Can you remove Claude's watermark?
Not with any of the tools currently marketed for it. Those strip invisible Unicode characters, which is not what Claude's text watermark is made of[1][8] — Anthropic says nothing is added to the text and there are no hidden characters[10]. The mark lives in the word choices themselves, so the only thing that removes it is changing the words. Anthropic's own statement of the threshold: light editing probably will not remove it completely, a complete rewrite replacing every word will[10]. Anything between those two is untested[5].
This page explains the mechanics honestly. We do not sell a remover, and we would be lying if we did, because the only detector that could test one is Anthropic's and it is not available to us[1].
What does nothing
| Method | Effect on Claude's text watermark |
|---|---|
| Stripping zero-width characters (U+200B, U+200C, U+200D, U+FEFF) | None The mark is not made of characters |
| Replacing em dashes with hyphens | None Punctuation is a style tell, not a watermark |
| Changing fonts, or pasting through Notepad | None Affects formatting only |
| Retyping the text by hand, unchanged | None Same words, same mark |
The first two are what the tools ranking for this query actually do. One of them describes its own method as scanning for zero-width spaces, zero-width joiners and other invisible Unicode characters, and replacing em dashes[8]. That is a genuine feature; it is just unrelated to what it is being sold for.
What plausibly works
A statistical watermark is a pattern across many token choices. Break enough of those choices and the pattern falls below the detection threshold. So:
- Substantial rewriting in your own words. Not synonym swapping — genuinely restructured sentences.
- Translation out and back, through a different model. Reported as something the mark may not survive[6]. Note the direction of travel: a translation Claude produces carries a watermark, because Claude chooses every word of it[10]. Translating with Claude adds a mark; translating away from Claude is what might shed one.
- Retelling the content from scratch after reading it. At that point you have written it, which rather resolves the question.
Anthropic itself acknowledges the mark may not survive heavy editing, format conversion, translation or screenshots as reported[6], and says it "may persist through some editing"[1]. There is a second reason removal may be beside the point: the watermark is already sparse wherever the wording was forced. Factual passages, proofreading and code outside the comments carry little or none of it[10].
The only detector for Claude's text watermark is Anthropic's, and since 1 September 2026 it has been in private preview for organisations obliged to check under EU law[1]. No vendor selling removal has it, so none can demonstrate that its tool worked, and you cannot verify it either. You would be paying for an unfalsifiable claim. If your reason for wanting this is an accusation, that page is more use, because the strongest thing you can cite is Anthropic's own statement that a mark is not authorship[1].
What about the C2PA metadata on images?
That one genuinely is removable, and trivially. C2PA provenance data is metadata attached to the file[1]; re-encoding an image, screenshotting it, or running it through most social platforms discards it. That is a known limitation of metadata-based provenance rather than a trick. More on C2PA.
Sources
Last verified against primary sources: