Starting in August 2026, the text Claude writes for you will carry an invisible mark you can’t see, can’t remove by copy-pasting, and that Anthropic says “may persist through some editing.”
That’s the practical upshot of Anthropic signing the EU AI Act Code of Practice on transparency for AI-generated content. New Claude models will embed watermarks in text and attach signed provenance metadata to files.
The EU wrote the rule, but the border doesn’t hold it
Claude models that launch in the EU on or after August 2, 2026, ship with the labeling baked in. The interesting part is what happens outside the EU: nothing different. Anthropic is applying it globally, across the API, Claude, Claude Code, Claude Cowork and Claude Tag.
Generated text gets embedded watermarks. Generated files get digitally signed provenance metadata. Existing models fall under a transition period in the law, but Anthropic says it’s already working on retrofitting them.
Verification tools are coming for users and third parties to check the labels. Anthropic hasn’t said when. That’s a meaningful gap: a mark nobody can read yet is a mark that does nothing yet.
Developers building Claude into their own products have to work out for themselves which Article 50 requirements apply to what they’ve shipped, according to Anthropic.
Two different labels doing two different jobs
The text watermark is invisible and, Anthropic says, doesn’t affect the meaning, quality or readability of the output. It survives copy and paste. It’s applied at the model level, so the Claude product you’re using is irrelevant to whether it’s there.
Files are handled differently. Supported formats, including .svg, .png and .jpg images, get signed provenance metadata built on C2PA, the open standard from the Coalition for Content Provenance and Authenticity. The signature says Claude touched the file and can expose later tampering.
Text watermarks should also come through cloud partners including AWS, Google Cloud and Microsoft Foundry. Signed metadata might not, since those platforms may not support it.
Anthropic is unusually blunt about what this proves
A detected watermark does not mean Claude wrote the content. That’s Anthropic’s own caveat, not a critic’s. People run Claude for proofreading, translation and summarizing all day, which means a human’s ideas can come out the other end wearing a machine’s label.
The absence of a watermark tells you even less. The model may predate watermarking. The text may have been heavily edited or translated. The passage may be too short to detect reliably. The metadata may have been stripped by a format conversion, or by someone taking a screenshot.
So you have a signal that can be a false positive when a person did the thinking, and a false negative for at least five ordinary reasons. Read the fine print before you treat a detection result as evidence of anything.
The one thing worth testing
Everything hinges on how these marks hold up under editing, reformatting and translation. Those are exactly the three things a student or a spammer does by reflex.
If the watermark survives them, checking for a known mark beats what’s on the market now. Detectors like Pangram use proprietary methods that don’t tell you what triggered a result. A watermark you’re looking for on purpose is a different kind of claim than a probability score with no receipts. Third-party detectors could add support for Anthropic’s mark and get a firmer signal out of it.
Google shipped this already, and OpenAI is sitting on something better
Google Deepmind open-sourced SynthID and built it into the Gemini models. It nudges probability values during token prediction to leave a watermark without degrading text quality, and it works across languages. It struggles with text edited after generation, which is the same weak point Anthropic will have to answer for.
OpenAI’s position is the strange one. The company has had a text detector with 99.9 percent accuracy for roughly two years and hasn’t shipped it. The stated reasons: users beat it easily through translation or rewriting, and it risks stigmatizing certain groups. The unstated one, most likely, is that a public detector would be bad for OpenAI’s business.
Why the accuracy caveats matter more than the standard
In education, an unreliable detector isn’t a nuisance. It’s a false cheating allegation attached to a real person’s record.
And there are real reasons to want to know how much AI went into a piece of work. Studies show heavy reliance on these tools can weaken critical thinking and writing skills, especially among students using them as a shortcut instead of a learning aid. It’s not confined to campuses either: scammers are enrolling fake students at US colleges, running the coursework through AI and collecting financial aid.
The customers most likely to notice
Claude has a following in knowledge work and among school and college students, partly because even the older models write prose that doesn’t read like a machine wrote it. That’s the exact user base a working detector makes uncomfortable.
Schools and universities are already fighting over AI in academic work. Anthropic just volunteered to make its own output easier to flag, globally, ahead of any requirement to do so outside the EU. Watch whether the verification tools arrive before August 2026 or after. Until they do, every Claude output after that date carries a mark that only Anthropic can read.
