Work your team drafts with Claude now carries a hidden digital mark
Anthropic has extended its invisible watermark to older Claude models to meet EU rules, and it applies worldwide. A mark can suggest AI use but can't prove who wrote the work, so check your client contracts now.
The short version
- Anthropic now adds a hidden watermark to text from its Claude AI models, in every country where Claude is offered.
- You can't see the mark, and it stays with the text when you copy and paste. Anthropic did this to meet European Union rules on labeling AI content.
- Today only approved groups can check text for the mark. Update your AI policy and client contracts now, assuming AI use may be traced later.
What changed
Anthropic's help page now lists which Claude models mark their output. Every model on the list carries the text watermark on Anthropic's own services. The newer ones also carry it when you use Claude through the big cloud providers.
The change comes from the EU AI Act. Anthropic has signed the law's Code of Practice on labeling AI-generated content. Claude models launched in the EU from August 2, 2026 support marking from day one, and older models are now being added.
The Decoder, a trade publication, reports that Anthropic told customers about the latest step by email and tied it to a December deadline. It also reports that customers can't turn the mark off. There is one gap, it says: on 2 older models, developers can avoid the mark by changing a technical setting.
How the mark works
For text, Anthropic says it weaves a mark into the words themselves. It says the mark doesn't change the meaning or the quality. The Decoder describes it as a statistical pattern in word choice, with no hidden characters added.
The mark is part of the text. It travels when you copy and paste, and it may survive some editing.
Some writers dispute the claim about quality. They argue that a word picked to carry a mark isn't always the best word.
Files work differently. When Claude creates a supported picture file, it attaches a signed label called a Content Credential. That label follows an open industry standard, and Anthropic offers a free tool to check for it.
What the mark can't tell you
Anthropic is plain about the limits. A detected mark means Claude may have handled the text. People also use Claude to proofread or translate, so the ideas may be entirely human. The text may also have changed after Claude touched it.
A missing mark proves nothing either. Text may carry no detectable mark if it is very short, or if it has been heavily edited, paraphrased or translated.
OpenAI's published tests of its own watermark show how fragile these marks can be. Replacing 10% of the words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%. The Decoder notes that Anthropic hasn't published detection rates for Claude.
The 2 companies also apply their marks differently. OpenAI's is on by default for ChatGPT and Codex users in the EU only, and developers elsewhere must opt in. So you can't assume all AI text is marked.
What this means for a business in Indonesia
The law is European, but Anthropic's mark isn't limited to Europe. If your team drafts with Claude here, the mark goes with the draft. That could include proposals and contract clauses.
For now, only approved groups can check text for the mark. They include regulators, media, fact-checkers and researchers, plus enterprises with their own duties under the law. Anthropic plans to widen access. A client probably can't test your work today, but one might be able to later.
Artificial Lawyer, a legal trade publication, has flagged where this could bite. A client that bans AI on its work could discover you used it, even when the document is correct. A client pushing for lower fees on AI-assisted work could look for the same evidence. Marks can also carry over into reused templates.
Those concerns apply beyond law firms. Any business that sells written work could face the same questions.
If you build your own product on Claude, Anthropic says you should assess your own duties under the EU rule separately.
What to do next
- Write down where your teams use AI on work that leaves the building.
- Read your client and agency partner contracts for terms on AI use, and fix any mismatch with what your teams really do.
- Never tell a client or regulator that work had no AI help unless you have checked.
- Don't punish a staff member or drop a supplier on a detector result alone. It is only a signal.
- If you sell into the EU or ship AI features to users there, ask your lawyers about the EU's labeling rule now.
- Ask your other AI vendors what they mark, and where.