Article 50 AI Act · decision tree
Article 50 contains four obligations, split across two roles. Which one applies depends on what your AI does and on your role. Answer a few questions and see straight away which paragraph applies, how to implement it according to the Commission guidelines, and what that looks like in practice.
Question 1 · Art. 50(1) to (4)
What does the application do?
Choose the function you want to check now. Obligations can apply at the same time (Guidelines para. 8). Purely personal use and pure research fall outside Article 50 (Guidelines para. 19, 21). Your outcomes are collected below in one overview.
Article 50 at a glance
Paragraphs 1 and 2 rest on the provider, paragraphs 3 and 4 on the deployer. Paragraph 5 sets how and when you inform people for all four. One organisation can hold both roles.
Design systems that communicate directly with people so that they know they are dealing with AI, including agents that email or call.
Exception: not where this is obvious to a reasonably well-informed, observant and circumspect person, and not for systems authorised by law for law enforcement, unless the public can use them to report a criminal offence.
Mark synthetic audio, images, video and text in a machine-readable format and make them detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust and reliable as far as technically feasible.
Exception: not for standard editing that does not change the meaning, such as spelling, translation or transcription, not for short texts, source code or intermediate steps, and not for law enforcement authorised by law.
If the system was on the market before 2 August 2026: from 2 December 2026 at the latest.
Inform the people exposed to emotion recognition or biometric categorisation about its operation, and process their data in line with the GDPR.
Exception: not for law enforcement permitted by law, with appropriate safeguards. Check Article 5 first: emotion recognition at work and in education is prohibited, except for medical or safety reasons, and biometric categorisation that infers race, political opinions, religion or sexual orientation is prohibited.
Disclose that a deep fake has been artificially generated or manipulated, and that text you publish to inform the public on matters of public interest was generated or manipulated with AI.
Exception: for evidently artistic, satirical or fictional work a lighter disclosure suffices. Text that has been substantively reviewed by a person, with a publicly identified responsible party, falls outside. Law enforcement authorised by law is exempt.
Provide the information in a clear and distinguishable manner, at the latest at the first interaction or exposure, and meet the accessibility requirements.
Paragraph 6: paragraphs 1 to 4 leave the high-risk requirements unaffected and are without prejudice to other transparency obligations for deployers under Union or national law, such as consumer law. Paragraph 7: the Commission encourages codes of practice for detecting, marking and labelling AI content. The code of practice on transparency of AI-generated content was published on 10 June 2026 and is voluntary.
Examples
Nine applications we see most often in practice, with the role, the paragraph and what to do. If in doubt, walk through the decision tree for your own situation.
The supplier builds in the AI notice; you make sure your configuration or styling does not remove it. If you have the chatbot developed for you or substantially modify it, and put it into service under your own name or trademark, you become the provider.
Disclose in the first message that it is AI and answer truthfully when asked whether it is human. Mark the generated text in a machine-readable way, with a watermark as robust as technically feasible.
As long as it is not a deep fake, you have no disclosure duty of your own under Article 50; the provider marks the output. Note: a product image that makes the product look better than it is does count as a deep fake (Guidelines para. 116).
Label in the top right of the frame, at the start and after ad breaks. According to the guidelines, advertising is not automatically creative work, so the lighter regime usually does not apply.
Public, informative and on a matter of public interest? Then a label above the text, unless a subject expert reviewed the content and it is public who is responsible. An unreviewed AI translation can also fall under the obligation.
Text for a closed class or course group is probably not published: no text label. A public course on, say, health or climate can be, unless a subject teacher reviews it and the provider takes responsibility. A realistic AI video of a historical figure is a deep fake, in class too. The tutor discloses in child-friendly language that it is AI.
Based on voice or face: inform customers before first exposure and check the GDPR. Analysis of written text only is not emotion recognition. If the system also infers employees' emotions, that is prohibited. High-risk requirements apply from 2 December 2027.
Prohibited, except for medical or safety reasons, including a webcam feature that measures ‘engagement’. Informing people does not make it permissible. Pure fatigue detection that infers no emotions falls outside the definition.
Standard editing that does not change the meaning need not be marked (Guidelines para. 92). If the same tool also produces summaries or rewrites, you must mark that output.
Timeline
The Digital Omnibus did not postpone Article 50. Only the machine-readable marking of paragraph 2 has a transition for systems already on the market.
Voluntary code with concrete measures for marking and labelling.
The Commission's interpretation of the Article 50 transparency obligations.
All paragraphs apply. Only paragraph 2 has a transition for systems already on the market.
Applies nowMachine-readable marking also for systems on the market before 2 August 2026.
Questions
Yes. Article 50 has applied since 2 August 2026 and was not postponed by the Digital Omnibus. Only the machine-readable marking of paragraph 2 has a transition until 2 December 2026, and only for systems on the market before 2 August 2026. The chatbot disclosure is not covered by that transition.
No. Disclosure is required for deep fakes (image, audio or video) and for AI text you publish to inform the public on matters of public interest, unless that text has been substantively reviewed by a person and someone holds editorial responsibility (paragraph 4). Machine-readable marking is an obligation of the provider of the system that generates the content (paragraph 2).
Usually not as text: teaching material for a closed class or course group is probably not published, and much teaching content is not about a matter of public interest. A public course on, say, health or climate can fall under it, unless a subject teacher reviews the content and the provider takes responsibility. A realistic AI video of an existing person or historical event is a deep fake, in class too. An AI tutor falls under paragraph 1, and emotion recognition of pupils is prohibited.
You are a provider if you develop an AI system or have it developed and place it on the market or put it into service under your own name or trademark, whether for payment or free of charge (Article 3(3)). You are a deployer if you use an AI system under your authority, except for personal, non-professional use (Article 3(4)). Your employees are not separate deployers; your organisation is. One organisation can hold both roles.
Check per function. An AI assistant that talks and generates text falls under paragraph 1 and paragraph 2. Save each outcome and walk through the tree again for the next function; that way you build one overview.
Clearly, distinguishable from other information, at the latest at the first interaction or exposure, and accessibly (paragraph 5). For interaction at least at the start of every session, for content per output. A notice in the terms, a manual or a menu does not suffice (Guidelines paras. 141 to 144).
No. Content generated before 2 August 2026 need not be marked or labelled retroactively. Text generated then but published now must be labelled (Guidelines para. 154).
The code of practice on transparency of AI-generated content (10 June 2026) sets out concrete measures for marking and labelling. Signing is voluntary. Non-signatories must demonstrate compliance in another way, for example by comparing their measures with the code, and can expect more questions from the supervisor (Guidelines paras. 146 to 149).
A breach of Article 50 can lead to a fine of up to EUR 15 million or, for an undertaking, up to 3 percent of total worldwide annual turnover in the preceding financial year, whichever is higher (Article 99(4)). For SMEs, start-ups and small mid-caps the lower of the two applies (Article 99(6) and (6a)). Member States organise supervision and enforcement.
Sources
The questions, outcomes and practical measures were checked against the legal text, the Commission guidelines and the code of practice on 5 October 2026. A decision tree simplifies: it does not replace an assessment of your concrete situation.