China platform AI labels: confirmed, possible and suspected

China's labeling measures ask distribution platforms to sort AI content into three groups: confirmed, possible and suspected, and each gets a different on-screen label. Confirmed comes from an embedded implicit label, possible comes from a user report, and suspected comes from other evidence that the content is AI-generated.

This is from Inside Privacy's summary (read 2026-10-02) of the rules that took effect on 1 September 2025. TechNode (read 2026-10-02) shows the regulator enforcing labeling against apps in April 2026. I could not find a platform's own help page on how it applies the three groups, so I describe the rule, not any one app's behavior.

What are the three groups?

Per the summary, a platform that distributes content treats it as follows.

What does that mean for your uploads?

The practical split is between files that announce themselves and files that do not. A clip with an intact implicit label is confirmed with no argument. A clip stripped of metadata falls to reports and the platform's own detection, which means a viewer's flag can put a 'possibly AI-generated' note on it.

A visible label you burn in yourself does not change those categories, but it removes the guesswork for the viewer. Video captions burns authored cues into a finished clip at the times you set, and the job id from Jobs and results ties the file to its generation record.

What if the label is wrong?

I found no appeal procedure in the sources I read, so I cannot describe one. What is documented is that the label mechanism rests on the metadata, user reports and other evidence. If a real, non-AI clip is flagged, the useful evidence is the original camera file and edit history, which Sume does not hold for you.

Equally, if your clip is AI-generated and you want the label to be correct, add the implicit metadata and the visible label rather than hoping detection stays soft. Sume's docs describe what each tool returns and say nothing about a watermark or embedded provenance record on outputs, so treat the output as unmarked until you have checked the delivered file yourself.

Why does the three-way split exist?

Platforms cannot know the origin of every upload, so the rule gives them a ladder of confidence. Metadata is the strongest signal and earns a flat AI-generated indicator. A report from another user is weaker and earns a softer reminder. Other evidence, such as a platform's own detector, earns a 'suspected' label.

For a publisher this means the label a viewer sees is partly out of your hands. The one input you control is the file: an intact implicit label moves the clip to the top rung without any dispute, while a stripped file leaves the label to the crowd and the detector.

Enforcement also shapes platform behavior. The April 2026 action against apps shows the regulator holds the service responsible for labeling, so platforms have a reason to label generously rather than miss a clip. Expect false positives to be tolerated more than false negatives, though the sources I read do not say so explicitly.

Checklist for a China-facing clip

Use this as a working list and adjust it to your own pipeline and counsel's advice.

• Burn a short visible AI-generated notice into the pixels.
• Add provider name and content ID metadata in your own pipeline and test it survives your last export.
• Expect a platform to relabel a clip it detects or that users report.
• Keep the original generation record: job id, delivered file, date.
• Check the live platform help page, since I did not find one to cite.