Remove AI Watermark From Text Free: What Comes Off, What Does Not, and Why
Updated September 2026
Two things get called an AI text watermark. One is a set of characters you cannot see, planted in or picked up by the text. The other is a statistical pattern in the words themselves. The first comes off for free with proof. The second does not come off at all, and any page that says otherwise is selling a paraphrase and a hope.
Quick answer
Free, in your browser:
- 1Paste text and Scan. Every hidden character is listed by code point and shown in place.
- 2Strip. Zero-width characters, directional marks, tag characters and odd spaces are removed. Emoji and complex scripts are kept whole.
- 3Read the writing tells. Phrases and rhythm patterns are flagged so you can edit them yourself.
- 4Know the limit. A token-level statistical mark stays; nothing but rewriting changes it, and StripShot never rewrites.
What the free cleaner removes
| Carrier | Code points | Result |
|---|---|---|
| Zero-width space, non-joiner, joiner, word joiner | U+200B, U+200C, U+200D, U+2060 | Removed unless holding an emoji or a script together |
| Byte order mark, soft hyphen, combining grapheme joiner | U+FEFF, U+00AD, U+034F | Removed |
| Directional marks, embeddings, overrides, isolates | U+200E, U+200F, U+202A to U+202E, U+2066 to U+2069 | Removed |
| Unicode tag characters | U+E0000 to U+E007F | Removed unless part of a flag emoji |
| Variation selectors | U+FE00 to U+FE0F, U+E0100 to U+E01EF | Kept after emoji and CJK, removed alone |
| Non-breaking and typographic spaces | U+00A0, U+202F, U+2000 to U+200A, U+3000 | Replaced with a normal space |
| Look-alike letters | Cyrillic, Greek, full-width Latin | Replaced with the Latin letter (option) |
| Em dashes, curly quotes, ellipsis | U+2014, U+2018, U+2019, U+201C, U+201D, U+2026 | Normalized to keyboard punctuation (option) |
What it cannot remove, and why that is the honest answer
Google's SynthID-Text, published in Nature in 2024, and the model-level watermark Anthropic describes for Claude models launched from August 2026 both work by biasing which token the model picks at each step. The result reads as ordinary prose. There is no character to find. A detector holding the vendor's key counts how often the text landed on the keyed choices; nobody without the key can. Changing characters changes nothing about that count. Changing words does, which is what paraphrase does, and it weakens the mark without any way to confirm the result.
Why StripShot does not rewrite
Rewriting means sending your text to a model, accepting its changes to your meaning, and still not knowing whether the mark survived. That is not a service anyone can offer honestly, so it is not one StripShot offers. The writing-tells panel shows you the phrases that read as generated; the editing is yours.
When character cleaning is the whole answer
- Leaked-copy fingerprints. A pattern of zero-width characters unique to your copy is the classic way to trace a document. Cleaning removes it with certainty.
- Copy residue. Narrow spaces, soft hyphens and directional marks from web pages, PDFs and chat interfaces break search, code and matching.
- Smuggled instructions. Text hidden in tag characters is invisible to you and legible to the next AI tool you paste into. Finding it is a security step.
- Look-alike substitutions. Cyrillic letters in Latin words defeat spellcheck and search until they are swapped back.
Free text cleaner
Paste text to find and remove hidden carriers
Runs in your browser. Nothing is uploaded.
Options
Need unlimited strips?
Scans are free. Pro is unlimited.
Images, video, audio, documents, PDF and text in one tool. Pro adds unlimited strips, whole-folder batches and zip download with a JSON report.
Frequently asked questions
How do I remove an AI watermark from text for free?
Paste the text into StripShot and click Scan, then Strip. Hidden characters such as zero-width spaces, word joiners, directional marks and tag characters are removed, unusual spaces are normalized and look-alike letters are fixed if you turn that option on. The re-scan proves the count is zero. This is free and runs in your browser.
Does that remove ChatGPT, Claude or Gemini watermarks?
It removes any character-level carrier regardless of source. It does not remove a statistical watermark, the kind Google published as SynthID-Text and the kind Anthropic describes for Claude, because that lives in which words were chosen. No character tool can touch it and StripShot does not claim to.
Will an AI humanizer remove the watermark?
A humanizer paraphrases your text with another model. Paraphrase weakens statistical marks, and the research says so, but the humanizer cannot tell you whether the mark is gone, it changes your meaning along the way, and a rewrite by a marked model can stamp the text again. StripShot does not rewrite.
Is my text uploaded?
No. Scanning and cleaning run in your browser tab. Nothing is sent to a server or to a language model.
Can I see where the hidden characters are?
Yes. After a scan the text is shown with every hidden character marked as its code point in place, before anything is changed.