Three duties that are easy to confuse and must not be

Article 50 of the EU AI Act is short and it is regularly summarised as "you have to label AI content". That summary merges three distinct duties which fall on different parties, take different forms and are discharged in different ways. Merging them is the most common and most expensive mistake in this area.

Duty one: tell people they are talking to a machine

Article 50(1) requires that AI systems intended to interact directly with natural persons are designed and developed so that the persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a reasonably well-informed, observant and circumspect person, taking account of the circumstances and the context of use.

The exception is narrower than deployers would like. Obvious means obvious to an ordinary person in that situation, which is not the same as obvious to the team that built it. A chat widget whose responses are indistinguishable from a person's is not obvious merely because the industry knows such widgets exist.

Duty two: mark the output machine-readably

Article 50(2) falls on providers of AI systems that generate synthetic audio, image, video or text. They must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

This is a technical duty on the party that builds the generator. It is invisible to a human. It exists so that platforms, verification tools and downstream systems have something to check, and so that provenance can travel with a file rather than with a claim about a file.

Two dates matter. Article 50 applies from 2 August 2026 and was not deferred by the Digital Omnibus, which pushed back other parts of the Act. For systems already on the market before 2 August 2026, the marking grace period ends 2 December 2026. After that there is no grandfathering left.

Duty three: disclose a deepfake, visibly

Article 50(4) falls on deployers. A deployer of an AI system that generates or manipulates image, audio or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated. There is a limited exception where the content is evidently artistic, creative, satirical or fictional, and in that case the disclosure is made in an appropriate manner that does not hamper the display or enjoyment of the work.

Article 50(4) also reaches text: where AI-generated or manipulated text is published with the purpose of informing the public on matters of public interest, its artificial origin must be disclosed, unless the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.

The point on which organisations get this wrong

A deployer cannot discharge the Article 50(4) disclosure duty by pointing to the provider's machine-readable marking. This is worth stating plainly because the argument is attractive and it fails. The reasoning runs: the file carries an embedded marker, the marker says it is synthetic, therefore it has been disclosed.

The Commission's guidance on Article 50 requires a visible or audible label that is understandable without a detection tool. A marker a member of the public cannot perceive, and could not check even if they knew to look, does not inform anyone. The two duties do different work: the machine-readable mark serves systems, the visible label serves people. Discharging one leaves the other outstanding.

Article 50(5) ties the timing down: the information must be provided at the latest at the time of the first interaction or exposure, clearly and distinguishably. Not in a terms page, not on request, not afterwards.

Who bears what

If you buy a generation tool and publish its output, the provider owes the machine-readable marking and you owe the visible disclosure. Asking a vendor whether their tool is Article 50 compliant is therefore a question with a limited answer. They can tell you about 50(2). Only you can discharge 50(4), because only you know what you published, in what context, and to whom.

As of 24 September 2026.