EU AI Act Article 50: what agents must disclose, and when
The transparency obligations, read for teams that run AI agents. Four paragraphs, two roles, two dates, one fine — and the design question the law leaves to you.
Article 50 of the EU AI Act has applied since 2 August 2026. It does four things. People must be told when they are interacting with an AI system, unless that is obvious. Providers of generative systems must mark their outputs so machines can detect them as synthetic. Deployers must tell people when emotion recognition or biometric categorization is running on them. And deployers must disclose deep fakes and AI-written text published on matters of public interest, unless a person has reviewed it and holds editorial responsibility. An agent that chats, emails, calls or publishes on a company’s behalf is inside the first paragraph from the first message. The fine is up to €15 million or 3% of worldwide turnover. The law says what must be disclosed and when; where the disclosure appears so that a person actually sees it is a design decision, and that is the part this dispatch is for.
Most coverage of the AI Act is about high-risk systems, conformity assessments and the register. Those obligations touch a minority of companies. Article 50 touches almost everyone who has shipped an agent, because its first paragraph is not about risk class at all. It is about a person on the other end of a conversation, and whether they know what they are talking to.[1]
This dispatch reads the article for that case: the support agent, the outbound email agent, the voice agent that calls a supplier, the content agent that drafts the newsletter. It is not legal advice and it does not classify any system. It sets out what the text says, who it binds, when, what it costs to ignore, and where in an agent the disclosure has to live so that it counts.
01 · What Article 50 says, paragraph by paragraph
Article 50 sits in Chapter IV of Regulation (EU) 2024/1689 and is titled “Transparency obligations for providers and deployers of certain AI systems”.[1] Four paragraphs carry obligations; three more say how and to whom they apply.
| Paragraph | Who | Obligation | Exception |
|---|---|---|---|
| 50(1) | Providers | Design systems intended to interact directly with natural persons so that those persons are informed they are interacting with an AI system | Where it is obvious to a reasonably well-informed, observant and circumspect person in the circumstances; law-enforcement systems authorized by law |
| 50(2) | Providers, including of general-purpose AI systems | Mark synthetic audio, image, video or text outputs in a machine-readable format, detectable as artificially generated or manipulated; techniques effective, interoperable, robust and reliable as far as technically feasible | Assistive functions for standard editing, or outputs that do not substantially alter the deployer’s input; law enforcement |
| 50(3) | Deployers | Inform the people exposed to an emotion recognition or biometric categorization system of its operation, and process their data under the GDPR and related rules | Law enforcement, subject to safeguards |
| 50(4) | Deployers | Disclose that a deep fake has been artificially generated or manipulated; disclose that text published to inform the public on matters of public interest was AI-generated or manipulated | Artistic, creative, satirical or fictional works, where disclosure must not hamper the work; text that has undergone human review or editorial control, where a person holds editorial responsibility; law enforcement |
| 50(5) | Providers and deployers | Provide the information in paragraphs 1 to 4 clearly and distinguishably, at the latest at the time of the first interaction or exposure, and in accordance with accessibility requirements | — |
| 50(6) | — | These obligations add to, and do not replace, the high-risk obligations of Chapter III and other Union or national law | — |
| 50(7) | The AI Office | Encourages and facilitates codes of practice at Union level for the detection and labeling of AI-generated content; the Commission may adopt implementing acts or common rules if the codes are inadequate | — |
↔ scroll table · paraphrased from the regulation text; the article itself is linked in the sources[1]
Three details decide how much of this an agent team has to care about. First, paragraph 1 does not depend on risk classification: a low-risk support agent is inside it the moment it talks to a person. Second, the “obvious” exception is judged from the person’s point of view in the circumstances, not from the company’s intent; a chat window labeled “assistant” may be obvious, an email signed with a name is not. Third, paragraph 5 sets the timing: clearly, distinguishably and at the latest at the first interaction. A disclosure buried in terms of service or shown after the third message is late.
Which paragraph binds your agent
02 · Provider or deployer: which one are you?
