Skip to content
Coverage checker

What labels does China require on AI-generated content?

Two, since 1 September 2025, if your service is offered to users in mainland China: a visible label a person can see or hear, and an implicit label in the file's metadata, the AIGC field fixed by GB 45438-2025. A C2PA manifest does not replace that field. Your user agreement must explain the labels, and removing labels, or supplying tools to, is prohibited.

Evaluated on 18 October 2026 from SignUnder’s rule packs (ca-ai-transparency-act@2026.10.12, cn-ai-labelling-measures@2026.10.13, eu-ai-act-art50@2026.10.12, eu-code-of-practice@2026.10.12, in-it-rules-synthetic@2026.10.13, kr-ai-basic-act@2026.10.13, ny-gbl-396b@2026.10.12, platform-labels@2026.10.12). Your situation may differ: open these answers in the checker and change them.

What you need

  • China’s AIGC metadata field (GB 45438-2025)
  • A visible label on the output itself
  • Your user agreement explaining the labels
  • No way to remove labels or metadata

The law

These bind you.

The AIGC metadata field in every file

In force since 1 September 2025

Write an implicit label into the file's metadata: a field named AIGC with seven keys (the label value, your name or code as producer, your content number, and the reserved and distributor keys), as Annex E of GB 45438-2025 fixes. A C2PA manifest does not replace it, and it must be written before any manifest is signed. Watermarks are encouraged, not required.

“production element information such as the attribute of the generated content, the service provider's name or code, and a content number”

Not confirmed: Where the field goes in each file format is in TC260 practice guides we have not read. That the Measures reach only services offered to the public in mainland China is from memory.

A visible label on generated content

In force since 1 September 2025

If your generation service is offered to users in mainland China, generated text, images, audio and video carry an explicit label a person can see or hear, containing "AI" (or 人工智能) and "generated" (生成) or "synthesised" (合成). On images and video it is text at an edge or corner at least 5% of the shortest side high; audio gets a spoken notice or the Morse rhythm for "AI". Exported files keep the label unless the user asks otherwise under your agreement, and then you keep logs for six months.

“ensure the file contains an explicit label that meets the requirements”

Not confirmed: Which services must add the visible label is set by Article 17(1) of the deep synthesis provisions, which we have not re-read; the standard's Annex B lists text-to-image, music and video, cloned voices and face generation as typical. That the Measures reach only services offered to the public in mainland China is from memory.

SignUnder does the marking for you

Every image you send through SignUnder comes back with signed Content Credentials, an invisible watermark and a fingerprint on record, with a public page where anyone can check it is yours. Early access is free.

Get an API key

SignUnder is not a law firm and this is not legal advice. Every rule cites its primary source and the date we read it, so your own counsel can check our reading. The rule packs are published as JSON at /rules/index.json.