EU AI Act Article 50: The Complete Guide to Transparency Obligations

Article 50 of the EU AI Act is the Act's transparency chapter — the part that doesn't regulate how risky an AI system is, but whether people are told when they're dealing with one. It sits alongside the Act's risk-tier rules (prohibited, high-risk, limited-risk systems) as a separate, standalone set of disclosure duties that apply regardless of risk category, wherever AI-generated or AI-manipulated content reaches a real person.

It has applied since August 2, 2026, and penalties for non-compliance can reach €15 million or 3% of global annual turnover, whichever is higher — the same upper tier as the Act's prohibited-practices penalties, which tells you how seriously the Commission treats transparency.

Article 50 doesn't create one rule — it creates four distinct duties, on different parties, triggered by different content types. Mixing them up is the single most common source of confusion:

• Chatbot/interaction disclosure (Art. 50(1)) — providers of AI systems that interact directly with people (chatbots, voice assistants, support bots) must make clear the person is talking to AI, at first interaction, unless it's already obvious.
• Provider content marking (Art. 50(2)) — providers of AI systems that generate synthetic audio, image, video, or text must mark outputs in a machine-readable format detectable as artificially generated, where technically feasible.
• Deployer deepfake disclosure (Art. 50(4), first paragraph) — deployers of AI systems that produce a deepfake (see our deepfake disclosure guide) must disclose that the content is artificially generated or manipulated, clearly and perceivably, to anyone exposed to it.
• Deployer public-interest text disclosure (Art. 50(4), second paragraph) — deployers publishing AI-generated text on matters of public interest must disclose that too, unless a named human editor genuinely reviewed it first.

See our provider vs. deployer guide if you're not sure which role applies to you — it changes which of these duties, if any, is yours.

A provider's machine-readable mark (Art. 50(2)) and a deployer's disclosure duty (Art. 50(4)) are separate obligations on separate parties. If you use a third-party AI image tool that embeds a valid C2PA manifest, that satisfies the tool provider's obligation — it does not, by itself, satisfy your obligation as the deployer to clearly disclose the content to the people who see it. You need your own visible, perceivable disclosure, independent of whatever metadata is (or isn't) embedded in the file.

Satirical and fictional content is not exempt from disclosure — it gets an adapted form (unobtrusive labeling, credits, a description) that doesn't interrupt the work, rather than a waiver. The only real exemption in Article 50(4) is narrow: it applies exclusively to text, only on matters of public interest, and only where a named human editor genuinely reviewed the content first. It never extends to images, audio, or video.

If you're a deployer publishing or distributing AI-generated images, the practical question is almost always: does this specific image meet the Article 3(60) deepfake test, and if so, do you already have a clear, perceivable disclosure in place? Our free checker walks through exactly that, image by image.