The article binds two roles, and most agent teams are one of them without having decided which. 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.[1]
| Setup | Provider | Deployer | Which paragraphs bind you |
|---|---|---|---|
| You buy an agent platform and run it with your data | The platform vendor | You | 50(3) and 50(4) as deployer; you still need the vendor’s 50(1) and 50(2) behavior to work in your channels |
| You build your own agent on a model API and put it into service under your name | You, for the agent system; the model vendor for the model | You | All four may apply to you; the marking in 50(2) depends on what the model vendor’s outputs already carry |
| An agency builds an agent for you and you run it | Depends on whose name it is placed into service under — decide this in the contract | You | 50(3), 50(4), and 50(1) if you are the provider |
| You use an agent only for internal work, no outside person | Vendor or you | You | 50(1) still applies to the employees it interacts with, unless obvious; 50(4) if its text is published |
↔ scroll table · auxfirst’s reading of the definitions in Article 3; the Commission’s draft guidelines address the boundary cases[7]
The case that catches teams is the third one. When an agency or a contractor builds an agent that goes live under the client’s brand, someone is the provider of that system, and the contract usually does not say who. It should. The role decides who has to design the first-interaction disclosure and who answers for the marking of generated content, and “we assumed the vendor did it” is not a defense available to a deployer whose agent talked to ten thousand customers without saying what it was.
03 · The dates
From publication to application
Two things are worth saying plainly, because both are being reported loosely. The first-interaction disclosure in paragraph 1 has applied since 2 August 2026 to every system in scope, new or old; there is no grace period for it. And the 2 December 2026 date is Article 111(4): it gives providers of generative systems that were already on the market a few more months to implement the marking in paragraph 2.[4] A company whose agent went live in September 2026 does not get it.
04 · Where the disclosure has to appear in an agent
The law fixes the standard: clearly, distinguishably, at the latest at the first interaction, from the point of view of a reasonably well-informed person in the circumstances. It does not say where in an interface that is. For an agent, that is the whole problem, because an agent does not have one interface. It has a chat window, an email it sends, a voice line it calls on, a document it hands over, a ticket it closes. Each is a first interaction for someone.
In auxfirst’s practice this is not a new question. Agent identity and disclosure is the first of the six supervision primitives we score commercial platforms on in the Agent Supervision Register: can a human tell which agent acted, on whose authority, and that it was not a person? It is the first thing on the list because an agent that cannot say what it is cannot be governed either. It is also the first of the AUX heuristics, visibility of agent intent and action, and the ninth, clarity about who is acting in a system with several humans and agents. Article 50(1) makes the last part of the primitive a legal obligation.
Where the disclosure lives, by channel
The grid follows the same rule as everything else we write about agents: a recommendation is not a control. A system prompt that says “always disclose that you are an AI” is a recommendation. The disclosure rendered by the channel, logged per interaction, is a control. Under the AI agent governance framework, disclosure belongs in the mandate of every action that leaves the building, which on the Action Heat Ladder is every action with exposure above zero.
The exception is real: a product called “AI Assistant” that opens with a robot icon in a chat window is probably obvious to a reasonably well-informed person. The temptation is to stretch it. An agent that sends email under a person’s name, or calls with a natural voice, or answers in a shared inbox next to human colleagues, is not obvious, however well the company knows what it is. The test is the person’s view in the circumstances. When in doubt, disclose, and keep the record.
05 · A disclosure checklist for agent teams
Seven steps, in order
06 · The fines
Article 99(4) puts non-compliance with the transparency obligations of Article 50 in the middle tier: administrative fines of up to €15,000,000 or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher.[2] For comparison, prohibited practices under Article 5 carry up to €35 million or 7%, and supplying incorrect information to authorities up to €7.5 million or 1%. For SMEs and start-ups, Article 99(6) applies the lower of the two amounts in each tier.[2]
Fines are imposed by national market surveillance authorities, which member states had to designate by August 2025, so enforcement practice will vary by country and will take time to form. The more immediate cost is the one the law does not price: an agent that was found to have talked to customers without disclosure is an agent whose every other claim about itself is now open to question.
07 · What is still open
- How “obvious” will be read. The Commission published draft guidelines on the scope of the Article 50 obligations in May 2026 and consulted on them;[7] the way national authorities apply the exception to agents in email and voice is the question that matters most for agent teams, and there is no case practice yet.
- Whether the code’s methods become the de facto standard. The Code of Practice on Transparency of AI-generated Content is voluntary, final since June 2026, and had about 190 signatories by the end of July 2026; the Commission and the AI Board have confirmed it as an adequate way to demonstrate compliance with 50(2), (4) and (5).[5] A provider that does not sign still has to meet the obligations by some method.
- Marking that survives the pipeline. Machine-readable marks are only as good as the last system that touched the file. Agents that generate, edit, export and send are a chain of such systems; which link is responsible when the mark is gone is a question the article does not settle.
- National enforcement. Who enforces, with what resources and what priorities, differs by member state. The first fines will tell more than the text does.
The AUX takeaway
Article 50 is the part of the AI Act that most agent teams are already subject to and least prepared for, because it looks like a compliance question and is mostly a design question. The law says people must be told, clearly, first. Whether they were is decided by the opening message, the email body, the first seconds of the call, and the receipt that proves it.
