ChatGPT Watermark Remover and Hidden Character Cleaner
Paste text to highlight hidden characters and look-alike letters. Clean them in one click and copy the result.
0 / 100,000 characters · runs in your browser
Paste text to find hidden characters
- Paste your text into the box. Hidden characters are counted by type and marked in a copy of your text below it.
- Hover over a mark to see exactly which character it is, for example “U+200B zero-width space”.
- Click Remove hidden characters, then Copy text.
All of this happens in your browser. Nothing is uploaded, there's no daily limit and you don't need an account. If you want to see how the text reads to an AI detector, click Check for AI and it opens in our AI detector.
Does ChatGPT add a watermark to its text?
Not the kind you can delete. A text watermark, where one exists, is a statistical pattern in which words a model picks. There are no extra characters, so a cleaner like this one can't find or remove it. Here's what each company says, as of November 11, 2026:
- OpenAI says verification of its text marks is available only to approved organizations, such as research and academic institutions.
- Anthropic says recent Claude models, including Fable 5.1 and the 5.5 versions of Opus, Sonnet and Haiku, add a text watermark; older ones such as Sonnet 4.5 don't. Our Claude detector page has the details.
- Google watermarks text from the Gemini app and website with SynthID.
Hidden characters are a separate thing. Zero-width spaces, odd spaces and soft hyphens can come along when you copy text from a chat window, a PDF or a web page. They aren't a signature, but they can break code, formatting and word counts.
Our own detector doesn't check for a watermark either; it judges the writing. Our explainer on how AI detection works covers how the two approaches differ.
Characters this tool finds and removes
| Character | Code | What you see | What the cleaner does |
|---|---|---|---|
| Zero-width space | U+200B | Nothing | Removed |
| Zero-width non-joiner | U+200C | Nothing | Removed |
| Zero-width joiner | U+200D | Nothing | Removed, except inside emoji like 👨👩👧 |
| Word joiner | U+2060 | Nothing | Removed |
| Byte order mark | U+FEFF | Nothing | Removed |
| Soft hyphen | U+00AD | Nothing, or a hyphen at a line break | Removed |
| Thin, hair, en, em and other fixed-width spaces | U+2000 to U+200A | A space of a slightly different width | Replaced with a normal space |
| Narrow no-break space | U+202F | A narrow space | Replaced with a normal space |
| Medium mathematical space, ideographic space | U+205F, U+3000 | A space | Replaced with a normal space |
| Cyrillic or Greek letters inside English words | e.g. U+0430 а | Identical to a, e, o, p and others | Replaced with the Latin letter |
| No-break space | U+00A0 | A normal space | Kept: it often belongs there, as in “Mr. Smith” |
Where possible, a letter and a separate accent mark are combined into one character. The cleaning rules are the same ones our humanizer uses on rewritten text.
What cleaning can't do
It can't remove a statistical watermark. We also tested whether cleaning changes an AI score on our own detector.
We took 100 AI-written texts and made a copy of each with hidden characters added: zero-width spaces, unusual spaces, soft hyphens and a few look-alike letters, about 59 per text. We scored the copy with hidden characters and the cleaned version 3 times each, without using saved results, and compared the averages.
| Measure | Points out of 100 |
|---|---|
| Median change | 0 |
| Middle 80% of texts | -0.7 to 0 |
| Largest drop and rise | -33.0 and +0.3 |
| Texts within 5 points | 96 of 100 |
| Texts whose verdict changed | 3 of 100 |
For most texts, cleaning made almost no difference. Where it did, the score went down, not up: in 4 texts the version with hidden characters had scored higher. Scores also vary a little from one run to the next, which is why we averaged three. This is what we observed on our detector, not a promise about any other.
If you want to change how the text reads, edit the wording itself: our tool to humanize AI text rewrites only the flagged sentences. The ChatGPT humanizer page shows which phrases were more common in our November 2026 GPT samples than in human writing.
Where hidden characters cause real problems
- Code. A zero-width space inside a variable name or a non-standard space in an indent can cause syntax errors that are almost impossible to see.
- Web pages. Text pasted into a CMS can wrap oddly or refuse to wrap, because some special spaces don't allow a line break.
- Word counts and limits. Different tools count these characters differently, so a 500-word answer can come out over a form's limit.
- Find and replace. Searching for a word fails if a zero-width character or a look-alike letter sits inside it.
How to see hidden characters in Word and Google Docs
Word. On the Home tab, click Show/Hide ¶ in the Paragraph group. Spaces show as dots, and optional hyphens and other formatting marks appear. On a Mac you can choose which marks to show under Word > Preferences > View. Zero-width characters and look-alike letters still don't show.
Google Docs. Click View, then Show non-printing characters. Google's help page says this shows paragraph, line, section, page and column breaks, tabs and spaces in blue. It doesn't mention zero-width characters, and look-alike letters look like normal letters.
For those, paste the text into the box at the top of this page. (Both help pages checked November 11, 2026.)
Frequently asked questions
Does ChatGPT put invisible characters in its text?
- OpenAI hasn't said that it adds hidden characters as a watermark. Odd spaces and zero-width characters do turn up in text copied from chat apps, PDFs, websites and word processors, so it's worth checking text from any source.
Is this a real ChatGPT watermark remover?
- It removes hidden characters and look-alike letters. It can't remove a statistical watermark, which is built into which words were chosen, and no character cleaner can.
Will removing hidden characters lower my AI score?
- Usually not. In our test of 100 AI texts, cleaning moved our detector's score by 5 points or less for 96 of them. In 4, the copy with hidden characters scored higher, so cleaning brought the score down. Other detectors may behave differently.
Does it change my visible text?
- Look-alike Cyrillic or Greek letters inside English words are replaced with the Latin letters they imitate. Removing hidden characters can also change spacing, where lines break and how some letters join. Words themselves aren't rewritten.
Can it remove Claude's or Gemini's watermark?
- No. Their watermarks are statistical patterns in word choice, not characters, so there's nothing for a cleaner to delete.
Is my text sent to your server?
- No. The check and the cleaning run in your browser. Text only leaves the page if you click Check for AI, which sends it to our AI detector.
Is there a word limit?
- Up to 100,000 characters at a time, with no daily limit and no sign-up.
Does it work on text from other AI tools?
- Yes. It doesn't care where the text came from: Claude, Gemini, a PDF or a website all work the same way.