EU AI Act Article 50: what it requires
From 2 August 2026, providers of generative AI must mark their outputs in a machine-readable format, detectable as artificially generated or manipulated[7]. This is the reason Claude's watermark exists, and the reason it arrived on that specific date[1]. Systems already on the market before that date have until 2 December 2026, and only for the marking and detection obligation[7].
What does Article 50 require?
Article 50 sets transparency duties on providers and deployers of certain AI systems. The part that produced this month's news is the marking requirement: output must be machine-readable and detectable as artificially generated or manipulated[7].
What does the Code of Practice add?
Alongside the law there is a Code of Practice on Transparency of AI-generated Content, published 10 June 2026 after drafting began in November 2025, with a compliance deadline of 2 August 2026[2]. It is voluntary — signing it is optional, complying with Article 50 is not — and the Commission and AI Board have confirmed it as an adequate voluntary tool for demonstrating compliance[2]. Around 190 companies and organisations had signed by late July 2026[2].
Its Section 1 requires that provider outputs be marked in a machine-readable format and detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust and reliable as far as technically feasible[2].
"Detectable" is doing a lot of work, and on 1 September 2026 Anthropic showed how it reads the word. It moved from promising forthcoming detection documentation to running a detection API in private preview, and scoped access to the bodies the law obliges to check: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations, EU civil society groups, and enterprises with the same obligation[1][10]. That is a defensible reading of a rule written for those bodies. It also means "detectable" has been satisfied without the public being able to detect anything, which is a different thing from what most readers assume the word promises. We are not going to pretend to settle whether it is enough.
Why does it reach outside the EU?
It largely does not have to, and providers are doing it anyway. Anthropic applies marking wherever Claude is offered, worldwide[1], and has now given the reason: it does not yet have a durable way to scope watermarking by region, and says it will keep evaluating approaches[10]. That is the familiar pattern by which an EU rule becomes the global default — the same one GDPR produced with cookie banners — but note it is described as a current technical limit rather than a permanent policy.
Who has actually complied, and how.
Sources
Last verified against primary sources: