The EU AI Act & Watermarks

What it means for you, the individual AI user

I'm sure you've all seen people talking about Claude's new "watermark," and you've probably also seen a lot of confusion about what that actually means. But this isn't a feature Anthropic dreamed up on its own — it's Claude's response to a legal requirement, one that applies to every AI provider doing business in the EU, not just Anthropic: Meta, OpenAI, Google, and more. The requirement comes from the EU AI Act, specifically Article 50, and it went into effect on August 2, 2026.

So before getting into what Claude actually does, it's worth understanding what "watermark" means here — because it's not what most people picture when they hear that word.

Nobody is stamping a little "AI" logo into the corner of your image or sprinkling visible tags through your text, or even a tag in metadata. What's actually happening is quieter, more technical, and — frankly — a lot more interesting.

It's not a watermark in the traditional sense we're all accustomed to

Before getting into what's actually happening, it helps to look at the word "watermark" itself — because it already means a few different things, and none of them quite match what's going on here.

Most people picture the visible kind: a translucent logo across a stock photo, or the corner tag some AI image generators add. Authors also know a quieter version — the invisible ID publishers embed in ebook files, often in the metadata, to trace a leaked copy back to its source. This is leading to the assumption causing most of the confusion: if it's invisible, people assume it must be sitting in the metadata somewhere — a tag that gets wiped the moment you copy the text into a new document, convert the file, or strip the metadata out.

That assumption is wrong, at least for text. The new text watermarks aren't a metadata tag at all — they're woven directly into the word choices and patterns of the output itself. Copying the text, retyping it, or converting the file doesn't touch it, because there's no separate tag to strip; the mark is part of the content.

What the law actually says

The requirement comes from a law called the EU AI Act — specifically a piece of it called Article 50, which went into effect on August 2, 2026.

Stripped of the legal phrasing, it says this: if an AI system generates text, it has to be marked in a way that a machine can detect as AI-generated. Not a visible tag a person reads — a machine-readable signal a detection tool can pick up on.

That second part matters, because the law is actually pretty clear that a visible label by itself doesn't count. So the little "AI-generated" disclaimers you might have seen tacked onto some content? Those don't satisfy this requirement on their own. The law wants something a computer can verify, not just something a human can read.

The same law covers images, video, and audio too — the marking just tends to work differently there, usually through metadata attached to the file rather than something baked into the content itself the way text works. That's a bigger topic on its own, so for the rest of this post, we're sticking with text — the part causing the most confusion.

This applies to every AI provider offering these tools in the EU — it's not optional,. Technically, the law only requires this for content reaching EU users. But most AI providers are applying it worldwide anyway — because geofencing by country is a pain, easy to dodge with a VPN, and building two separate versions of a product is way more trouble than it's worth.

So what is the watermark, actually?

The pattern isn't added after the fact — it's built into the writing process itself. When AI generates text, every word it chooses already carries a subtle, consistent lean baked in at the moment it's written. The choosing and the marking happen as the same act, not as two separate steps.

It doesn't change what the text says, and it doesn't make the writing worse. You're not reading "watermarked" sentences that look different from normal ones — the quality is the same either way. The pattern is just there, underneath, invisible to a reader, but there for a tool built to look for it.

What happens to that pattern depends on what you do with the text afterward. If you leave the AI's wording as written, the pattern stays intact. If you edit or rewrite a sentence yourself, you're replacing the AI's word choice with your own — and your choice doesn't carry that lean. So editing doesn't erase an existing mark; it just swaps out marked AI text for unmarked human text wherever you changed something. Lightly edited AI text still holds plenty of pattern in the untouched parts. Heavily rewritten text holds less and less, because more of the original wording has been replaced.

This is built into Claude at the model level, so it applies wherever Claude generates text: the chat app, the API, Claude Code, Cowork, all of it, worldwide. Newer models have it built in from the start; older ones are being updated to include it.

AI watermark "removal" tools

You may have also heard about tools claiming to "remove" AI watermarks. Worth being clear about this: those tools work by stripping metadata — deleting the hidden tags some files carry. That approach does nothing here, because there's no metadata tag to strip in the first place. The pattern is woven into the actual words, not attached alongside them as a separate piece of data.

The only way to genuinely eliminate the pattern is to substantially rewrite the text yourself — which isn't really "removing a watermark" so much as it is replacing the AI's word choices with your own, the same editing effect already covered above. There's no tool that scans a finished piece of AI text and quietly erases the pattern while leaving the wording untouched.

What an AI watermark does (and doesn't) prove

Worth remembering: the detection tools themselves aren't broadly available yet, as covered above. But even once they are, a couple of things are worth knowing so the watermark doesn't get treated as more definitive than it actually is.

No detectable watermark doesn't mean the text wasn't AI-generated. It might have come from a model that predates marking, or it might have been heavily rewritten afterward, in which case the pattern's mostly gone for the reasons already covered above. Either way, "nothing detected" isn't proof of human authorship — just an absence of a signal.

And on the flip side, finding a watermark doesn't confirm the whole piece was AI-written. Someone could paste a single AI-generated paragraph into an otherwise entirely human-written document and that paragraph could carry the watermark — and since detection tools for this aren't public yet, it's not even clear whether a checker would isolate that one paragraph or just flag the document as a whole.

Either way, a watermark tells you AI was involved somewhere, not necessarily that AI wrote everything. So the watermark is a data point, not a verdict one way or the other.

The Bottom Line... how it affects you

None of this is a stamp saying "AI wrote this" slapped visibly across your text. It's a quiet, statistical signal woven into the writing itself — invisible to read, with no detection tool publicly available yet, and telling you at most that AI was involved somewhere, not the full story of how a piece came together.

And if you use AI as part of your writing process — whether you're working on a novel, a blog post, a news piece, or just an email at work — this watermark isn't a label slapped on your work. Current law does not require you personally to disclose anything about your process.

The watermark is a compliance requirement on the companies building these tools, not on the people using them.

Sources:

• How Claude Marks AI-Generated Content — From Claude's Help Desk

• High-Level Summary of the AI Act — A plain language overview

• EU Artificial Intelligence Act, Article 50: Transparency — If you want to delve into the Act itself with all its legalese