AI-Generated Content Disclosure: EU Article 50 Duties

Article 50 splits obligations between the entity that builds a system, the provider, and the entity that puts it in front of end users, the deployer. Those are frequently different companies.

• Article 50(1): deployers of directly interactive AI systems, chatbots and similar, must inform natural persons they are interacting with an AI system, unless that is already obvious
• Article 50(2): providers must mark synthetic output in a machine-readable format, detectable as artificially generated
• Article 50(4): deployers must label deepfakes, meaning realistic AI-generated or manipulated image, audio, or video content resembling existing people, objects, places, or events
• Article 50(5): deployers publishing AI-generated or manipulated text on matters of public interest must disclose that it was artificially generated, unless a human exercised editorial responsibility and control over the publication

Only one of those four is the provider's. The other three land on whoever deploys the system.

A model provider making its output machine-detectable does not satisfy a deployer's separate duty to tell the human on the other end. Both obligations can apply to the same piece of content, held by different parties, which is exactly the split that trips people up when they assume a watermarked file has handled transparency on its own.

Cooley flags three terms that remain undefined even after the effective date:

• The scope of the "already obvious" exception to interactive-AI disclosure
• What counts as a "public interest" matter triggering the deployer text-disclosure duty
• The criteria for "substantive human editorial review" that can exempt text from disclosure

That third one matters most for publishers and content platforms. The consistent view across firms is that a token human glance at AI-generated text is not enough to claim the editorial exemption, and no primary EU source has drawn a bright line for what is.

Jones Day adds a more forgiving carve-out worth knowing about. The Code explicitly allows lighter or adapted labelling for artistic, fictional, or satirical works, recognizing that a disclosure regime built for disinformation should not land identically on a parody video.

Say you make a realistic video of a person who does not exist, with no intention of deceiving anyone. Greenberg Traurig reads Article 50(4) as indifferent to both intent and to whether the person on screen is real, so a label is still owed, and this carve-out changes only how light it can be.

Treat any confident claim about where either line sits, the editorial-review threshold or the creative-work adaptation, as informed legal commentary rather than settled law. No Commission guidance or enforcement decision has narrowed them yet.

Take the most common real-world case. A company puts an AI chatbot on its support or sales page.

Article 50(1) says users need to be told they are talking to an AI, unless that is "already obvious." The Commission has not precisely defined that phrase.

Veeam field CTO Edwin Weijdema, writing in the days after the rule took effect, raised a question that sounds basic and is not answered anywhere. Does an AI agent embedded in a ticketing or procurement portal even count as "directly interactive" the way Article 50(1) means it, or does its more transactional, less conversational framing put it outside the duty entirely?

Nobody tracking the rule claims to know for certain. And no confirmed enforcement action under Article 50 had been publicly reported as of 11 August 2026, nine days after it took effect.

So there is no test case yet showing how strictly regulators will read "already obvious," "directly interactive," or anything else on the undefined list. Companies rolling out customer-facing AI right now are making judgment calls ahead of the case law that will eventually settle them.