AuditImage
Guides

What the AIGC implicit label looks like

The metadata label required by the AI-generated content labelling rules, its fields, where it lives in a file, and how platforms actually write it.

Regulatory background

The Measures for Labelling AI-Generated Synthetic Content were issued jointly by four regulators (the cyberspace, industry, public security and broadcasting authorities), in force since 1 September 2025. The mandatory national standard GB 45438-2025 (Cybersecurity technology: Labelling method for AI-generated synthetic content) took effect the same day, followed by practice guides from TC260, including one on implicit labels in file metadata.

The measures require two kinds of label:

  • Explicit: visible text, badges or audio prompts.
  • Implicit: written into file metadata, invisible to people, readable by machines.

The implicit label

The standard specifies a JSON object in file metadata with fixed field names:

{
  "Label": "AIGC",
  "ContentProducer": "name or code of the generation service provider",
  "ProduceID": "content production ID",
  "ReservedCode1": "reserved",
  "ContentPropagator": "name or code of the distribution platform",
  "PropagateID": "content distribution ID",
  "ReservedCode2": "reserved"
}

Label is always AIGC and is the anchor for detection. The generation provider fills ContentProducer and ProduceID; when a platform redistributes the content it adds ContentPropagator and PropagateID.

Where it lives

The standard specifies a location per file type:

  • Images: an extension field in the metadata. In practice PNG uses a tEXt or iTXt chunk keyed AIGC; JPEG mostly puts it in XMP, sometimes in the COM comment segment.
  • Video and audio: the container's metadata area.
  • Text files: the file header.

Because implementations differ, AuditImage does not fix a location. It searches PNG text chunks, the XMP packet, EXIF description fields, JPEG comment segments and unknown APP segments for the "Label":"AIGC" pattern and parses the full JSON when found.

What platforms actually do

At the time of writing, the major generators covered here (Tongyi Wanxiang, Jimeng, Kling, Yige, Hunyuan) write the implicit label into exports, usually with a company name or registration code. Douyin, WeChat and Weibo detect the label on upload and add the propagator fields. Formats are still settling and the same platform can produce slightly different output over time.

Relationship to C2PA

The goals are similar, the mechanisms different. The AIGC implicit label is plain JSON with no signature; C2PA has a signature and a content hash. The label's credibility comes from the legal obligation to write it, not from cryptography. AuditImage shows them in separate sections: "AIGC implicit label" and "C2PA Content Credentials".

Manual check

exiftool -a -G1 -s image.png | grep -i aigc
exiftool -XMP:all image.jpg | grep AIGC