Marking AI Content Under the EU AI Act: 'Machine-Readable' Explained

Q: What does "machine-readable marking" mean?

There are two layers, and they are not the same. A visible label for people. A deepfake or an AI text you show to the public has to be recognisable as AI. Think of a caption or a badge the reader simply sees. A machine-readable marking inside the file. Separately, the content has to carry a technical marker that software can read: metadata or a watermark that signals "this was created or edited by AI". The marking obligation is mainly about that second layer. A person doesn't have to see the marker; a platform, browser or supervisory authority does need to be able to check it without having to ask you anything.

Q: Which content falls under the AI marking obligation?

Article 50 targets synthetic output: content that an AI system creates or significantly edits and that is meant for the public. Text you publish to inform the public. Images, from product photos to illustrations. Audio, such as generated voices or sound clips. Video and deepfakes, where a person or thing is imitated realistically. The chatbot notice is a separate obligation under the same article. If an AI assistant runs on your site, the duty to disclose it at first contact applies on top of this. We cover that in a separate article on whether your chatbot has to say it's AI.

July 12, 2026

"Machine-readable marking" means AI content carries a label that not only a person can see but a computer can also read. Under article 50 of the EU AI Act, AI-generated or significantly edited text, images and audio must carry such a marking from 2 August 2026. Generative systems that were already on the market before that date get a grace period until 2 December 2026.

What does "machine-readable marking" mean?

There are two layers, and they are not the same.

A visible label for people. A deepfake or an AI text you show to the public has to be recognisable as AI. Think of a caption or a badge the reader simply sees.
A machine-readable marking inside the file. Separately, the content has to carry a technical marker that software can read: metadata or a watermark that signals "this was created or edited by AI".

The marking obligation is mainly about that second layer. A person doesn't have to see the marker; a platform, browser or supervisory authority does need to be able to check it without having to ask you anything.

Which content falls under the AI marking obligation?

Article 50 targets synthetic output: content that an AI system creates or significantly edits and that is meant for the public.

Text you publish to inform the public.
Images, from product photos to illustrations.
Audio, such as generated voices or sound clips.
Video and deepfakes, where a person or thing is imitated realistically.

The chatbot notice is a separate obligation under the same article. If an AI assistant runs on your site, the duty to disclose it at first contact applies on top of this. We cover that in a separate article on whether your chatbot has to say it's AI.

Machine-readable in practice: C2PA and metadata

The marking is applied at the point of creation, by the tool that generates the content. The best-known standard is C2PA (Content Credentials): a piece of signed provenance data that travels with the file and records how it was made. Alongside it there are metadata fields and invisible watermarks that serve the same purpose.

One caveat: the marking stays reliable only as long as the metadata survives the journey. A screenshot, a re-export or heavy compression can wipe the marker. So it isn't a matter of switching it on once, but of building your whole content chain around it.

The grace period until 2 December 2026 and the rest of the timeline

Since 2 February 2025: AI literacy (article 4). Staff who work with AI have to understand enough about it. This applies already.
From 2 August 2026: the marking obligation and the chatbot notice.
Until 2 December 2026: a grace period for generative systems that were already running before 2 August 2026. You get slightly more room for existing content flows, not for new ones.

Unlike the European Accessibility Act, there is no exemption for small businesses. The obligation applies regardless of company size. Breaches carry fines of up to 15 million euro or 3% of worldwide annual turnover. In each EU country, a national supervisory authority enforces this.

What to ask your tooling

Concretely, with every supplier that produces AI content for you:

Do you apply machine-readable marking? Ask explicitly about C2PA or metadata on the output, and whether it's on by default.
Does the marking survive export? Test whether the marker stays put after you place the image or text into your CMS or online store.
Where do I record this? Keep track of which tools you use, for what, and what marking they provide. That overview will be your evidence later.

And on what's actually machine-detectable: no one can reliably prove after the fact that a given text was made by AI. That's precisely why the law places the obligation on marking during creation, not on detection afterwards. Handling it at the source is the only approach that fully holds up.

Sort out the marking and the evidence in one go

Seviranta's AI transparency check delivers the notice text for your chatbot and an AI register that records which AI you use and where, as evidence for a supervisory authority. As part of this we detect which chat widget runs on your site; whether it's AI-driven is something you confirm yourself. An automatic test of whether your notice and marking are genuinely switched on is the announced next step. For 29 euro per month or 290 euro per year, excl. VAT, as part of your Seviranta plan or on its own.

Want to know where your site stands first? Scan your site for free.

Sources: Article 50, EU AI Act · AI Act implementation timeline