metadata schema from day one. Retrofitting an identifier across a media library is far more expensive than writing it at generation time.
- Do not treat a label as a defence for a false claim. Build the claim-substantiation review separately, because the regulator does.