How do I remove an AI watermark from text?
You can remove hidden Unicode characters, zero-width spaces, bidirectional controls, tag characters and unusual spaces, completely, using the cleaner below. You cannot remove a statistical watermark such as Google's SynthID or Claude's confirmed mark, because those live in word choice, not in characters.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
That distinction decides whether this page can help you. If your text carries invisible characters, the cleaner strips them in one pass and hands back a corrected copy. If your text was generated by a model that applies a statistical watermark, no amount of character cleaning changes the signal, and any tool claiming otherwise is selling something it cannot deliver.
What the cleaner removes
Every category below is handled in a single code-point pass, so a document with several different problems is fixed in one go rather than one class at a time.
| What is in your text | What the cleaner does |
|---|---|
| Zero-width space, zero-width no-break space, word joiner, invisible maths operators | Removed |
| Soft hyphen, combining grapheme joiner, deprecated format controls | Removed |
| Left-to-right and right-to-left marks, embeddings, overrides and isolates | Removed and flagged as high risk |
| Tag characters (the invisible ASCII block) | Removed, unless they form a valid flag emoji |
| C0 and C1 control characters, except tab, line feed and carriage return | Removed |
| No-break space, thin space, figure space, ideographic space and the rest | Replaced with an ordinary space |
| Line separator and paragraph separator | Converted to a normal line break |
| Hangul fillers and blank braille cells used as invisible padding | Removed when they sit outside Korean or braille text |
What the cleaner deliberately leaves alone
A cleaner that deletes every invisible character is not thorough, it is careless. These stay, and the x-ray view labels them in green so you can see the decision being made:
- Zero-width joiners inside emoji. The family emoji is three separate people joined by two invisible characters. Remove them and you get three separate people.
- Variation selectors after an emoji or a CJK ideograph. U+FE0F is what makes a symbol render in colour, and it is also part of every keycap sequence.
- Tag characters inside a flag emoji. Regional flags are built from an invisible tag sequence closed by U+E007F.
- Zero-width non-joiners in Persian, Urdu and Indic scripts. There they are orthography, not noise.
- Khmer inherent vowels, Mongolian vowel separators and Hangul fillers inside their own scripts.
- Blank braille cells adjacent to other braille.
How the decision is made. For each candidate the engine looks at the characters on either side and tests their Unicode script and emoji properties. Same code point, different verdict, depending on where it sits. The invisible Unicode list documents the rule for every code point in the table.
The honest scope of this page
What this tool cannot do. It cannot remove Google's SynthID from Gemini output, or any statistical token-level mark another vendor deploys now or later. It cannot strip C2PA provenance metadata from generated files, and it is not designed to. It does not rewrite, paraphrase or alter your wording in any way. The cleaned text has the same words in the same order as what you pasted.
What it does is repair text. If you are preparing a data import, committing code, filling a form that keeps rejecting a value, or publishing content that was assembled from several sources, the invisible debris is a genuine problem with a genuine fix, and that fix is above.
Where the debris comes from
Understanding the source is how you stop it coming back. In practice almost all of it arrives through one of five routes: PDF text extraction, word processors, spreadsheet exports, web page copy-paste, and localisation or translation files. Each leaves a recognisable signature, described in the hidden characters guide.
Model output is a sixth route, but a smaller one than the internet suggests. When a chat interface renders a response into HTML and you copy it from the page, you are copying the page's typography, including whatever no-break spaces and thin spaces the renderer inserted. That is a property of the copy, not of the model.
Frequently asked questions
Does removing hidden characters change how my text reads?
No. Every character the cleaner removes is invisible on screen, and every unusual space is replaced by an ordinary space rather than deleted, so words never run together. The visible words and their order are unchanged.
Why can a statistical watermark not be removed by a cleaner?
Because there is nothing to delete. A statistical watermark biases which of several acceptable words the model picks, spread across hundreds of tokens. The signal is the pattern of choices, not a character in the text, so removing characters leaves it intact. Only rewriting enough of the passage to change the word choices disturbs it.
Can I clean a whole document at once?
Yes. Paste the full text into the input box. The cleaner processes everything you paste, however long. For very long documents the x-ray view stops drawing chips after a few hundred to keep the page responsive, but the cleaned output and the tally still cover the entire text.
Is it safe to paste confidential text here?
The processing itself never leaves your browser: nothing is uploaded, stored or logged, and the page makes no network request while you type. As with any web page, apply your organisation's own policy about what may be pasted into a browser tab.
Should I turn on the quote and dash options?
Only if you want visible punctuation changed. Straightening curly quotes and converting em and en dashes to hyphens are editorial preferences, not corrections, which is why both are off by default. They are useful when pasting prose into code, YAML or a system that only accepts ASCII punctuation.
Related pages
- The full AI watermark remover and checker with the x-ray view and the complete explainer.
- How to remove AI watermarks, step by step - the walkthrough, from working out which kind you have to cleaning it.
- Claude's confirmed watermark - the newest deployment, and exactly what cleaning cannot touch.
- Statistical versus character watermarks - the mechanics behind that limit.