The Claude watermark: how it works and how to remove it

In August 2026 Anthropic said that Claude models released after 2 August 2026 write text with a watermark: a way to estimate how likely it is that Claude wrote a text. From that date the EU AI Act requires AI providers serving the EU to mark what their systems generate in a way machines can detect, and Anthropic says several other major providers are doing the same.

From Anthropic's explainer:

• It applies worldwide, not only in the EU. Anthropic says it doesn't yet have a durable way to limit it by region.
• Models released before 2 August 2026 fall under a transition period in the EU law, and Anthropic is working to add the watermark to them too.
• It carries no identifying information and can't be traced to a person, an organisation or a chat.
• Text Claude translates is watermarked as well, because Claude chooses every word of the translation.

When Claude writes, there are usually several good choices for the next word. The watermark uses a secret key to give some of those choices a slight edge. Any one choice still looks natural, but over a few hundred words the favoured choices add up to a pattern that a detector holding the key can measure. Without the key, watermarked text looks like any other.

Because the pattern is in the words themselves, it survives copying, pasting, retyping and reformatting. It's also why short texts are hard to judge: Anthropic notes that detection doesn't work well on small samples, where there are fewer word choices to go on.

The method is Anthropic's version of SynthID-Text, which Google DeepMind published in Nature in 2024. More on that in SynthID-Text, Google's watermark for AI text.

• Deleting hidden charactersStays

  The watermark isn't a character, so there's nothing to delete.

• Fixing a few wordsStays

  Anthropic says light editing probably won't remove it completely.

• Asking Claude to rewrite or translate itStays

  The new text is Claude's own choice of words, so it carries a watermark of its own.

• Retyping it by handStays

  The same words make the same pattern.

• Rewording much of it yourselfWeakened

  Each word you change takes some of the pattern with it, but what's left may still be enough to detect.

• A complete rewrite by a model that adds no watermarkRemoved

  Anthropic says a complete rewrite where every word is replaced will remove it.

Only Anthropic can read the watermark, because only Anthropic holds the key. It offers a detection API, in private preview, to the kinds of organisations EU law gives that role: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups.

There's no public Claude watermark checker. A site that claims to detect the watermark doesn't have Anthropic's key, so it's checking something else, usually hidden characters or writing style. More on that in Can you check text for an AI watermark?

1. Paste the text, or upload a .txt or .md file, on the glyphwash home page.
2. Choose Standard, which rewords sentence by sentence and keeps your structure, or Deep, which rebuilds each passage from its facts.
3. glyphwash takes out hidden characters, rewrites every passage with an open-weight model that adds no watermark of its own, and checks that the meaning, names and numbers came through.
4. Read any passage it flags for review, then copy or download the result.

Nobody outside Anthropic can test for the watermark, glyphwash included. So instead of a pass or a fail, each result shows how much of the original wording survived, since that's where the pattern lives, and how well each passage kept its meaning.

Inside Claude

If you write with Claude, add glyphwash as a connector at https://glyphwash.com/mcp, or in Claude Code run claude mcp add --transport http --scope user glyphwash https://glyphwash.com/mcp. Then ask Claude to wash what it has just written, and use the washed text exactly as it comes back: any rewording by Claude can bring the watermark back. Setting it up in Claude and other AI apps.