AI Watermark Cleaner
Clean AI text from Claude, ChatGPT or Gemini in one paste: hidden Unicode characters, zero-width spaces, exotic spaces, smart quotes and dashes are gone, and the phrases that read as AI are flagged. Your words stay exactly as they were.
Paste your text here:
(e.g. in conclusion, it's worth noting that, let's dive in)
What it removes, and what it keeps on purpose
| Category | Codepoints | Action |
|---|---|---|
| Zero width space, word joiner, BOM | U+200B, U+2060, U+FEFF | removed |
| Soft hyphen, mongolian vowel separator, combining grapheme joiner | U+00AD, U+180E, U+034F | removed |
| Invisible math operators | U+2061 to U+2064 | removed |
| Unicode tag characters (hidden payloads) | U+E0000 to U+E007F | removed |
| Exotic spaces and blanks (nbsp, thin space, ideographic space, line and paragraph separator) | U+00A0, U+2000 to U+200A, U+202F, U+205F, U+3000, U+2028, U+2029 | replaced with a normal space |
| Smart quotes, ellipsis, bullets, dashes | U+2018 and friends | normalised to ASCII |
| Lookalike letters (Cyrillic і inside a Latin word) | per rulebook | flagged, replaced only when you switch it on |
Kept, because removing them breaks real text: the zero-width joiner next to an emoji or inside Arabic, Persian and Indic scripts; bidi marks in documents that contain right-to-left text; non-breaking spaces next to a digit or a unit (12 000, 10 %, 5 kg) and inside spaced abbreviations. Between two Latin letters, all of them are watermarks and go.
Clean whole folders from the terminal
The tool above cleans one text you paste. The same cleaner also runs as a command in your terminal, and there it cleans every text file in a folder at once: a blog with 60 articles, a docs folder, a book. Each changed file gets a backup next to it. On top of the text rules it strips the metadata from images and Word documents (author, camera, location), and it can run before every git commit, so a watermark never reaches your site again.
Install once, with Node or with Python:
npm install -g watermark-cleanerpipx install watermark-cleanerThen, inside any folder:
watermark-cleaner check .watermark-cleaner fix .check only reports what it would change. fix cleans the files in place. Without installing anything: npx watermark-cleaner check . Source, tests and the full rulebook: github.com/
What the AI watermark in text is, and what this tool does about it
Two different things get called an AI watermark. The first is what you can find in the characters: a zero-width space between two letters, a word joiner, a non-breaking space between two ordinary words, typographic quotes and dashes where you typed straight ones. They come from copy-and-paste chains, from tools that insert them, and for a few weeks in 2025 from ChatGPT itself. Detectors count them, and so does anyone who pastes your text into a code editor. This cleaner removes them and normalises the typography, with the same rules as the open-source command line tool, in your browser.
The second is the official kind. Since August 2026 Anthropic marks Claude output with a statistical watermark: a pattern in which words the model picks, readable only with Anthropic's key. Google's SynthID works the same way. No character tool can remove that, and this one does not claim to. Anthropic says the mark weakens through heavy editing, paraphrasing or translation, which is why the command line tool offers DeepL back-translation as an option, off by default, because it changes your wording.
What you can do here is the part that shows: the characters, the typography, and the AI phrases and sentence shapes flagged in the details for you to rewrite. Nothing is rewritten for you.
How this was built
I built the cleaner with AI, in public, after Claude started watermarking everything I publish. Two independent AI review rounds before the first release, 65 Python and 48 Node tests green. The build is episode 008 of the journal, with the video, the prompts and the files.
AI watermark FAQ
Not the official one. Since August 2026 Anthropic marks Claude output with a statistical watermark: a pattern in which words the model picks, detectable only with Anthropic's key. No character tool can remove that, and Anthropic itself says the mark weakens through heavy editing, paraphrasing or translation. What this cleaner removes are the traces AI text carries in its characters: zero-width spaces, word joiners, exotic spaces, typographic quotes and dashes, and it flags the phrases that read as AI. That is what detectors and readers notice first.
No. It removes characters you cannot see and normalises quotes, dashes and ellipses. AI phrases like "in today's fast-paced world" are flagged in the details so you can rewrite them yourself; the tool never rewrites a sentence.
No. The cleaning runs in your browser with the same rules as the open-source command line tool. There is no server, no account and no log.
Zero-width joiners next to emoji and inside scripts that need them (Arabic, Persian, Indic scripts), bidi marks in right-to-left documents, and non-breaking spaces next to numbers and units, like "12 000" or "10 %". Removing those would break legitimate text.
No, on purpose. A statistical watermark like Claude's or Google's SynthID sits in the choice of words, not in hidden characters, and the only way to weaken it is to rewrite or translate the text. Tools that offer this send your text to a server and hand it back in a different voice. This cleaner keeps your wording exactly as it is. If you want the rewrite anyway, the command line tool has an optional DeepL back-translation step, off by default.
Yes. The same rules ship as a command line tool for Python and Node: npx watermark-cleaner check . reports, npx watermark-cleaner fix . cleans every text file in place with a backup. It also strips metadata from images and Office documents, which the web version does not do.