Guide · 2026-09-21T00:00:00.000Z · reviewed 2026-09-21T00:00:00.000Z
Article 50, writing an AI disclosure that holds up
What each paragraph of Article 50 requires, where the disclosure has to appear, the exceptions people misread, and the five mistakes we see most.
The five paragraphs, in plain terms
| Paragraph | Who | Duty |
|---|---|---|
| 50(1) | Providers of AI that interacts with people | Tell people they are dealing with AI, unless obvious to a reasonably well-informed person |
| 50(2) | Providers of generative AI | Mark generated audio, image, video and text as artificial, machine-readably, “effective, interoperable, robust and reliable as far as technically feasible” |
| 50(3) | Deployers of emotion recognition or biometric categorisation | Inform the people exposed; process biometric data under GDPR |
| 50(4) | Deployers | Disclose deepfakes; disclose AI-generated public-interest text unless a person reviewed it and holds editorial responsibility |
| 50(5) | Everyone above | Disclose clearly and distinguishably, accessibly, at the latest at the first interaction or exposure |
Applies since 2 August 2026. Not deferred by the Digital Omnibus, which only added paragraph 7 on codes of practice for detecting and labelling synthetic content.
Where the disclosure has to be
“At the latest at the first interaction” means the notice is visible before the person types, not in a footer they may never scroll to, and not in the second message. For a chat window that is the opening message plus a persistent label. For generated content it is a visible label on the content itself and intact machine-readable marks underneath. For emotion recognition it is a notice at the point of capture.
Accessibility applies: the notice must work with screen readers and on small screens.
The exceptions people misread
- “Obvious” (50(1)) is judged from the point of view of a reasonably well-informed, observant person. A bot named “Anna” with a photo is not obvious.
- Assistive editing (50(2)): spell-check, small retouching, and edits that do not substantially alter the content are outside the marking duty. A generated hero image is not assistive editing.
- Creative work (50(4)): for evidently artistic, satirical or fictional work the duty shrinks to disclosing that the content is artificial, in a way that does not hamper the work. It does not disappear.
- Editorial responsibility (50(4)): AI-written public-interest text is exempt only if a person reviewed it and takes editorial responsibility. “We use AI” in a site footer is not that.
Five mistakes we see most
- Disclosing after the first exchange.
- Stripping provider marks by exporting to a flat file before publishing.
- Treating a “powered by AI” logo as the disclosure.
- Forgetting the Korean and US variants when the same chatbot serves those markets.
- Not recording, per system, that the disclosure exists. The inventory column
disclosure_in_placeis the evidence.
Texts you can use
The disclosure generator produces the first-message notice, the persistent label, the answer to “am I talking to a person?”, content labels, deepfake notices, emotion-recognition notices and the Korea, Utah and Colorado variants. The Compliance Kit adds a 21-point checklist per system.