Every Claude watermark checker, compared by what it actually claims

The short version

If you want to find copy-paste artifacts — the HTML and invisible characters that actually prove text came out of a chat interface — any of these tools will do it, including ours, and the free ones do it fine.

If you want to know whether Anthropic's watermark is present — nobody can tell you. Not us, not the paid tools. The key is Anthropic's and has not been released. Any product that claims this is describing a check it cannot run.

If you want to know whether Turnitin or GPTZero will flag you — that is a different question again, and unrelated to the watermark. Those tools judge writing style.

What each tool claims, in its own words

stealthgpt.ai — *“Detect and remove watermarks from AI generated text in just a few clicks… 99% human output when verified by AI detection.”* Overclaims. Detection of the watermark requires Anthropic's key. The “99% human output” figure is about AI detectors, which is a different system, sold in the same sentence.

claudewatermark.com — *“a free tool to remove the watermarks for AI systems like Claude”*, and separately *“eliminates hidden zero-width characters, invisible spaces, and other unicode markers.”* Self-contradicting on one page: the second describes character stripping, which cannot affect a mark carried in word choice. Also promises *“clean, undetectable text”* — undetectable by what, given no detector exists?

gptcleanup.com — title sells a *“Claude Watermark Remover”*, body hedges to *“Detect and remove supported hidden Unicode characters.”* Mixed. The qualifier is doing real work and the body never claims the statistical mark.

founderstoday.org — *“It does not represent a guaranteed watermark-removal percentage.”* Careful.

haimaker.ai — *“Merging those findings into a single ‘watermark removed’ result would claim more than the evidence supports.”* Careful, and the most rigorous write-up in the category.

github.com/guillaumemeyer/watermarks-remover — *“It cannot certify that vendor detectors will fail.”* Careful. Free, open source, and the most honest documentation of the lot, despite Reddit calling it snake oil.

Where we fit, and where we do not

We check for artifacts and clean them, free and unlimited, in your browser. Nothing is uploaded for that. There is a paid rewrite for the statistical layer, and we say plainly that nobody can verify whether it worked, us included.

What we do that the others do not: our detection engine is open source, so you can read exactly what it looks for rather than taking our word. And we publish measurements — including ones that undercut us. We found that non-breaking spaces, which we flag, appear routinely on bbc.com and gov.uk, so that class is not evidence of a chat interface, and we corrected our own tool when the measurement showed it.

What we do not do: detect Anthropic's watermark, remove it verifiably, or tell you that any text is undetectable. Nobody in this table can do those things and we would rather lose the sale than pretend.

How to choose, honestly

For a one-off paste, use whatever is free — that includes us, gptcleanup and the GitHub tool. The artifact-stripping part is not hard and nobody has a meaningful edge on it.

For repeated work on long documents, the rewrite step is where money is involved, and the honest framing is that you are paying for a model pass, not for watermark removal. Judge it on whether the output reads like you.

Ignore any tool promising certainty. The whole category's defining fact is that verification is impossible from outside Anthropic, and the tools that admit this are the ones telling you the truth about everything else too.

Questions

What is the best Claude watermark checker?

For copy-paste artifacts, any of the free tools work, including ours and the open-source watermarks-remover repository. For Anthropic's statistical watermark, none of them work, because detection requires a key Anthropic has not released. Choose on honesty about that limit rather than on claimed accuracy.

Do any of these actually remove the Claude watermark?

None can prove it. Substantially rewriting text does degrade a distributional watermark, because the mark is carried by the specific word choices, but no third party can verify the result. Tools claiming confirmed removal are describing a check nobody can run.

Which tools overclaim?

By their own published wording, stealthgpt.ai and claudewatermark.com. Both advertise removal or detection of the watermark while describing mechanisms that cannot achieve it. Quotes and dates are in the comparison above.

Is a free tool good enough?

For artifact checking, yes. Finding HTML class names and invisible characters is straightforward and free tools do it correctly. Money only becomes relevant for model rewriting on long documents.