EU AI Act Article 50, explained
Article 50 is the transparency article of the EU AI Act. From August 2, 2026, providers of generative AI systems must ensure their output is marked in a machine-readable format and detectable as artificially generated or manipulated, and deployers must disclose deepfakes and AI-generated text published to inform the public on matters of public interest. The Commission's Code of Practice on Transparency of AI-Generated Content, final on June 10, 2026 and judged adequate on July 8, 2026, turns that into a two-layer commitment: secured metadata such as C2PA plus an imperceptible watermark. The 2026 watermark launches from Anthropic and OpenAI are responses to this article. It binds providers and deployers, not individual users.
The provider duty
Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and detectable as generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities of content types, costs and the state of the art. That last clause matters: the law does not demand a perfect mark, only a serious one. It is why providers publish their failure numbers rather than hiding them, and why OpenAI's announcement reads as a compliance document as much as a product note.
The deployer duty
Article 50(4) covers people and organisations that use AI systems. Deployers who generate or manipulate deepfakes must disclose that the content was artificially generated. Deployers who publish AI-generated text to inform the public on matters of public interest must disclose it, unless the text underwent human review or editorial control and a natural or legal person holds editorial responsibility. For a newsroom or a public body, that exception is the operative rule: a reviewed, bylined article is not a labelled one.
What the Code of Practice adds
The Code is the Commission's voluntary route to compliance. Section one, for providers, sets the two-layer approach: secured metadata that records origin and is cryptographically signed, plus an imperceptible watermark that survives when metadata is stripped. Signatories commit to keeping marks non-removable to the best of their abilities, to making detection available to appropriate parties, and to working toward interoperable methods. Section two, for deployers, covers labelling of deepfakes and published text. The Commission's July 8, 2026 opinion found the Code adequate for Article 50(2), (4) and (5), which gives signatories a presumption of conformity. In practice that presumption is why the large providers aligned their launches with the Code rather than inventing their own approach.
Dates
Article 50 applies from August 2, 2026. Systems placed on the market before that date benefit from a transition period, which is why Anthropic said models released before August 2, 2026 would gain the mark over the coming months rather than on day one. The final Code was published June 10, 2026 and found adequate July 8, 2026. National market surveillance authorities enforce the article, with penalties in the Act's general tier.
What Article 50 does not do
It does not prohibit a person from editing, paraphrasing or translating AI text. It does not require platforms to preserve marks; that idea appears in the Code's provider commitments and, in the United States, in California's platform duty from 2027. It does not give the public a detector; access is left to providers, who have so far limited it to approved organisations. And it does not define a pass mark for detection confidence, which is why the Code's language about appropriate parties and the state of the art will carry the weight in any dispute.
Why it matters outside Europe
Because Anthropic could not scope its watermark by region, Claude's mark applies everywhere. OpenAI chose an EU-only default for ChatGPT and Codex but opened opt-in watermarking to API customers worldwide. A single EU article therefore changed what a writer in Ohio or Osaka receives from one provider and not another. The model-by-model guide tracks the current state.
Questions people ask
When does Article 50 apply?
From August 2, 2026, with a transition period for systems placed on the market before that date.
Does Article 50 require a watermark specifically?
It requires machine-readable, detectable marking. The Code of Practice interprets that as secured metadata plus an imperceptible watermark.
Does Article 50 apply to users who paraphrase AI text?
No. The duties fall on providers of generative systems and on deployers who publish deepfakes or public-interest AI text.
Is the Code of Practice mandatory?
No. It is voluntary, but signatories gain a presumption of conformity with Article 50.