AI-generated content
AI-generated content labelling, marking and Article 50
Article 50 includes transparency obligations for certain AI-generated and manipulated content. Implementation can involve machine-readable marking by providers and additional visible or audible disclosure obligations for deployers in specific use cases.
Updated 7 October 2026 · Technical implementation guidance, not legal advice.
Separate provider marking from deployer disclosure
The Commission guidance distinguishes technical marking obligations for certain provider outputs from deployer-facing disclosure duties that can apply to deepfakes and other specified content. Product teams should map which role applies to each release path.
Test the actual generated output
Do not stop at a platform capability statement. Generate an example output and verify what metadata, provenance signal, watermark, content credential or visible label survives the route users actually take.
- Generate representative outputs from the production or release-candidate flow.
- Inspect downloaded/exported files and any public rendering path.
- Record which marking or provenance mechanism is present and where it can be detected.
- Capture known transformations that remove or alter the signal.
Deepfake and synthetic-content disclosure
Where a deployer disclosure obligation applies, the visible or audible disclosure should be reviewed in the context of the content experience—not only in general terms or policy text.
Create an evidence handoff
Retain representative output files, screenshots, metadata/provenance inspection results, the implementation ticket and acceptance criteria used to approve the release.
Common questions
Does Article 50 require all AI-generated content to carry the same label?
No. The obligations vary by role, content type and use case. Provider marking and deployer disclosure are distinct concepts, and exceptions can apply.
Are C2PA or watermarks the only possible implementation?
The Commission's framework is technology-neutral. The appropriate technical approach should be evaluated against the applicable obligation, state of the art and the actual output path.
Can TraceNotice test public generated-content outputs?
Yes. TraceNotice can inspect observable output paths and turn those observations into implementation findings, acceptance criteria and evidence-retention steps.