EU AI Act Article 50 Explained: Marking & Disclosure Rules

Article 50 of the EU AI Act sets transparency obligations for certain AI systems: users must be told when they interact with AI, synthetic audio, image, video and text must be marked in a machine-readable and detectable format, and deepfakes and AI-generated public-interest text must be disclosed to recipients.

50(1) — Telling people they are talking to AI

Systems that interact directly with people must make that interaction evident, unless it is obvious to a reasonably well-informed person from the context. The disclosure has to be given at the first interaction, not buried in terms of service. This is the easiest clause to satisfy and the one most often satisfied only cosmetically.

50(2) — Machine-readable marking of synthetic output

Providers of systems that generate synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The text specifically points at technical solutions that are effective, interoperable, robust and reliable as far as technically feasible. That wording matters: a single fragile mechanism is not a defensible answer, because robustness is judged against how content actually travels.

50(4) — Deepfakes and public-interest text

Deployers who generate or manipulate image, audio or video that resembles real people, objects, places or events must disclose that the content is artificially generated. Deployers who publish AI-generated text on matters of public interest must disclose it too, unless the output went through human review with editorial responsibility. There are narrower carve-outs for artistic, satirical and fictional works, where disclosure must not spoil the work but still has to be present.

What 'machine-readable' realistically means

In practice it means the file itself, or a record tied to the file, can be read by software without human interpretation. Embedded credentials, watermarks and externally registered declarations each solve part of the problem and fail in different ways. Embedded data is stripped by many platforms. Watermarks survive re-encoding but carry little detail. External records survive everything but need a way to link back to the file. Serious deployments layer all three.

What Article 50 does not require

It does not require proof that content is true, does not require identification of the human author, and does not mandate any single vendor's format. It also does not turn a probabilistic detector into compliance evidence. Read narrowly, the obligation is about declaration and detectability — which is why the durable answer is a declaration you can show was made, not an after-the-fact analysis.

Frequently asked questions

Who is responsible under Article 50 — the model provider or the publisher?

Both, in different ways. Providers must mark generated output as artificially generated; deployers who publish deepfakes or AI-generated public-interest text must disclose that to recipients.

Does Article 50 apply to AI-assisted editing?

It applies where content is generated or substantially manipulated by AI in a way that resembles reality. Routine corrections that do not materially alter the content are treated differently.

Are satire and art exempt?

Not exempt, but the disclosure is lighter: it must be made in a way that does not hamper the display or enjoyment of the work.

What happens if metadata is stripped by a platform?

Stripping is expected, which is why marking should not rely on one carrier. Pairing embedded credentials with a watermark and an externally registered declaration keeps the disclosure recoverable.

See a machine-readable declaration

Inspect how a marked asset is checked from the outside.