Article 50 is the part of the EU AI Act that applies to almost everyone — including businesses whose AI is nowhere near high-risk. Four disclosure duties, who owes each one, and the exact moment each one is due.
Most EU AI Act guides start with risk categories — prohibited, high-risk, limited, minimal. Transparency cuts across them. Article 50 lists specific situations where someone must be told an AI is involved, and those situations are defined by what the system does in public, not by its category. That is why a friendly support chatbot with no high-risk features still owes its visitors one clear sentence, and why a marketing team publishing AI video owes a label even though the tool behind it is unremarkable.
Work out your category first if you have not yet — our risk classification guide walks through the 7 questions — then use this page for the disclosure layer that applies either way.
Providers build or supply the AI system and put it on the market under their own name. Deployers use a system in their business, under their own authority — the shop that installs a chatbot, the publisher that posts AI video. Article 50 splits its duties along exactly this line: duties 1 and 2 below fall on providers, duties 3, 4 and 5 on deployers. Many small businesses are deployers of several systems at once, and a business that substantially modifies or rebrands a system can step into provider shoes for it. Buying a tool never transfers your deployer duties to the vendor.
If your AI system is meant to interact directly with people — a support chatbot, a voice agent, an AI receptionist — the people using it must be informed that they are interacting with an AI system. The only escape is when it is already obvious from the context, and “obvious” is judged from the person’s side, not the vendor’s.
The notice has to arrive at the latest at the first interaction, and it has to be clear and distinguishable — not buried in terms of service nobody opens.
If your system generates synthetic audio, images, video or text, its outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable and robust as far as technically feasible.
This is a duty about the output itself — metadata, watermarking or an equivalent signal — and it sits with the provider of the generative system. If you deploy someone else’s generator, ask the vendor, in writing, how their marking works; “we handle compliance” is not an answer you can show anyone.
If you deploy an AI system that generates or manipulates image, audio or video content that resembles real people, objects, places or events — a deepfake — you must disclose that the content was artificially generated or manipulated.
For works that are evidently artistic, creative or fictional, the duty narrows: disclose the existence of the generated or manipulated content in a way that does not hamper displaying or enjoying the work. A satire label or a credit line can be enough; silently passing synthetic media off as footage is exactly what the rule targets.
If you deploy an emotion-recognition system or a system that categorises people from biometric data, the people exposed to it must be informed that it is operating. Their personal data also has to be processed in line with GDPR and other applicable EU data law.
Separately, remember the boundary from the classification guide: emotion inference in workplaces and education is on the prohibited list, with narrow exceptions. A notice does not make a banned use legal.
If you deploy AI to generate or manipulate text that is published to inform the public on matters of public interest, the text must disclose that it is AI-generated — unless one of two things is true: the use is authorised by law, or the content has passed human review or editorial control and a named person or business holds editorial responsibility for it.
In practice: an AI-drafted news-style post that nobody reviewed needs the disclosure; the same draft reviewed, edited and published under an editor’s responsibility falls under the exception. Keep a record of which route each piece took.
1. Hiding the notice in the terms. Article 50 requires the information to be clear and distinguishable at the time of first interaction or exposure. A clause on page nine of a legal document is neither.
2. Disclosing to readers but skipping the machine-readable mark. The visible label and the embedded signal are two different duties owed by two different roles; doing one does not cancel the other.
3. Assuming the vendor handled it. Deployer duties — deepfake labels, emotion-recognition notices, public-interest text — stay with the business that publishes or operates the system.
4. Labelling once, at launch, and never again. Every new chatbot surface, voice agent or content format restarts the question. A disclosure that exists only on the homepage misses the chat window where the interaction actually happens.
5. Treating “artistic” as “exempt.” Evidently creative or fictional works still require disclosure of the generated or manipulated content — just in a lighter form that does not hamper the work.
The EU AI Act Classification Report maps your specific system against the Act: risk category with reasoning, the obligations that follow — including its Article 50 transparency duties — your deadlines and a step-by-step plan in plain English. $49 one-time, delivered by email. Optional Act Watch is $19/month and re-checks your classification as the rules move. Not legal advice.
Article 50 of the EU AI Act sets disclosure duties that apply regardless of risk category: people must be told when they are interacting with an AI system, AI-generated or manipulated content must be marked and disclosed, deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them, and deepfakes must be labelled as artificially generated or manipulated.
Both, for different duties. Providers (the businesses that build or supply the system) owe the interaction disclosure and the machine-readable marking of generated content. Deployers (the businesses that use the system with their own customers or staff) owe the deepfake, emotion-recognition and public-interest text disclosures. If you buy an AI tool and put it in front of your customers, you keep deployer duties even though the vendor built the system.
The Article 50 transparency duties have applied since 2 August 2026. A grace period for the machine-readable marking of content from generative systems already on the market ends on 2 December 2026 (information as of November 2026).
Yes — one narrow thing. If visitors interact with your chatbot, they must be informed they are interacting with an AI system, in a clear and distinguishable way, at the latest at the first interaction, unless it is obvious from the context. A short line in the chat window itself is the usual way to do it.
It means the generated output itself carries a detectable signal — for example metadata or watermarking — showing it was artificially generated or manipulated, in a format machines can read. A visible caption alone is a disclosure to readers; the marking duty is about the file or output carrying the signal too, as far as technically feasible.
No. This guide and the SkillQuesters classification report are educational information based on the public text of the EU AI Act (Regulation (EU) 2024/1689, Article 50). They do not replace a lawyer.
SkillQuesters is independent and is not affiliated with the European Commission. Primary source: Regulation (EU) 2024/1689, Article 50, and the European Commission’s AI Act pages. Timeline information as of November 2026. This page is educational information, not legal advice.