China AI Labelling Measures · GB 45438-2025 In force since 1 Sep 2025

An example of how to comply with China's AI-content labelling rules using JPEG Trust

China's Measures for Labelling AI-Generated Synthetic Content, together with the mandatory standard GB 45438-2025, require both visible labels and hidden machine-readable metadata. The rules are already in force and actively enforced. JPEG Trust can carry both a C2PA manifest and the prescribed TC260 metadata.

What the rules require

China's system is broader and more operationally detailed than the EU's or California's. It covers text, images, audio, video and virtual scenes, and distinguishes between explicit (visible) and implicit (metadata) labels — generally requiring both.

Explicit labels
Visible / audible

A visible or audible notice on the content — a mark on an image, a notice around a video player, an audible cue in audio — that should travel with downloads.

Implicit labels
File metadata

Machine-readable information in the file's metadata: whether it is AI-generated, the provider, a content ID, and later distribution information.

Platform duties
Distribution

Distributing platforms must inspect metadata, label or flag content, let users declare AI use, and add their own propagation information.

User duties
Declaration

Users publishing AI content must actively declare it. Removing, altering, forging or concealing labels — or providing tools to defeat them — is prohibited.

Enforcement date 1 September 2025 (already in force)
National standard GB 45438-2025 (mandatory)
Related rules Deep Synthesis Provisions (Jan 2023) · Interim GenAI Measures (Aug 2023)
Scope Text, image, audio, video and virtual scenes
Enforcement Active — the CAC took action against non-compliant apps in November 2025

Explicit vs implicit labels

Unlike California's offered visible label, China generally requires both an explicit label that a viewer can see or hear and an implicit machine-readable label in the file. A single C2PA manifest alone may not be enough.

Visible / audible

Explicit label

A disclosure perceptible to the person consuming the content, appropriate to the medium.

  • A mark placed appropriately on an image
  • A label in a video's opening frame and around the player
  • Should remain with downloaded, exported or copied content
Machine-readable

Implicit label

Structured metadata embedded in the file, following the TC260 AIGC schema.

  • Whether the content is AI-generated or synthetic
  • Provider name or registration code and a content ID
  • Later distribution platform and distribution ID

TC260 AIGC metadata structure

Technical guidance associated with GB 45438-2025 defines a specific XMP metadata structure for images. JPEG Trust can write and read these fields alongside its normal provenance.

Namespace http://www.tc260.org.cn/ns/AIGC/1.0/
XMP field Purpose
LabelWhether the content is AI-generated (the core flag)
ContentProducerThe content producer's name or registration code
ProduceIDA unique identifier for the produced content
ReservedCode1Reserved field for future use
ContentPropagatorThe platform that later distributed the content
PropagateIDA distribution identifier added by that platform
ReservedCode2Reserved field for future use

For JPEG images, the XMP packet is stored in the APP1 segment; PNG and TIFF use their corresponding metadata locations.

The propagation chain

China requires more than reading the creator's original manifest. The distribution platform becomes part of the provenance chain by adding its own propagation information.

Create

Content producer

An AI provider generates and signs the content, embedding the explicit and implicit labels.

Distribute

Distribution platform

A platform inspects the metadata, preserves the labels and records its own propagation event.

Verify

Downstream consumer

Anyone downstream can verify the labels, the producer and the propagation history.

How JPEG Trust supports it

A standard C2PA manifest alone may not be enough for a Chinese implementation. JPEG Trust can carry both representations in the same JPEG without one replacing the other.

C2PA signed provenance

Cryptographically signed assertions, action log and tamper-evident hash — JPEG Trust's normal representation.

+
TC260 AIGC XMP

The prescribed implicit-label fields in the TC260 namespace, stored in the JPEG APP1 segment.

TC260 implicit-label XMP

<!-- TC260 AIGC implicit label, stored in the JPEG APP1 (XMP) segment -->
<rdf:Description xmlns:aigc="http://www.tc260.org.cn/ns/AIGC/1.0/">
  <aigc:Label>1</aigc:Label>              # 1 = AI-generated
  <aigc:ContentProducer>Acme AI</aigc:ContentProducer>
  <aigc:ProduceID>urn:uuid:7f3a...</aigc:ProduceID>
  <aigc:ContentPropagator>PlatformA</aigc:ContentPropagator>
  <aigc:PropagateID>dist-20260902-001</aigc:PropagateID>
</rdf:Description>

CLI example

# Embed both a C2PA manifest and the TC260 AIGC implicit label
jpegtrust embed ai_output.jpg cn_metadata.json -o labelled.jpg \
    --profile cn-aigc \
    --aigc-label 1 \
    --aigc-producer "Acme AI" \
    --aigc-produce-id urn:uuid:7f3a...

# Inspect both representations
jpegtrust extract labelled.jpg

How JPEG Trust can verify labelled content

A platform or consumer can verify both the TC260 implicit label and the C2PA signature to confirm the content is labelled and unaltered.

China AIGC profile checks
✓AIGC labelThe TC260 Label field is present and indicates AI generation
✓Content producerContentProducer and ProduceID are populated
✓Content IDA unique produced-content identifier is present
✓Propagation infoDistribution platform and PropagateID are recorded where applicable
✓Signature chainThe C2PA signature chains to a trusted issuer
✓Pixel hashThe file has not been tampered with after signing

Limitations

  • JPEG Trust does not render the explicit on-screen or audible label or moderate platform decisions — those remain the provider's and platform's responsibility.
  • Providers must retain certain records for at least six months and handle user-declaration workflows separately.
  • GB 45438-2025 and its technical guidance evolve; field names and encodings should be checked against the current standard.
  • Detecting AI content that carries no metadata at all is outside JPEG Trust's provenance model.
Last legally reviewed: 30 June 2026 — verify against current standards before relying on this.
Labelling Measures (CAC) TC260 / GB 45438-2025

Get started

JPEG Trust is available as a Python package, a browser-based tool, and an HTTP API. Choose the integration that fits your workflow.

Python package

Install with pip. Embed, extract, and verify from any Python script or backend.

pip install jpegtrust

AI Content Transparency Around the World

See how China's labelling framework lines up against the EU AI Act and California.

How the three regimes compare →

Verification

Validate signatures and the TC260 implicit label from the package, CLI or API.

Verification guide →