An example of how to comply with EU AI Act Article 50 using JPEG Trust

What Article 50 requires

Article 50 of the EU AI Act creates transparency obligations for anyone who builds or deploys AI systems that generate or manipulate content. It entered the final text of Regulation (EU) 2024/1689 and takes effect on 2 August 2026.

Why AI-generated media needs labelling

Generative AI can now produce photorealistic images, convincing audio, and fluid video that are indistinguishable from human-created content. Without labelling, there is no reliable way for viewers, journalists, fact-checkers, or automated systems to know what they are looking at.

Article 50 addresses this by requiring that the provenance of AI-generated content travel with the content itself — not as a disclaimer on a webpage, but as metadata embedded in the file. This makes the label machine-readable, persistent, and verifiable downstream regardless of where the content ends up.

The EU's recommended approach

The European Commission's draft Code of Practice for AI-generated content (published December 2025, final version expected mid-2026) specifies a multi-layered approach. No single technique is sufficient on its own:

JPEG Trust implements the first and most critical layer: cryptographically signed provenance metadata using the JPEG Trust Trust Record. This is the layer that makes AI labelling machine-readable, interoperable, and legally auditable.

How JPEG Trust embeds compliance metadata

When an AI system generates or modifies an image, JPEG Trust embeds the required provenance assertions directly into the JPEG file. These assertions declare the AI involvement using standardised vocabularies that JPEG Trust tools can read, and that any C2PA-compatible tool can also read.

Digital source types for AI content

Python example: embed AI provenance

from jpegtrust import embed_image, create_profile, save_profile

# 1. Create an AI-labelling profile
profile = create_profile(template="ai-labelling", name="Article 50 Compliance")
save_profile(profile, "article50.yml")

# 2. Embed provenance into an AI-generated image
embed_image(
    "ai_generated_photo.jpg",
    "ai_metadata.json",
    "labelled_output.jpg",
    digital_source_type="trainedAlgorithmicMedia"
)

# The output file now contains:
#   - digitalSourceType "trainedAlgorithmicMedia" on the c2pa.created action
#   - the assertions from ai_metadata.json, e.g. c2pa.actions.v2 (generation
#     action log) and cawg.training-mining (AI training/inference permissions)

CLI example

# Embed AI provenance metadata
jpegtrust embed ai_photo.jpg ai_metadata.json -o labelled.jpg \
    --digital-source-type trainedAlgorithmicMedia

# Inspect what was embedded
jpegtrust extract labelled.jpg

How JPEG Trust can verify compliance

Downstream consumers — platforms, newsrooms, regulators, or automated pipelines — can verify that an image carries valid Article 50 metadata using the ai-labelling trust profile. The profile defines exactly which assertions must be present and valid.

Python example: verify compliance

from jpegtrust import verify_image

# Verify an image against the AI-labelling profile
report = verify_image("labelled_output.jpg", "article50.yml")

print(f"Verdict: {report.status}")           # 'pass'
print(f"AI label: {report.ai_label}")         # 'trainedAlgorithmicMedia'
print(f"Checks passed: {report.passed_checks}/{report.total_checks}")

if "ai_declaration" in report.failed:
    print("WARNING: No AI training/inference permission declared")

if "digital_source_type" in report.failed:
    print("WARNING: Missing IPTC digital source type")

CLI example

# Verify against the ai-labelling profile
jpegtrust verify labelled.jpg article50.yml

# Batch-verify a folder of AI-generated content
jpegtrust batch verify ./ai_outputs article50.yml

Example end-to-end compliance pipeline

Here is a typical workflow for an organisation that generates AI content and needs to comply with Article 50 — from generation through to downstream verification.

output = model.generate(prompt="sunset over lake")
output.save("raw_output.jpg")

jpegtrust embed raw_output.jpg \
  ai_meta.json -o labelled.jpg \
  --digital-source-type \
  trainedAlgorithmicMedia

# The labelled JPEG can be:
# - Published on a website
# - Shared via messaging
# - Uploaded to a platform
# - Sent to a client

jpegtrust verify labelled.jpg \
  article50.yml
# verdict: pass
# ai_label: trainedAlgorithmicMedia
# checks: 7/7 passed

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.