Article 50 of the EU AI Act: transparency obligations explained

Of all the obligations introduced by the EU AI Act (Regulation (EU) 2024/1689), the one on transparency is probably the most visible to end users and one of the most concrete to implement. It's the obligation that answers a simple but increasingly urgent question: *am I talking to a person or to an AI? Is this content real or artificially generated?*

Article 50 of the AI Act puts this principle in writing. In this article we look at what it requires exactly, whom it applies to, how it translates into concrete examples, and how to implement a disclosure that is both human-readable and machine-readable.

What Article 50 requires

Art. 50 sets transparency obligations for certain AI systems. In short, providers and deployers of systems that interact with people or that generate or manipulate content — text, audio, images, video — must inform users clearly and distinguishably that they are interacting with an AI, or that the content is artificially generated or manipulated.

Two elements deserve attention. First, the *timing*: the information must be given at the latest at the moment of first interaction or exposure. It's not enough to bury it at the bottom of a terms-of-use page; it must arrive when the user encounters the AI or the content. Second, the *clarity*: the communication must be clear and distinguishable — recognizable without interpretive effort.

Note that this part of the transparency obligations falls within the tranche of measures becoming applicable in 2026, according to the Regulation's staggered timeline. But, as with the other obligations, it pays to prepare earlier.

Whom it applies to

Art. 50 concerns two broad families of cases.

The first: AI systems designed to interact directly with people. This is the typical case of chatbots and conversational agents. Here the obligation is to inform the user that they are interacting with an AI system — unless it's already obvious from the context to a reasonably observant person.

The second: systems that generate or manipulate content. Text, audio, images, and video produced or altered by AI must be recognizable as such. This is the part that directly touches the world of synthetic content and so-called deepfakes.

The obligations involve both providers (who build the system) and deployers (who use it under their own authority), with responsibilities distributed along the chain. For anyone putting a generative AI agent into production, this means transparency isn't a design detail: it's a requirement.

Why transparency matters

Behind the formal obligation there's a substantive reason. When a person doesn't know they're interacting with an AI, they make decisions on the wrong basis: they attribute to a machine the reliability they'd reserve for a human being, or they take as authentic content that has been generated or manipulated. The transparency of Art. 50 serves to restore an elementary condition of informed trust: the person has the right to know what they're dealing with, so they can judge accordingly.

This also explains why the legislator insists on *timing* and *clarity*. Information given too late, or phrased ambiguously, doesn't produce the intended effect: the person has already interpreted the situation as if the AI weren't there. Transparency, to be useful, must arrive while it can still shape judgment.

Practical examples

Let's see what changes concretely.

A customer-support chatbot. An AI agent that answers users on a website must make it evident, from the start of the conversation, that the interlocutor is an AI. A clear opening line ("You're talking to a virtual assistant") satisfies the obligation far better than an ambiguous icon.

Content generated by an agent. If an AI agent automatically produces product descriptions, summaries, images, or audio intended for the public, that content falls within the scope of Art. 50 when it is artificially generated or manipulated. The exposed user must be able to recognize it as such.

An agent operating behind the scenes. Even when the AI doesn't "speak" directly with the user but produces output that will reach people, it's worth asking whether and how to signal the artificial nature of what is being communicated.

The common thread is that the obligation activates at the point of contact with the person: where a human interacts with the AI or is exposed to its content.

How to implement a disclosure

An effective disclosure works on two complementary planes.

The human-readable plane is the communication aimed at the person: a message, a label, a notice that clearly says "this is an AI" or "this content is artificially generated." It must be visible at the right moment — first interaction or exposure — and phrased so that anyone understands it.

The machine-readable plane is the marking that makes the information legible to systems: metadata, markers, or technical signals embedded in the content, allowing other platforms and tools to automatically recognize the synthetic nature of what circulates. It's the plane that makes transparency scalable beyond a single human glance.

The two planes aren't alternatives: a good implementation keeps them together. The person sees a clear notice; downstream systems read a consistent signal. A third aspect, often overlooked, is consistency over time: the disclosure must remain valid as the agent evolves, changes version, or expands its functions. Documenting where and how the information is shown helps the deployer demonstrate, if asked, that it met the obligation in the right place and at the right time. In all cases, the golden rule is the same: clear, distinguishable, at the right moment.

One last operational tip: verify that the disclosure actually works in the real context. A notice can be technically present but effectively invisible — hidden below the fold, shown for a fraction of a second, phrased in jargon the user can't decipher. The right question isn't "is the notice there?" but "does the average user notice and understand it, at the right moment?" Only by answering this question does transparency move from being a formal box-tick to being an effective safeguard.

Transparency and autonomous agents

Autonomous AI agents add a particular difficulty. A static chatbot declares its nature once and that's the end of it. An autonomous agent, on the other hand, can change behavior, generate ever-new content, activate different functions depending on the context. The disclosure must hold up to this variability: showing it once isn't enough if the agent then starts producing content that falls within the scope of Art. 50 without signaling it.

Then there's the question of the chain: an agent may integrate components from several providers, and the final output may arise from the combination of multiple systems. In these cases transparency must be thought of at the level of the user's overall experience, not the single component: what matters is that the person, in the end, knows they're facing an AI or artificial content. Reasoning in terms of user experience, rather than individual technical module, is the most robust way to avoid gaps.

DAMM and the transparency block

When an AI agent produces an output through a decision framework, transparency can be built into the process instead of being added afterward. In DAMM's case, every report generated via API includes a transparency block: a section that makes explicit that the analysis is produced by an AI system. It's one piece that helps the deployer respond to the spirit of Art. 50 at the point where the output reaches a person.

Let's be honest: a transparency block in a report covers one specific requirement — signaling the artificial nature of that content — but does not by itself exhaust all transparency obligations, nor compliance with the AI Act as a whole, which touches roles, risk classification, and documentation. If you want to go deeper, our page on how DAMM helps with the transparency required by the EU AI Act explains where the framework intervenes and where other measures remain necessary.

*This article is for informational purposes only and does not constitute legal advice.*