AI Content Transparency EU & CA: Aug 2026 · China: in force

AI Content Transparency Around the World

Governments are introducing requirements for identifying AI-generated and synthetically altered content. The legal details differ, but the direction is consistent: content should carry reliable machine-readable provenance, people should receive understandable disclosures, and platforms should preserve and communicate authenticity information.

Overview

Three major regimes now set the pace for AI-content transparency: the EU AI Act, California's AI Transparency Act, and China's AI-generated synthetic content labelling framework. They share a common core — signed, machine-readable provenance plus human-facing disclosure — but diverge on visible labels, prescribed metadata fields, detection tooling and platform duties. JPEG Trust provides technical building blocks that map onto all three.

JPEG Trust provides provenance, verification and disclosure capabilities designed to support compliance with emerging AI-content transparency requirements. Use of the technology does not by itself guarantee legal compliance. This mapping should receive legal review before being relied upon as definitive advice.

Jurisdictions

🇪🇺

European Union

Applies 2 Aug 2026
Key lawEU AI Act, Articles 50 & 53
Main focusMachine-readable marking and detectability of synthetic media, visible disclosure for deepfakes and certain public-interest content, and GPAI documentation.
JPEG Trust roleSigned provenance, detection, interoperability and user-facing verification reports.
Art. 50 — Transparency → Art. 53 — GPAI Models →
🇺🇸

California

Applies 2 Aug 2026
Key lawCalifornia AI Transparency Act (SB 942 / AB 853) and AB 2013
Main focusLatent disclosures, a free public detection tool with an API, optional visible labels, platform preservation (2027) and capture-device provenance (2028).
JPEG Trust roleVerification API, signed provenance, platform badges and capture history.
California AI Transparency Act →
🇨🇳

China

In force & enforced
Key lawAI-Generated Synthetic Content Labelling Measures & GB 45438-2025
Main focusBoth explicit (visible) and implicit (metadata) labels, prescribed TC260 fields, user declarations and platform propagation information.
JPEG Trust roleTC260 metadata support, visible labels, signatures and propagation-chain verification.
China AI Labelling Measures →

How the three regimes compare

A high-level comparison of the obligations across the EU, California and China. Exact wording and exceptions vary; treat this as orientation, not legal advice.

Requirement 🇪🇺 EU Art. 50 🇺🇸 California 🇨🇳 China
Machine-readable label Required Required (covered providers) Required
Visible label In deployment cases Offered as an option Generally required
Text covered Yes Image/video/audio focus Yes
Prescribed metadata fields Not fixed in Art. 50 Provider, model, time, ID Detailed TC260 fields
Public detection tool Not structured this way Required, incl. API Platform detection
User declaration workflow Not a core requirement Not a core requirement Required
Platform preservation Via interoperability From 2027 Explicit duties
Capture-device provenance Not central From 2028 Can form part
Training-data transparency Article 53 (separate) AB 2013 (separate) Other provider rules
Current application date 2 Aug 2026 2 Aug 2026 1 Sep 2025

Required Partial / conditional Not a core requirement

How JPEG Trust fits

Assuming JPEG Trust verifies signed provenance metadata and presents it through a browser extension or API, here is where it lands against these regimes.

Strong fit
  • Read machine-readable AI provenance
  • Verify digital signatures
  • Show originating provider or capture device
  • Show model, version and edit history
  • Detect broken or invalid provenance
  • Provide a browser badge and detailed report
Partial fit
  • Public detection tool and API
  • Visible disclosure labels
  • Robustness to cropping and re-compression
  • Platform propagation records
  • China's exact TC260 XMP structure
  • User declaration workflows
Outside the core
  • Training-dataset transparency (AB 2013 / Art. 53)
  • Detecting AI content that never carried provenance
  • Platform moderation decisions
  • Six-month activity logs
  • Regulatory filings and security assessments
  • Legal determination of "clear and conspicuous"

Get started

Explore each jurisdiction in detail, or install the package to embed and verify provenance yourself.

European Union

Machine-readable marking and detectability of synthetic media, visible disclosure for deepfakes and certain public-interest content, and GPAI documentation.

Art. 50 — Transparency →

California

Latent disclosures, a free public detection tool with an API, optional visible labels, platform preservation (2027) and capture-device provenance (2028).

California AI Transparency Act →

China

Both explicit (visible) and implicit (metadata) labels, prescribed TC260 fields, user declarations and platform propagation information.

China AI Labelling Measures →