EU AI labelling law: Article 50 of the AI Act explained

Article 50 of the AI Act (Regulation (EU) 2024/1689) sets the “transparency obligations for providers and deployers of certain AI systems”. For images, three of its seven paragraphs matter: 50(2) for the makers of AI generators, 50(4) for the professionals who use them, and 50(5) for how to tell people.

A deep fake (Article 3(60)) is AI-generated or manipulated image, audio or video content “that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”. Our Deepfake label guide applies that definition case by case.

• Deployers label. A deployer uses an AI system under its own authority, professionally: a brand, an agency, a newsroom, a shop, a freelancer. If you use AI to make a deep fake and publish it, the visible disclosure is your job, not the AI tool's. Employees aren't separate deployers; their employer is. Companies outside the EU are covered when they intend their deep fakes to be seen in the EU, posting them on the open internet included (Guidelines, paras 13 and 14).
• Providers mark. The makers of generative AI must embed machine-readable marks in the output, such as signed metadata and invisible watermarks: see Detect AI-generated images.

Purely personal use is excluded (Article 2(10)): deep fakes of your household, or ones you post on social media in a personal capacity. But earning money from it on a regular basis, or any business, trade or freelance work, counts as professional (para 19). Platforms that only pass other people's content on aren't deployers either (para 16); what they show is in AI labels on social media.

Not every AI image needs a label under the law: only deep fakes do. The Guidelines (para 113) read four conditions into the definition, all of them required:

1. Resemblance: highly similar, not necessarily identical.
2. Existing: something that exists or could plausibly exist. Dragons and elephants driving cars are out.
3. Persons, objects, places, entities or events, avatars, buildings, products and animals included.
4. Would falsely appear authentic or truthful to a person.

No intent to deceive is needed, and the foreseeable audience counts, children included (paras 114 and 115). Photorealism makes a deep fake likelier but doesn't decide alone.

Minor edits usually aren't deep fakes (para 116): removing passers-by, colour correction, noise reduction, aesthetic background changes in a product advert. Evidently artistic, creative, satirical or fictional works still need a disclosure, one that doesn't hamper the work, such as in the credits; “evidently” is read strictly (para 122). Public-interest text made with AI must be disclosed too, unless it was reviewed under someone's editorial responsibility. Outside the law, labelling stays good practice, and platforms may require it.

Article 50 has applied since 2 August 2026. The Digital Omnibus on AI didn't postpone it for deployers.

Nothing retroactive: deep fakes made before 2 August 2026 needn't be labelled (para 154), though the Commission encourages it. Older AI text published from that date on must be.

The AI Act sets no design, size or corner, only a standard (Article 50(5)), which the Guidelines explain:

• Clear and distinguishable: noticeable and separate from its surroundings, not hidden in menus or terms of use (para 142).
• At first exposure: for every person who sees it, including while scrolling a feed (para 143).
• Visible without tools: the provider's machine-readable marks don't count for the deployer (para 117).
• Accessible: the Commission's icons page suggests alt text or ARIA labels saying the content is AI-generated.

The Code of Practice adds detail for its signatories: the EU icon or an equivalent label, placed where no other overlay intervenes, “e.g., in the top right corner of an image or video deep fake”; embedded in the content; visible against any background; resized only in proportion, with “AI” in English. In user testing, versions with a word (“generated”, “modified”) did significantly better. The icons are optional and free, and using them doesn't establish compliance by itself: see EU AI icons.

Breaching Article 50 can cost up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)); for SMEs, start-ups and small mid-caps, whichever is lower. EU institutions face up to EUR 750,000. Each country sets its penalties, warnings included, and authorities may treat commitments under the Code as a mitigating factor. The often-quoted EUR 35 million or 7% is for prohibited AI practices, not for labelling.

National market surveillance authorities enforce, on their own initiative or after a complaint, which anyone with grounds can file (Article 85). The Commission's list of 7 September 2026 doesn't say which one handles Article 50; for France, Germany, Italy and Spain, it names:

Two Commission documents fill in Article 50. Neither is binding.

• The Code of Practice on Transparency of AI-Generated Content (final on 10 June 2026) is voluntary: Section 1 for providers' marking and detection, Section 2 for deployers' labelling, Annex 1 for the icons. The Commission (8 July 2026) and the AI Board (9 July) found it adequate, and about 190 organisations had signed by the end of July. Non-signatories must show compliance another way, for example with a gap analysis against the Code (para 148).
• The Guidelines on Article 50 (approved on 20 July 2026) explain the definitions with examples. Formal adoption follows once all language versions exist; in mid-October 2026, only the English version was listed.

In English, on the EU's own websites. Only the texts in the Official Journal are authentic.