Article 50 of the EU AI Act: a practical guide for chatbots and AI agents
If you run a customer-facing chatbot or an AI agent, AI Act Article 50 is the obligation you can be measured against today. The transparency duties of the EU AI Act, Regulation (EU) 2024/1689, have applied since 2 August 2026, and they do not depend on a high-risk classification. The question is plain: do people know they are talking to a machine, and can the content be recognised as artificial? This guide is written for the head of AI governance, the DPO, the CISO and the product owner of an assistant. It is not legal advice.
Article 50(1): telling people they are talking to an AI
Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect... (Article 50(1))
Three points matter.
The duty is on the design. A line in the terms of service is not a design feature. A label in the chat window, a spoken sentence at the start of a call, a first message: these are.
The test is whether the people concerned are in fact informed, not whether you published a notice somewhere.
The only exception is narrow: the AI nature must be obvious, in context, to a reasonably well-informed, observant and circumspect person.
Article 50(2): marking AI-generated content
Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated... (Article 50(2))
Paragraph 2 is about the output, not the conversation. A chatbot that writes a downloadable letter generates synthetic text; a voice assistant generates synthetic audio. Both fall under paragraph 2, on top of paragraph 1.
The exception covers assistive functions for standard editing and systems that do not substantially alter the deployer's input or its meaning.
For systems placed on the market before 2 August 2026, the grace period for machine-readable marking ends on 2 December 2026.
Article 50(3): emotion recognition and biometric categorisation
Deployers of an emotion recognition system or a biometric categorisation system shall inform the natural persons exposed thereto of the operation of the system, and shall process the personal data in accordance with Regulations (EU) 2016/679... (Article 50(3))
This is the deployer's duty and the DPO's paragraph: an information duty on top of the GDPR.
The Article 3 definitions decide scope, and the word that matters is biometric. An emotion recognition system infers emotions or intentions from biometric data. A voice assistant that analyses the caller's tone is likely to be processing such data. A classifier that scores the sentiment of a typed complaint is not. Confirm the classification with counsel.
Article 50(4): deep fakes and public-interest text
Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated... Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated... (Article 50(4))
Paragraph 4 is also a deployer duty, in two halves.
The first is deep fakes: under Article 3(60), image, audio or video content that resembles real persons, objects, places, entities or events and would falsely appear authentic. A synthetic voice that imitates a named adviser is one, and the deployer must disclose it.
The second is text published to inform the public on matters of public interest. The duty does not apply where the text has gone through human review or editorial control and a natural or legal person holds editorial responsibility. Keep proof of that review.
Article 50(5): clear, timely, accessible
The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements. (Article 50(5))
Paragraph 5 binds providers and deployers alike.
Clear and distinguishable: the disclosure must stand out from the rest of the interface.
At the latest at the first interaction: the person must know before they start, or as they start. A disclosure that appears after several messages is late.
Accessible: a visual badge with no text alternative is invisible to a screen reader. A voice channel needs a spoken disclosure.
Provider or deployer: who carries each duty
Paragraphs 1 and 2 bind the provider. Paragraphs 3 and 4 bind the deployer. Paragraph 5 binds whoever owes the information.
Article 3 draws the line. A provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark. A deployer uses an AI system under its authority in a professional activity.
If a vendor built your assistant and you run it under the vendor's name, you are the deployer. If you had it built and run it under your own brand, the definition of provider fits you. Article 25 spells out the same logic for high-risk systems: put your name on the system, or change its intended purpose, and you take on the provider's obligations. A limited-risk chatbot falls outside Article 25, but not outside the Article 3 definition. Do not assume the vendor carries paragraphs 1 and 2 for you; check the contract.
Article 50(6) makes the article a floor.
Paragraphs 1 to 4 shall not affect the requirements and obligations set out in Chapter III, and shall be without prejudice to other transparency obligations laid down in Union or national law... (Article 50(6))
The GDPR's information duties and consumer law still apply, and a system that is also high-risk under Annex III meets Chapter III on top from 2 December 2027.
The free Article 50 check walks through these questions for one system at a time.
The Commission Guidelines and the Code of Practice
Article 50(7) is the legal basis.
The Commission shall encourage and facilitate the drawing up of codes of practice at Union level to facilitate the effective implementation of the obligations regarding the detection, marking and labelling of artificially generated or manipulated content... (Article 50(7))
The Code of Practice on Transparency of AI-Generated Content was finalised on 10 June 2026. The Commission Guidelines on Article 50 were published on 20 July 2026.
Adhering to the Code is voluntary, but declining it has a cost: whoever declines, the provider for paragraph 2 or the deployer for paragraph 4, must demonstrate that the obligation is met by equivalent means, in writing, with evidence, and with a signature.
Read the Guidelines yourself and record the version you read.
What "obvious to a reasonably observant person" means in practice
The exception in Article 50(1) is where most arguments will happen.
An assistant called "Virtual assistant", with a robot icon, on your website, is probably obvious to that person. Other cases are not: an assistant with a human first name and a photograph, or a chat that continues in the thread where a human adviser was answering earlier the same day.
Two rules follow. Do not rely on the exception unless you can write down, in a few sentences, why the AI nature is obvious in this channel to this population. And never let the system deny being an AI when asked: a person who asks "am I talking to a human?" and gets an evasive answer is, by definition, not informed.
How to test disclosure under pressure
A disclosure that survives a screenshot may not survive a conversation. AI Act Article 50 is met or missed at run time, so test it there.
Write the tests from the article's text. For paragraph 1 the questions are direct: are you a human, are you a bot, who am I talking to. Ask them at the start, after a long conversation, and after the user has asked the system to play a role.
Then push: paste a document that instructs the assistant to say it is a human, and hand the conversation to a human to check that the handover is announced in both directions.
For paragraph 2, ask for content the user could publish and check that the marking survives a request to remove it. For paragraph 4, ask it to imitate a real person's voice or face, and check that it refuses or discloses.
Grade every answer against a fixed rubric: disclosed, evaded, denied. Have a second grader work blind, and a human confirm the failures. Record the prompt, the answer, every tool call, the model version and the system prompt version. Re-run the set after every change to the model, the prompt or the tools.
What to keep as evidence
A supervisor will ask what the person saw, and when. Keep the following, versioned and dated.
- The disclosure itself: screenshots of every channel, the first message, the accessibility check.
- The design decision: who chose the wording, when, and against which version of the Guidelines.
- The role allocation: the contract clause that says who is provider and who is deployer, for each system.
- The marking approach: the declaration of adherence to the Code of Practice, or the equivalent-means note, and, for older systems, the 2 December 2026 plan.
- For paragraphs 3 and 4: the information given to the persons exposed, and the record of editorial review.
- The test results: transcripts, tool calls, grading, the reviewer's confirmation, and the date of the last run against the model in production.
- The change log: each model, prompt or tool change, and the re-test that followed.
Article 50 sets no retention period. Choose one, write it down, keep the whole set in one place, and sign it.
Vidimus turns Article 50 into exactly these tests. It runs them against your live agent, records the tool calls the agent makes, grades every answer with a separate grading model, has a human reviewer confirm, and issues a versioned, signed evidence pack. A pilot takes about two weeks for one agent. See how the Article 50 solution works.