Guide
How to remove AI watermarks from your text
Paste the text, press Clean text, take the clean version back. This page is the rest of the answer: what those marks turn out to be, three ways to strip them, and the one kind no cleaner can touch.
Published · 7 minute read
What an “AI watermark” is, in practice
When someone says their text has an AI watermark, what they have found is almost always a character. Assistants, editors, content pipelines and copy-paste itself leave marks in a string that render as nothing at all: a space with no width, an invisible joiner that forbids a line break, an unusual space where the space bar belongs, a Cyrillic letter standing in for a Latin one.
You cannot see any of it, and all of it survives copy-paste. It breaks exact-match search, splits a word for a spell checker, and turns up in a diff as a change to a line nobody edited. That is the thing a character cleaner removes, and it is what the rest of this page is about.
The ones that turn up most often:
U+200BZERO WIDTH SPACE — a space with no width and no glyph — the classic carrier for hidden data.U+2060WORD JOINER — invisible, forbids a line break, and slips past filters that only look for the zero-width space.U+00A0NO-BREAK SPACE — a space that is not the one on your keyboard, so search and comparison both miss the word beside it.U+E0000–U+E007FTAG CHARACTERS — a whole invisible copy of ASCII: a message can be written in these and pasted inside ordinary prose.U+0430CYRILLIC SMALL LETTER A — a Cyrillic letter wearing a Latin letter’s face, inside a word that mixes scripts.
None of this proves a machine wrote your text. It proves the text passed through something — a chat window, a CMS, a translation tool, someone else’s clipboard. That is worth knowing, and it is worth cleaning up before you send the text on.
See what is in your own text
The cleaner on the front page shows every one of them: line and column, the codepoint, the official Unicode name, and what was done about it. It is free, there is no account and no character limit, and nothing is uploaded — the engine is JavaScript running in your own tab, and it keeps working with the network off.
If you have nothing to hand, press Load sample in the tool. The demo paragraph is seeded with a zero-width space, a word joiner, tag characters inside a flag emoji, a look-alike Cyrillic letter and a right-to-left override, so you can watch each one get its own verdict instead of taking ours.
Three ways to remove them
The same engine in all three. Pick the one that fits where your text already is.
Paste it into the web cleaner
Open the cleaner, paste, and one press cleans the text and puts it back on your clipboard. Free and unlimited, no account, nothing uploaded. This is the whole job for a single draft. Open the cleaner
Let the Chrome extension clean as you paste
The extension cleans pasted text in place, in the box you are already typing in, so a paste never has to be routed through a website first. It is free, it collects no data, and clean-as-you-paste is opt-in, one site at a time. Chrome extension
Clean anywhere on your Mac with a hotkey
The menu-bar app cleans whatever is on your clipboard from any application — a mail client, a code editor, a form in a browser you do not use. The app is licensed; without a licence the hotkey and clean-as-you-copy still run 20 times a week, and the report of what was found is never gated. Mac app
What this does and does not remove
Three different things get called an AI watermark. A character cleaner removes exactly one of them, and being clear about which is the point of this section.
Character-level marks: removed
Everything in the table is a character the engine can see, name and act on. Most of it goes in the default pass because it is invisible and doing no work. The rest sits behind an option, because removing it would change text you can read — and that is a decision you make, not one we make for you.
| Family | Codepoints | What happens |
|---|---|---|
| Zero-width and invisible | U+200BU+2060U+00ADU+FEFFU+2061U+2062U+2063U+2064U+034FU+180EU+2800U+3164U+FDD0–U+FDEF | Removed by default |
| Bidirectional overrides | U+202EU+202DU+202AU+202BU+202C | Removed by default |
| Bidirectional marks | U+200EU+200FU+061CU+2066–U+2069 | Removed unless it is doing a job |
| Joiners, selectors and fillers | U+200DU+200CU+FE00–U+FE0FU+E0000–U+E007FU+180B–U+180DU+17B4–U+17B5U+115F–U+1160 | Removed unless it is doing a job |
| Private-use characters | U+E000–U+F8FF | Needs “Private-use characters” |
| Exotic spaces | U+00A0U+202FU+205FU+3000U+1680U+2000–U+200A | Needs “Spaces” |
| Line separators | U+2028U+2029U+0085 | Needs “Line endings” |
| Control characters | U+0000–U+001FU+0080–U+009F | Needs “Control characters” |
| Typographic punctuation | U+2014U+2019U+2026 | Needs “Punctuation” |
| Lookalike letters | U+0430U+043EU+0435U+0441 | Needs “Lookalike letters” |
Every codepoint above has a page of its own: what it is for, when it is doing a real job in someone’s script, and what we do about it. Character index
Statistical watermarks: not removed, by anyone
The other kind of watermark is not made of characters at all. Schemes like Google’s SynthID-Text bias which words the model picks, so the mark lives in the wording — in ordinary letters, ordinary spaces, ordinary punctuation that nothing about the characters would ever flag. No character cleaner can see it, and this one does not claim to. A tool that tells you it strips that kind is selling you something.
Your own writing habits: not a character at all
The third thing people mean is how the prose reads: sentences that are all the same length, stock transitions, a high em-dash rate, no specifics. That is not in the characters either, and no amount of cleaning changes it. The prose-habits report measures those habits and hands you the numbers — no score, no probability, nothing to argue with — so you can edit them yourself.
Questions people actually ask
Will cleaning the characters change how my writing is scored?
Almost certainly not. Tools that score writing read the wording — how predictable each word is, how much that varies — and invisible characters are not part of that calculation. Cleaning changes the characters and leaves your sentences exactly as they were. If you want the wording to read as yours, that is an editing job, and the prose-habits report is where it starts.
Is there really a hidden watermark in ChatGPT output?
There is no published character-level watermark in ChatGPT’s text. Zero-width characters do turn up in text that has been through a web editor, a CMS or a long copy-paste chain, and people find them afterwards and call them a watermark. What OpenAI and Google have actually published is the statistical kind, which lives in word choice and which no character cleaner touches.
Is the cleaner free, and does my text get uploaded?
Free and unlimited, with no account and no character cap. The engine is JavaScript in your browser: turn off your network and the page still cleans. Nothing you paste is sent anywhere, because there is no server to send it to — the site is a static export.
Will cleaning break my emoji or my non-English text?
No. The same zero-width joiner is contraband between two Latin letters and load-bearing inside a family emoji or a Persian word, so the engine decides per character with the neighbours in view. It keeps the ones doing a job, and the findings table tells you which those were and why.
What about a Word document or a PDF?
Text is only half of a document. Take a .docx, .odt or .html file to the documents page: it reports hidden runs, white-on-white text, tracked changes, comments and authoring metadata as well as the characters, and it never deletes those for you, because that changes the document. PDF is not supported.
Do em dashes count as a watermark?
No. An em dash is visible typography, not a hidden mark, and people who write well use them. The cleaner leaves dashes alone at Standard and converts them only if you ask for it. There is a separate guide on doing exactly that, and on what comes with a pasted draft that you cannot see.