Remove Claude Watermarks

Detect and strip invisible Unicode marks, look-alike spaces, bidi overrides and homoglyphs from text you own. Everything runs in your browser — nothing is uploaded, and the page makes no network requests.

Text

0 characters

Report

empty

Paste some text to scan it.

This scans for invisible characters only. It cannot detect or remove a statistical watermark baked into the words themselves, so a clean report here is not a clean bill of health.

Options

What this does, and what it cannot do

What it removes

The deterministic, detectable half of AI provenance marking: zero-width and format characters, Mongolian/Hangul/Khmer fillers, variation selectors, Unicode tag characters, noncharacters, reserved invisible codepoints, private-use characters, bidi overrides, and look-alike spaces and Latin homoglyphs. These are the carriers that survive copy-paste and are used for steganographic marking and for fingerprinting pasted text.

What it cannot remove

Claude's published text marking is applied at the model level as a statistical (token-sampling) watermark — a bias in which tokens get chosen, not a character hidden in the string. There is no deterministic removal for that; the honest answer is that it is weakened by rewriting, paraphrase, translation, or mixing with other text, and that short texts carry less signal than long ones. This tool does not pretend otherwise.

Conservative by design

Load-bearing invisibles are kept unless you ask otherwise: the zero-width joiner in 👨‍👩‍👧, the variation selector in ⚖️, the joiners inside Persian and Devanagari words, complete flag-tag sequences, and bidi marks in genuinely mixed-direction prose. Replacing those would visibly change the text, which is worse than leaving them.

Use it on your own text

Removing provenance marks to pass off someone else's work, to defeat an audit, or to misrepresent where content came from is not what this is for. It exists for privacy and hygiene on material you own or have the right to alter.