For most companies the transparency obligations under Art. 50 are the most tangible duty in the AI Act. They apply not only to high-risk systems but to ordinary chatbots, image generators and AI-assisted text. They have applied since 2 August 2026; the Digital Omnibus, Regulation (EU) 2026/1744, did not move that date. This article describes who must disclose what and what implementation looks like.
Four duties, two roles
Art. 50 contains duties for providers and deployers. Providers must ensure that people learn when they interact directly with an AI system, unless this is obvious from the circumstances. Providers of generative systems must mark their outputs — text, image, audio, video — in machine-readable form so that they are detectable as artificially generated. Deployers of emotion recognition or biometric categorisation systems must inform the people exposed. And deployers must disclose deepfakes as well as AI-generated text published to inform the public on matters of public interest.
The chatbot notice
A support chat, an in-app assistant, a phone bot: anyone talking to an AI must know it. The notice belongs at the start of the interaction and must be clear and recognisable, not hidden in terms of use. The exception is where it is obvious to a reasonably informed person that they are interacting with an AI. Anyone relying on that should record the assessment; when in doubt, a short notice is the safer and trust-building solution.
Machine-readable marking
Marking under Art. 50(2) falls on the providers of generative systems, the makers of the models and tools, not on the companies using them. It must be as effective, interoperable and robust as the state of the art allows, for example through metadata, watermarks or combinations. The Commission published a code of practice on marking AI-generated content in June 2026 and adopted guidelines on Art. 50 in July 2026. Generative systems placed on the market before 2 August 2026 have a transition period until 2 December 2026; new systems must mark from day one.
Deepfakes and AI text
Deployers generating or manipulating image, audio or video content that appreciably resembles real people, places or events must disclose that it is artificially generated or manipulated. For evidently artistic, satirical or fictional works, a disclosure that does not hamper the display is enough. For AI-generated text, the disclosure duty applies when it is published to inform the public on matters of public interest — unless a human has reviewed it and holds editorial responsibility. A company blog whose posts are reviewed by a responsible person is therefore assessed differently from an automatically published news feed.
Before and after August 2026
Content generated before 2 August 2026 need not be marked retroactively. AI-generated text published after that date is subject to the disclosure duty even if it was generated earlier. As for the transition on machine-readable marking: it concerns only providers and only systems already on the market. Deployers’ disclosure duties apply without any transition.
Penalties
Breaches of Art. 50 fall into the middle tier of the fines catalogue: up to EUR 15 million or 3 percent of worldwide annual turnover, for SMEs the lower amount in each case. Enforcement has been carried out by national market surveillance authorities since August 2026.
Implementation in five steps
1. Record every point at which people interact with AI or AI content is published: chat, phone, email replies, website, social media, marketing material.
2. Determine the role at each point: are you the provider of the system or the deployer of a third-party tool?
3. Implement the chatbot notice visibly at every interaction point.
4. Set a rule for published AI content: who reviews, who holds editorial responsibility, when is it labelled?
5. Clarify with tool vendors whether their systems mark output in machine-readable form or fall under the transition period, and document the result.
The duties are summarised in the glossary entry on transparency obligations under Art. 50; how Art. 50 fits into the whole is shown in our AI Act 2026 guide. For classifying individual systems, AI Risk Check fits.