Article 50: telling people when they are dealing with AI
The transparency duties have applied since August 2nd, 2026. They are the part of the AI Act most likely to reach an ordinary business, and they do not depend on your system being high risk.
The date passed, and the duty is now live
Until August 2026 the transparency rules were something to prepare for. They are now something to comply with, and a supervisory authority can ask about them.
Applies now
All four transparency duties, for systems placed on the market or put into service from August 2nd, 2026 onwards.
One transition, until December 2026
For systems already on the market before August 2nd, 2026, the machine-readable marking duty under Article 50(2) applies from December 2nd, 2026. Nothing else in Article 50 waits.
Untouched by the Omnibus
Regulation (EU) 2026/1744 deferred the high-risk obligations to December 2027 and August 2028. It left the Article 50 date where it was.
Content generated before August 2nd, 2026 does not have to be labelled retrospectively. The duty runs forward, not backward.
Four duties, and they do not all land on the same person
Two fall on the provider of the system and two on the deployer using it. The most common mistake is assuming the vendor has taken care of all of them.
Say that it is an AI system
A system that interacts directly with people has to be designed so they are informed they are dealing with AI, unless that is obvious to a reasonably well-informed and observant person. This is the one that catches website chatbots, voice agents and automated support.
Falls on the providerMark synthetic output so a machine can read it
Systems generating synthetic audio, image, video or text must mark their output in a machine-readable format that is detectable as artificially generated or manipulated. The marking has to be effective, interoperable and robust as far as is technically feasible, which is a technical requirement rather than a notice on a page.
Falls on the providerTell people exposed to emotion or biometric systems
Anyone operating an emotion recognition or biometric categorisation system has to inform the people exposed to it that it is running, and to keep the data protection obligations that come with it. This duty sits with whoever uses the system, not whoever built it.
Falls on the deployerDisclose deep fakes and published AI text
Deployers generating or manipulating image, audio or video content that resembles real people, places or events must disclose that it is artificial. The same goes for AI-generated text published to inform the public on matters of public interest, unless a person has reviewed it and someone holds editorial responsibility for it.
Falls on the deployerArticle 50(5) sets the manner: the information has to be clear, distinguishable and given at the latest at the moment of first interaction or exposure. A disclosure buried in terms of service, or added after the conversation, does not satisfy it. Accessibility requirements apply to it as well.
Where the duty stops
The exemptions are real, and reading them properly saves work. They are also narrower than they sound, and each one has to be argued on the facts of your system rather than assumed.
Article 50 also sits alongside the rest of the Regulation. Meeting it does not discharge the high-risk obligations in Chapter III, or any transparency duty you already owe under other Union or national law.
Check which duties reach youCarve-outs in the text
- Interaction that is obvious to a reasonably well-informed, observant and circumspect person
- Assistive editing functions and changes that do not substantially alter the input
- Artistic, creative, satirical and fictional works, where disclosure need only note that the content exists and must not spoil the display of the work
- Published text that a person genuinely reviewed and for which someone holds editorial responsibility. A spell check is not a review
- Uses authorised by law to detect, prevent, investigate or prosecute criminal offences, with the safeguards that go with them
The situations that come up most often
None of these need a high-risk system. They are ordinary uses of ordinary tools, and each one carries a duty under Article 50.
A support chatbot on your website
It interacts with people directly. If a visitor could reasonably take it for a colleague, it has to say what it is, at the start rather than on request.
Marketing copy, images and video from a generator
Marking sits with the provider of the tool. Publishing content that looks like a real person, place or event brings the deployer disclosure into play as well.
A synthetic voice on your phone line
Direct interaction and generated audio at the same time, which is why voice deployments usually need both a spoken notice and a written one.
Sentiment or attention analysis on customers or staff
Emotion recognition. The people exposed to it have to be told it is operating, and the data protection questions arrive at the same moment.
AI-written articles on your news or blog
Text published to inform the public on matters of public interest needs a label, unless a person reviewed it and someone is answerable for it editorially.
An AI plugin someone installed on a client site
The duty follows the system, not the invoice. Agencies and builders are frequently the deployer, and sometimes the provider, without having agreed to be either.
Breaches of Article 50 sit in the tier reaching fifteen million euro or three per cent of total worldwide annual turnover, whichever is higher. Proportionality applies to small and medium-sized organisations, but the tier is the tier.
Five things worth doing this month
Article 50 is one of the cheaper parts of the Regulation to get right, provided it is done deliberately and written down.
List every system that talks or generates
Chatbots, assistants, voice systems, content generators, anything embedded in a product you sell. Shadow tools your teams adopted on their own belong on the list too.
Decide what you are for each one
Provider or deployer, system by system. The answer decides which of the four duties is yours, and it can differ across two systems in the same office.
Write the notice, and put it where it lands first
Short, plain, and present at the first interaction. This is a drafting job worth doing once properly rather than four times badly.
Ask your vendors about the marking
Machine-readable marking under Article 50(2) is the provider duty you are most likely to be relying on somebody else for. Get the answer in writing.
Diary December 2nd, 2026
The transition for systems already on the market before August 2026 closes then, along with the new prohibited practices added by the Omnibus.
Not sure which of the four is yours?
The free assessment asks about how your systems are used and comes back with the provisions that actually apply, Article 50 among them.