Remove
Remove Claude Watermark Text
Remove Claude watermark text and you are attacking a statistical pattern in word choice, not a hidden character you can delete.
That distinction decides which methods work. Almost every tool marketed for this job strips invisible Unicode, which the Claude text watermark does not use.
Only substantial rewriting weakens the text watermark. Character scrubbing does nothing to it.
Runs in your browser. No signup, no upload, nothing stored.
What the mark actually is
Anthropic describes Claude as weaving an imperceptible watermark directly into the text itself, one that travels with the text when it is copied and pasted.
Technical reporting describes the mechanism as biased token sampling. At each step where several words would serve equally well, a pseudo-random rule seeded on a private key nudges the model toward one group of candidates over another.
There is no character to find. The signal is the distribution of the words you can already see, measured across the whole passage.
Why character scrubbers do nothing
A zero-width space is a character. You can search for it, delete it, and prove it is gone with a single regular expression.
That property is exactly why it would be useless as a provenance mark, and it is why Anthropic did not use one. A watermark that any user can strip in a second provides no provenance at all.
Tools advertising removal of zero-width joiners and invisible spaces are cleaning something real. It is just not this watermark, and running one changes nothing about your exposure.
What Anthropic says weakens the signal
The documentation lists the conditions under which a mark may not be detected. Read from the other direction, that list is the removal manual:
- Content that has been heavily edited
- Content that has been paraphrased
- Content that has been translated
- Content mixed with substantial human-written material
- Passages too short for a reliable statistical test
- Output from models that predate the rollout
The honest ceiling on any claim here
Anthropic has not released a public detector, and has not said when it will.
Nobody — including every site selling a remover, and including this one — can currently measure whether a given passage still carries the mark. Any tool claiming verified removal is claiming something it has no way to test.
Treat rewriting as a reduction in probability, not as clearing a known threshold. That is an unsatisfying answer, and it is the accurate one.
The practical sequence
Rewrite the passage substantially in your own voice, working paragraph by paragraph rather than swapping individual words. Larger units give you room to restructure while holding the meaning.
Blend the result with material you wrote yourself. Mixed provenance appears on Anthropic's own list of things that weaken detection, and it is the cheapest item on it.
Then read the rewrite against the original and repair any drift. Automated paraphrasing routinely loses a qualifier or inverts a hedge, and those errors become yours the moment you publish.
Common questions
- Does deleting invisible characters remove the Claude text watermark?
- No. The text watermark is a statistical pattern in token selection, not a hidden character. Character scrubbing removes formatting artefacts, which is useful for other reasons, but it leaves the watermark untouched.
- Does light editing remove it?
- Anthropic says the mark may persist through some editing. Light edits are unlikely to be enough; heavy rewriting, paraphrasing or translation are the conditions Anthropic itself names.
Source
The claims on this page are drawn from primary documentation and reporting rather than from other tools’ marketing copy.
Anthropic: How Claude marks AI-generated content →Related
Check what your text and files actually carry
Runs in your browser. No signup, no upload, nothing stored.