AI Watermark Remover for Text
Paste text and DetectMark strips the marks that are safe to strip by default, flags the rest for you to decide, and shows a receipt either way. Statistical watermarks get labeled, not faked.
- Hidden characters
- Available: Available Safe set by default; the rest opt-in
- Clipboard artifacts
- Available: Available
- Lookalike characters
- Unavailable: Unavailable Flagged, never removed
- Statistical watermark
- Unavailable: Unavailable Not locally removable
- Verification
- Re-scan of the cleaned text
- Uploaded
- 0 bytes of your text
What this remover changes
The one-click clean is deliberately conservative. It removes the characters that are inert almost everywhere they appear: zero-width spaces (U+200B), soft hyphens (U+00AD), interior byte-order marks (U+FEFF), and the class names and attributes a rich-text paste drags along behind it.
Everything else is detected and left for you to decide. Non-breaking and narrow no-break spaces (U+00A0, U+202F) can be replaced with ordinary spaces when you opt in. Zero-width joiners and non-joiners, variation selectors and bidi controls are flagged and individually removable, each with a warning about what removing it can break: joiners hold multi-part emoji together and are real spelling in Persian, Arabic and Indic scripts.
Every removal is listed by code point and position. The cleaned text is then scanned a second time, so the receipt reflects what actually left rather than what was requested.
Nothing is paraphrased. The wording you paste is the wording you get back.
What it does not change
A statistical watermark lives in the pattern of word choices rather than in the characters, so no tool running on your machine can find it or take it out. Anthropic put it plainly: "Nothing is added to the text and there are no hidden characters."
That is why the receipt always ends with an unchanged note for the statistical category. A cleaner that stays quiet about the thing it could not touch is telling you something false by omission.
Removing characters is not beating a classifier
Turnitin, GPTZero and Originality.ai score writing style. They do not read watermarks, and they never looked at your invisible characters in the first place. Stripping U+200B changes nothing about sentence rhythm or word choice, so it changes nothing about how those tools score a document.
FAQ
Not by cleaning. The cleanup removes hidden characters and clipboard fragments and labels the statistical watermark Unknown. What does remove the mark is a thorough rewrite through a model that is not Claude, because the mark lives in Claude's word choices, and this page can run that rewrite through OpenAI on your explicit request, with a receipt of how much wording changed. Nobody outside Anthropic's private preview can verify the result either way.
No. The cleanup removes non-printing characters and clipboard styling that can be identified exactly, and leaves every word alone. Separately, and only when you ask, the page can send your text to OpenAI for a full rewrite and show the result beside the original with a count of how much wording changed. It never rewrites silently and never claims what the rewrite removed.
Unlikely. GPTZero, Originality.ai and Turnitin classify by writing style, not by hidden characters. Removing invisible Unicode changes neither the style nor a real watermark.
Any plain or rich text. A rich paste has its styling stripped and the underlying characters checked, and the clipboard fragments it carried are reported as their own category.
No. There is no login, no paid tier, and no server that receives your text.
Every check runs in your browser: how it works. The limits are written down too: what DetectMark cannot remove.