Which AI outputs require transparency under Article 50?
The answer depends on who provides or deploys the system and the use: direct interaction, synthetic content, emotion or biometric categorisation, deepfakes, and AI-generated public-interest text have distinct duties and exceptions.
Operating sequence
- 01
Classify the interaction and content types each AI feature can produce.
- 02
Place disclosure where people encounter the system or content, in an accessible and timely form.
- 03
Preserve machine-readable marking where required and document editorial or legal exceptions.
Failure modes to avoid
- Using one generic “powered by AI” notice for every duty.
- Removing provider markings during downstream publication.
- Assuming human review always removes the public-interest disclosure duty without checking the rule.
The source remains the authority.
European Commission — Article 50 transparency Q&A