EU AI Act Watermarking: Article 50(2) Machine-Readable Marking Explained

“EU AI Act watermarking” is the everyday name for Article 50(2) of the AI Act (Regulation (EU) 2024/1689). It says that providers of AI systems, including general-purpose AI systems, “generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”

The duty sits with the provider of the system that generates the content. It is one of several transparency rules in Article 50. For the full picture, see Article 50 explained.

The Commission's Article 50 guidelines (C(2026) 7702 final, 5 November 2026, non-binding) read Article 50(2) as having two parts. The 20 July 2026 text (C(2026) 5054 final) was a draft submitted for approval and said the same. Outputs must be marked in a machine-readable format, and they must be detectable as AI-generated or manipulated. Doing only one of the two “will not suffice” (paras. 69–70).

“Machine-readable” means software can identify and extract the mark without a person stepping in (para. 71). A visible label is not ruled out, but it is a complement, not a replacement.

Detection is the second half. Providers must make the means of detection available to the people exposed to the content, and the results should be human-readable, in line with Article 50(5) (paras. 75–77).

Recital 133 lists techniques that may be used: “watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques, as may be appropriate”. Marking can be built in at the level of the AI system or the AI model, and a provider can use one technique or a combination (guidelines paras. 72–73).

Providers “are not required to record or keep a full provenance chain”, but they may use provenance methods (para. 73). They may also rely on a marking solution from an upstream model provider or a third party, such as an open standard; that doesn't remove their own responsibility (para. 74).

Article 50(2) says providers' technical solutions must be “effective, interoperable, robust and reliable as far as this is technically feasible”, taking into account the type of content, the costs of implementation and “the generally acknowledged state of the art”. Paragraph 79 of the guidelines defines each term:

Paragraph 81 adds that a provider is not obliged to use a solution that has not yet been developed or made available on the market, or that is technically unfeasible.

Article 50(2) does not apply “to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof”. It also has a law-enforcement exception. Whether a feature falls under the editing exception depends on what it actually does to the input; the wording is narrow.

Article 3(3) defines a provider as someone who develops an AI system, or has one developed, and places it on the market or puts it into service “under its own name or trademark, whether for payment or free of charge”. A deployer, under Article 3(4), is someone “using an AI system under its authority”; the guidelines describe this as deciding to deploy the system and how to use it and its outputs (para. 12).

Most website owners who use a third-party image generator, writing assistant or video tool are deployers of that tool. In that case the Article 50(2) marking duty belongs to the tool's provider.

The picture changes if you build a generative feature and offer it under your own name. The guidelines say a company that puts an image generator on the EU market under its own name is a provider, and so is a company that takes an existing generative system, modifies it and puts it into service under its own name (paras. 10–11). Providers outside the EU are covered if the output is used in the EU. For how the rules split, see Article 50 questions.

Breaches of Article 50 fall under Article 99(4): fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher; for SMEs and small mid-caps, whichever is lower (Article 99(6) and (6a)). See Article 50 fines.

Article 50(2) is about marks that software can read. It does not require a visible “AI-generated” label, and it does not remove any Article 50(4) duty. Paragraph 111 of the Commission's Article 50 guidelines, adopted 5 November 2026 (C(2026) 7702 final; non-binding), says the deployer labelling duties apply in addition to, and without prejudice to, the provider's marking duty. Visible or audible disclosure under Article 50(4) binds deployers in two cases: image, audio or video content that is a deep fake, and AI-generated or manipulated text published to inform the public on matters of public interest, unless a person has reviewed it and someone holds editorial responsibility. A deep fake is a cumulative test (guidelines paragraph 113): resemblance of an existing subject, and a false appearance of authenticity. Photorealism alone is not enough (paragraph 114). Under Article 50(5), that disclosure must be clear and distinguishable at the latest at the first exposure. Paragraph 117 says it should be perceivable without a special tool or a dedicated action, so a deployer cannot rely on the provider's machine-readable mark as the visible label.

For what that means on a page, see AI-generated content labelling. For wording you can adapt, see the AI image and video disclosure template.

Under Article 50(7), as replaced by the Digital Omnibus, the Commission encourages codes of practice on the detection, marking and labelling of AI-generated content. The Code of Practice on Transparency of AI-Generated Content was published on 10 June 2026 (guidelines, footnote 43). Providers may show compliance with Article 50(2) by adhering to a code assessed as adequate (paras. 84, 146–147); those who don't sign must show compliance “through other adequate means” (para. 148). The Omnibus recital 41 notes that such codes “do not grant a presumption of conformity”. More in the EU AI Act Code of Practice.

For images, audio and video, the code's Measure 1.1 asks providers for two layers of machine-readable marking: digitally signed metadata (Sub-measure 1.1.1), such as a C2PA manifest, and an imperceptible watermark (Sub-measure 1.1.2), with an exception for short text. The Commission found the code adequate in its opinion C(2026) 4839 of 8 July 2026.

The Article 50 scanner loads one public page in a real browser. Run a free scan to try it. Among other checks, it reads up to 8 of the page's main images and looks in their metadata for an IPTC digital source type or a C2PA (Content Credentials) manifest saying the image was AI-generated. It reports an image only if no label, caption, alt text or other text on the page says so (English wording). A missing marker is never a finding.

The scan cannot check whether your AI tool's provider marks its outputs: that is the provider's duty, and it can't be seen from a public page. The full scan report is €499, one-off; see pricing. The scan is not legal advice, and a clean result is not a compliance clearance.