China: labelling AI-generated content, one year of enforcement

On 28 April 2026, China’s cyberspace regulator (China · CAC) had three generative AI services summoned, ordered to comply and warned: the Jianying editing app, the Maoxiang app and the Jimeng AI website. The reason: inadequate implementation of the rules on labelling AI-generated content, in force since 1 September 2025. Two days later, the CAC launched a four-month campaign against misconduct by AI applications. Its first phase, across all grounds, dealt with more than 14,000 AI websites, apps and agents.

One year on, the Chinese rule is clear and its enforcement is real. Every link answers for the label: the generator, the platform, the app store and the user. Since August 2026, Europe has applied a similar obligation, over a narrower scope.

What is already settled

• The labelling measures (China · CAC, with the ministries of Industry and Public Security and the radio and television administration): published on 14 March 2025, in force since 1 September 2025.
• A hidden label for all generated content (China · CAC): file metadata must state the nature of the content, the provider’s name or code and a content ID. Digital watermarks are encouraged.
• A visible label when the public could be misled (China · CAC): text or dialogue imitating a person, synthetic or imitated voices, generated or swapped faces, realistic immersive scenes, and any other service that generates or significantly alters content. It must remain after download, copying or export.
• Three tiers for distribution platforms (China · CAC): content “confirmed” by metadata, “possible” if declared by the user, “suspected” if traces are detected. Each case is flagged to the public.
• Duties for app stores and users (China · CAC): check the labelling of generative apps before listing them; declare one’s generated content; never remove, falsify or conceal a label, nor supply tools to do so.
• The mandatory national standard GB 45438-2025 (China · SAMR and SAC, the standardisation authorities): published on 28 February 2025 and applied since 1 September 2025, it sets the labelling methods. The April 2026 campaign targets missing labels, but also labels whose size, position or transparency do not comply.
• Short videos (China · CAC, 12 May 2026): before posting, creators must pick one label out of six options, including “contains AI-generated content”. Twelve platforms had been piloting the feature since March 2026.
• A takedown (China · CAC, 15 September 2026): a WeChat mini-program offering AI chat and image generation was taken offline for lacking both visible labels and metadata.

What is still in progress

• The campaign’s second phase (China · CAC), focused on content: false information, deepfake impersonation, content targeting minors, automated fake accounts. As far as we know, no final review had been published as of 29 October 2026.
• Cross-platform recognition of hidden labels (China · CAC), which the CAC lists among the failings to fix, along with weak detection tools.
• Fines (China · CAC): as far as we know, no fine for failing to label had been publicly quantified as of 29 October 2026. Published measures are summons, warnings, compliance orders and takedowns.
• The European deadline (EU · Digital Omnibus on AI): generative tools placed on the market before 2 August 2026 have until 2 December 2026 to mark their outputs.

The real issue: Europe marks, China displays

Article 50 of the AI Act (EU · AI Act) has applied since 2 August 2026. Providers of generative systems must mark their outputs in a machine-readable, detectable format. Apart from the duty to tell people they are interacting with an AI, visible disclosure of content falls on deployers, for deepfakes and for text published to inform the public on matters of public interest, unless it has undergone human review under editorial responsibility. Fines can reach EUR 15 million or 3% of worldwide turnover.

China goes further on two points. It requires a visible label whenever a service could mislead the public. And it makes platforms detect and flag content they did not produce. In China, the platform also answers for content it did not generate. For a group active in both markets, the technical foundation is shared: robust metadata. The difference lies in display and in the chain of responsibility.

Three points to watch

1. Metadata gets lost on the way. Compression, screenshots, re-uploads: the hidden mark often disappears. The CAC faults platforms for weak detection, and the rules require the visible label to survive export. The whole chain needs testing, not just the generator.

2. “De-labelling” is targeted. The measures prohibit maliciously removing or concealing a label, and supplying tools to do so. The 2026 campaign also targets tutorials. Even without intent, an internal process that strips metadata destroys the evidence that you label.

3. The advertiser is also a user. The measures require users to declare their generated content and to use platforms’ labelling tools. A brand, its agency and the creators it pays are all concerned, especially on short videos.

What an advertiser or publisher active in China should do

1. List the generative tools producing content published in China: marketing, customer service, apps.
2. Require both labels from suppliers, visible and hidden, backed by metadata samples.
3. Preserve metadata throughout the production chain: export, compression, asset library.
4. Declare generated content on each Chinese platform with its own tools, including the dedicated short-video label.
5. Write labelling into contracts with agencies and creators, with a right to check.
6. Keep logs for at least six months if, as a provider, you deliver content without a visible label at a user’s request.
7. Build a single China–EU policy: machine-readable metadata everywhere, visible labels according to each market’s rules, ready before 2 December 2026.

A label that does not survive export is not a label.

The question to ask on Monday morning

To your head of marketing for China: “For every piece of AI-generated content we publish in China, can you show the visible label and the metadata, after export?”