Is Claude's Watermark Detectable?
Short answer: not by anyone outside Anthropic. Here is why, and what you can actually do.
Two different things get called "watermark"
When people say "Claude watermark," they usually mean one of two things. The first is a statistical token-level watermark: during generation the model subtly biases which words it picks, following a secret key. The bias is invisible to a reader but can be measured statistically, if you hold the key. The second is hidden characters: zero-width spaces, joiners, and similar invisible Unicode that can be inserted into text as a tracking mark. These are completely different, and only one of them is something a public tool can find.
Why the statistical watermark is not third-party detectable
Detecting a distributional watermark requires the same secret partition (the "green list") used to embed it. Without that key, the bias is indistinguishable from ordinary word-choice variation. Published research (Kirchenbauer et al. and later work) shows the detector needs the key; there is no reliable key-free detector. So any tool claiming a confident "92% watermark detected" score on arbitrary text is guessing. We would rather tell you that than invent a number.
What you can actually detect
You can reliably detect hidden Unicode characters. Our free scanner does exactly this: paste text and it reports any zero-width or invisible characters by name and count. If it finds them, you can strip them. If it finds none, your text is clean of invisible markers, though that does not rule out a statistical watermark, which, again, no one outside the model owner can read.
What about "removing" the statistical watermark?
The only established way to reduce a distributional watermark is to rewrite the text so its word choices change, paraphrasing with a different model. That reduces the signal, but it cannot guarantee a text will pass every detector, and you should not use it to misrepresent authorship where disclosure is expected. Our rewrite tool does this honestly, with that caveat stated on the tool itself.