Guide
Article 50 checklist
Article 50 of the EU AI Act splits into three separate obligations that land on different people depending on what they built or how they used it. Work out which one applies to you, then check it against what you actually need to disclose.
You build or provide a generative AI system
Article 50(2): mark the output and make it detectable
- Every synthetic audio, image, video, or text output your system produces carries a machine readable mark, not just a human readable label.
- The mark is technically effective. A detection tool can actually verify the output as AI generated, not just something a person could read if they knew to look.
- You have a documented view of how your marking and detection mechanism works, since the technical detail of what counts as adequate is largely being worked out through the Code of Practice on Transparency, not the bare text of the Act alone.
Exceptions
- Systems that only do assistive editing, like grammar or spelling correction, and don't substantially change what a person wrote.
- Systems whose main function isn't generation and that don't substantially alter the input data or its meaning.
- Systems used by law enforcement to detect, prevent, investigate, or prosecute a criminal offence.
In force since 2 August 2026 for new systems, with no transition period. Generative AI systems already on the market before that date have until 2 December 2026 to bring their marking and detection into conformity.
You use AI to create or manipulate a deepfake
Article 50(4), first part: disclose it
- If what you produced is image, audio, or video content that resembles a real person, object, place, entity, or event closely enough to look or sound authentic to someone who doesn't know otherwise, you disclose that it was artificially generated or manipulated.
- That disclosure reaches the person at the latest by the time they first see or hear it, in a way that's clear and distinguishable, not buried in a settings page or a footnote.
- The format fits the medium. A persistent visible label for video, an audible disclosure for audio, something that actually reaches the person experiencing the content rather than sitting only in metadata.
Exceptions
- Evidently artistic, creative, satirical, fictional, or similar work still needs disclosure, but in a lighter form that fits the work, not a disclaimer that breaks it.
- Use authorised by law for detecting, preventing, investigating, or prosecuting a criminal offence.
In force since 2 August 2026, no transition period.
You publish AI-generated text on a matter of public interest
Article 50(4), second part: disclose it, unless a person genuinely reviewed it
- What triggers this obligation is your purpose in publishing, informing the public on a matter of public interest, not just what the text happens to be about.
- Where it applies, you disclose that the text is artificially generated or manipulated, at the latest when a reader first sees it.
- If you're relying on the human review exemption, the review actually has to be substantive. Someone with real editorial responsibility read it and could have changed it. A glance or a rubber stamp doesn't qualify.
Exceptions
- Text that's gone through genuine human review or editorial control, where a named person or organisation holds editorial responsibility for it, doesn't need the disclosure.
In force since 2 August 2026, no transition period.
Code of Practice
Where the technical detail actually lives
Article 50 itself says outputs must be marked and detectable. It doesn't spell out exactly what counts as an adequate marking or detection mechanism. That detail is being worked out through the Code of Practice on Transparency of AI-Generated Content, a voluntary instrument drawn up by independent experts through a process the EU's AI Office ran with input from industry. Signing it isn't mandatory, but it's where the accepted technical approaches are actually being defined, and by the end of July 2026 around 190 companies and organisations had already signed on.
Non-compliance with Article 50 itself can expose a company to fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.
Where evidence comes in
A checklist tells you what applies. It doesn't check a file for you
Working out which obligation applies to you is a policy question. Confirming whether a specific asset actually carries the mark, credential, or watermark it's supposed to is an evidence question, which is what a Provenance Radar report is for. It shows what was found and how certain that finding is, as evidence relevant to your disclosure decision, never as a finding that a file complies with Article 50 or any other regime.
This is a general summary for orientation, not legal advice. Confirm current requirements for your situation with counsel before relying on it.