Article 50 of the EU AI Act: marking and labelling AI-generated content

Published 12 October 2026

Article 50 is the transparency article of the EU AI Act, Regulation (EU) 2024/1689. It is short, and most of it is about telling people when AI is involved: when they talk to it, when they see what it made, and when it reads their emotions. It has applied since 2 August 2026. This guide reads it paragraph by paragraph, then looks at what "machine-readable marking" means for an engineer.

The duties, paragraph by paragraph

The Act separates providers, who develop an AI system and place it on the market, from deployers, who use one under their own authority (Cooley, "EU AI Act: Transparency Obligations Take Effect 2 August 2026" (3 August 2026)). Article 50 gives each its own duties (Article 50 of the AI Act (artificialintelligenceact.eu)).

Paragraph	Who	Duty
50(1)	Providers	Systems that interact directly with people must tell them they are interacting with an AI system, unless that is obvious from the context.
50(2)	Providers	Systems that generate synthetic audio, image, video or text must mark their outputs in a machine-readable format, detectable as artificially generated or manipulated.
50(3)	Deployers	Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them.
50(4)	Deployers	Deployers must disclose deep fakes, and AI-generated text published to inform the public on matters of public interest, with the exceptions set out below.

Article 50(5) adds that this information must be given in a clear and distinguishable way, at the latest at the first interaction or exposure.

When each part applies

All of Article 50 has applied since 2 August 2026. The one exception is the marking duty in 50(2) for generative AI systems that were already on the market before that date. Their providers have until 2 December 2026 (Cooley, "EU AI Act: Transparency Obligations Take Effect 2 August 2026" (3 August 2026)). Content published before 2 August 2026 does not need to be labelled retroactively. The 2 December deadline covers this in detail.

The exceptions

The marking duty does not apply where a system only assists with standard editing, or does not substantially alter the input or its meaning.
For deep fakes that are part of an evidently artistic, creative, satirical or fictional work, disclosure is limited to saying the content exists, in a way that does not spoil the work.
AI-generated public-interest text needs no disclosure if it has had human review or editorial control, and a person or organisation holds editorial responsibility for it.
Several duties do not apply where the use is authorised by law to detect, prevent or investigate crime.

What "machine-readable" and "detectable" mean in practice

The Act does not name a technology. Article 50(2) asks that marking be effective, interoperable, robust and reliable, as far as technically feasible, given the kind of content, the cost and the state of the art. In engineering terms, that points to four properties.

Machine-readable: software, not just a person looking at the image, can tell the content was AI-generated. A visible caption alone is not enough.
Interoperable: tools other than yours can read the mark. That favours an open standard such as C2PA, whose Content Credentials any conforming reader can check (C2PA, the Coalition for Content Provenance and Authenticity).
Robust: the mark survives ordinary handling. Metadata alone is often stripped when a file is uploaded or re-encoded, which is why an invisible watermark in the pixels is usually paired with it. See C2PA and invisible watermarks.
Reliable: the mark is hard to forge. A C2PA manifest is cryptographically signed, so a changed manifest no longer validates.

No method is perfect, and the "technically feasible" wording accepts that. A determined attacker can remove both metadata and watermarks. What marking cannot do describes those limits for Tellmark's own marks.

The Code of Practice and the EU icons

The AI Office has published a voluntary Code of Practice on Transparency of AI-Generated Content. It includes a set of icons for labelling AI-generated content, which anyone may use without attribution (European Commission, EU icons for labelling AI-generated content). Cooley notes that signatories benefit from a degree of presumption that they meet the marking and detection duties. The European Commission also adopted guidelines on Article 50 on 20 July 2026.

Penalties

Cooley reports that breaching these obligations can bring fines of up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. The Act also applies to providers outside the EU whose AI outputs are used in the EU.

Where Tellmark fits

Tellmark provides the technical layer for 50(2) on images: a signed C2PA manifest stating the image was made or changed by AI, an invisible watermark that can find the record again if the manifest is stripped, and a free checker that anyone can use to detect them. For 50(1) it supplies a one-line chatbot disclosure. You stay the provider and stay responsible. Tellmark supplies technical measures, not a legal outcome.

Sources

Related guides

Describes technical measures. Not a legal assessment.