Article 50 of the EU AI Act sets four transparency duties for AI agents in any risk tier: tell people they are interacting with AI, mark synthetic content, tell people when emotion recognition or biometric categorisation is used on them, and disclose deep fakes and certain published text. Proof comes from a record of what was said.
What does Article 50 require, in plain words?
Article 50 is the EU AI Act's transparency article. It applies to systems in any risk tier, including those that are not high-risk, and adds to any high-risk duties a system also carries. Its four duties, in substance:
- Interaction. Providers must design systems that interact directly with people so those people are informed they are dealing with an AI system, unless that is obvious from the circumstances.
- Marking. Providers of systems generating synthetic audio, images, video or text must mark outputs in a machine-readable way so they are detectable as AI-generated, with exceptions for assistive editing.
- Emotion recognition and biometric categorisation. Deployers must inform people exposed to these systems.
- Deep fakes and public-interest text. Deployers must disclose deep fakes, and disclose AI-generated text published to inform the public on matters of public interest unless it has undergone human review and someone holds editorial responsibility.
Information must be clear and distinguishable, given at the latest at the first interaction or exposure, and accessible. The underlying idea is covered in what AI transparency means.
Who does each duty fall on?
Duties one and two fall on providers, who build the system. Duties three and four fall on deployers, who use it. A firm that builds its own agent is both and carries all four where they apply.
The split is less clean in practice. A deployer that renames a vendor's agent with a human name, or routes its output through templates that strip metadata, can undermine a provider's disclosure or marking. So deployers should treat the provider's transparency features as controls they must preserve.
What changes when the AI system is an agent that acts?
A chatbot waits for a person to start a conversation. An agent can start one. That creates cases a chatbot never raised.
Outbound agents. An agent that phones a supplier to chase an invoice, emails a hospital to request records or messages a customer about a missed payment is initiating interaction. The person on the other end must be told, at the start, that they are dealing with an AI system acting for your firm.
Agents acting for a customer. An agent that contacts a third party on a customer's behalf should disclose to that third party, too.
Agent-to-agent exchanges. Where two agents exchange messages with no person involved, the interaction duty, which protects natural persons, does not bite. Once a person reads or joins the thread, it does.
Mixed authorship. Agents draft content that people send. Whether the reader must be told depends on whether a person truly reviewed and took responsibility, as explained in Article 50 and agents that write to customers.
How do you meet Article 50 and prove it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Inform people they are interacting with AI | Disclosure fixed into every inbound and outbound channel at first contact | Channel scripts, templates and conversation records |
| Disclosure not undermined | Personas, names and voices reviewed so they do not suggest a person | Persona approvals |
| Mark synthetic content | Machine-readable marks on generated text, audio and images, backed by a signed record | Marked samples and the generation record |
| Inform people of emotion recognition or biometric categorisation | Agents not given these capabilities unless approved; notice built in where they are | Capability register and notice text |
| Disclose deep fakes and public-interest text | Generated public content labelled unless reviewed with named editorial responsibility | Publication records and reviewer sign-off |
| Prove it happened | Each disclosure and mark filed with the action that produced it | Append-only record per interaction |
For how to mark content so it survives conversion and forwarding, see Article 50 marking of AI-generated content. The proof row matters because Article 50 disputes are about a specific conversation. A policy that says "we disclose" is weaker than a transcript that shows the words. A generic log of system events rarely captures this, one reason we argue Article 12 logging is not enough for agents.
What will a supervisor ask?
- Which of your agents interact with people, and in which channels?
- Which agents start conversations, and what do they say first?
- Which outputs do you treat as exempt from marking, and why?
- Do any agents use emotion recognition, for example to detect frustration on calls?
- Show us the disclosure in a specific conversation from last month.
The emotion question catches teams out. Voice agents that adjust tone based on detected caller emotion may be using emotion recognition, which brings its own notice duty, and in some workplace and education settings is prohibited outright.
What does Article 50 not cover?
It does not decide whether an agent is high-risk; that comes from Annex III and the related rules. It does not require explanation of individual decisions, which sits with other provisions and with GDPR. It does not give people a right to a human. And it does not override sector rules on communications, such as consumer protection duties, which apply whether a person or an agent writes. For the full picture, see AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme files each disclosure and each generated output with the action that produced it, in the same append-only record as the verdict. That gives one place to answer what a person was told, what was marked, and what the agent was permitted to do in that conversation.
Questions people ask
Does Article 50 apply to agents that are not high-risk?
Yes. Article 50 applies across risk tiers. A low-risk customer service agent still has to disclose that it is an AI system.
Does an internal agent used only by staff need to disclose?
If staff interact with it directly, they should be informed, though this is often obvious from context, such as an internal tool labelled as AI. Generated content it produces may still need marking.
Can a firm rely on the provider for all of Article 50?
Not entirely. Providers build disclosure and marking in, but deployers control channels, personas and publication, and carry their own duties on emotion recognition, deep fakes and public-interest text.