An agent that cannot say what it is cannot be governed either. Disclosure is not the last control you add; it is the first.
Treat it the way you would treat any other action with exposure: in the mandate, rendered by the channel, logged per interaction, owned by a named person. Done that way, Article 50 costs an afternoon per agent. Done the other way, it costs the trust the agent was built to earn.
Terms used here
- Provider
- Under the AI Act, the party that 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. Article 50 paragraphs 1 and 2 bind providers.
- Deployer
- Under the AI Act, a natural or legal person, public authority, agency or other body using an AI system under its authority, except in a personal non-professional activity. Article 50 paragraphs 3 and 4 bind deployers.
- Deep fake
- AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)).
- Machine-readable marking
- The Article 50(2) obligation on providers of generative systems: outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated, with techniques that are effective, interoperable, robust and reliable as far as technically feasible.
- Code of Practice on Transparency of AI-generated Content
- A voluntary code drawn up by independent experts and facilitated by the European AI Office, published in June 2026, with a providers section on marking and detection and a deployers section on labeling. The Commission and the AI Board have confirmed it as an adequate tool to demonstrate compliance with Article 50(2), (4) and (5).
- Agent identity and disclosure
- The first of auxfirst's six supervision primitives: can a human tell which agent acted, on whose authority, and that it was not a person? Article 50(1) makes the last part a legal obligation for systems that interact directly with people.
Questions people are asking
What does Article 50 of the EU AI Act require?
Four transparency obligations. Providers must design AI systems that interact directly with people so that those people are informed they are dealing with an AI system, unless that is obvious. Providers of systems that generate synthetic audio, image, video or text must mark outputs in a machine-readable format and make them detectable as artificially generated. Deployers of emotion recognition or biometric categorization systems must inform the people exposed to them. Deployers must disclose deep fakes, and AI-generated text published to inform the public on matters of public interest, unless it has undergone human review and someone holds editorial responsibility.
When does Article 50 apply?
From 2 August 2026, under Article 113. One grace period applies: providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking obligation in Article 50(2), under Article 111(4). There is no grace period for the obligation to tell people they are interacting with an AI system.
Does Article 50 apply to AI agents, or only to chatbots?
It applies to any AI system intended to interact directly with natural persons, whatever it is called. An agent that chats with customers, answers a support ticket, calls a supplier by voice or sends an email on a person's behalf interacts directly with people. The exemption is where the AI nature is obvious to a reasonably well-informed person in the circumstances; for an email that reads like a colleague wrote it, it is not.
Am I a provider or a deployer?
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. A deployer uses an AI system under its authority in a professional activity. A company that buys an agent platform and runs it is a deployer; the platform vendor is the provider. A company that builds its own agent on a model API and puts it into service under its own name is the provider of that system as well as its deployer. The guidelines the Commission has drafted address the boundary; where you sit decides which paragraphs bind you.
What are the fines for breaching Article 50?
Under Article 99(4), non-compliance with the transparency obligations of Article 50 can be fined up to 15 million euro or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups the lower of the two amounts applies. Fines are imposed by national authorities, so the practice will differ between member states.
What is the Code of Practice on Transparency of AI-generated Content?
A voluntary code drawn up by independent experts in a process facilitated by the European AI Office and published in June 2026, with a providers section on marking and detection and a deployers section on labeling deep fakes and AI-generated text, plus a set of EU icons deployers may use. The Commission and the AI Board have confirmed it as an adequate tool to demonstrate compliance with Article 50(2), (4) and (5); about 190 organizations had signed it by the end of July 2026. Adherence is voluntary; the obligations themselves are not.
Is disclosure a legal question or a design question?
Both, and the design part is where most teams fail. The law says people must be informed clearly at the first interaction. Whether a disclosure in a footer, a system prompt or a terms page meets that is a question about what the person actually saw. In auxfirst's practice, agent identity and disclosure is the first supervision primitive and the first thing scored in the Agent Supervision Register, because an agent that cannot say what it is cannot be governed either.
Method · The regulation text, the penalties article, the application article and the implementation timeline were read on 26 October 2026 from the AI Act Explorer maintained by the Future of Life Institute, and the code of practice pages from the European Commission’s site on the same day. Table 1 paraphrases the article; the article itself is authoritative and is linked. The role table, the channel grid and the checklist are auxfirst’s reading for agent teams, not the Commission’s guidance. This dispatch is not legal advice and does not classify any system under the Act. Figures and dates will be revised if the guidelines or enforcement practice change them.