Article 50 Is Four Duties, Not One

Most writing about the transparency obligations, including some of our own earlier posts, treats Article 50 as one rule about telling people when AI was involved. It is closer to four separate duties with different triggers, different subjects, and different parties on the hook. Teams that read it as a single requirement tend to over-apply it in one place and miss it entirely in another.

The distinction that organizes all four is whether a duty binds the provider, meaning whoever builds and places the system on the market, or the deployer, meaning whoever uses it in the course of their work. Most professional services firms are deployers, which means two of the four duties are primarily somebody else’s problem and two of them are squarely theirs.

The two provider duties

The first is interaction disclosure. A system that deals directly with a person has to be built so that person is informed they are dealing with AI, unless it is obvious from the circumstances. The obligation sits with whoever designs the system, which is why a firm using a third-party chatbot is relying on that vendor having done it rather than adding a notice themselves.

The second is machine-readable marking. Providers of generative AI have to embed markings in synthetic audio, image, video and text output so the content is detectable as artificially generated by software further downstream. This is the duty carrying the grace period to 2 December 2026 for systems already on the market before August, and it is the one most likely to reach a professional firm indirectly, through whatever tools it has bought.

The two deployer duties

The third concerns emotion recognition and biometric categorization. A deployer running either has to inform the individuals exposed to it. For most consultancies, agencies and research teams this is simply out of scope, and it is worth saying so plainly rather than leaving teams to wonder whether a clause about biometrics somehow applies to document analysis.

The fourth is the one that actually lands on professional work. Deployers have to disclose AI-assisted or manipulated content, with the sharpest attention on synthetic media and on text published to inform the public on matters of public interest, unless a human editor took editorial responsibility for it. That final clause is the hinge, and it is where the regulation says something about its own intent: accountability substitutes for labeling. If a named person stands behind the output, the pressure to flag it as machine-produced eases, because what the rule protects is the reader’s ability to know whether anyone is answerable.

Why the split matters more than the count

Knowing there are four duties is trivia. Knowing which two are yours changes what you build. A firm that reads Article 50 as one undivided rule tends to respond by adding a disclaimer to everything, which addresses the fourth duty badly and the other three not at all, while creating a written claim about its process that it may not be able to substantiate.

The more productive response is to establish the split in writing, which is also what a client or an auditor will ask for. We publish our own view of which obligations sit with us and which sit with the organization using the platform, because a division of responsibility that lives only in a sales conversation is not much use when somebody needs an answer.

Qonera is the AI governance platform for professional teams, and the review and approval workflow is built around the deployer position specifically: oversight is a gate rather than an assumption, sign-off carries a name, and the record of what was checked is produced by the work instead of assembled afterward. The obligation-by-obligation division is set out in the responsibility matrix. Reading Article 50 as one rule is the fastest way to spend effort in the wrong place, and the split is not hard to establish once somebody asks the question.

This article is for general information only and does not provide legal advice. Organisations should consult qualified legal counsel about how Article 50 and the EU AI Act apply to their specific systems, workflows, and obligations, including whether they act as a provider or a deployer.