Remove AI Watermarks From Text: What Can and Cannot Be Removed

There are two very different things people mean by an AI text watermark. One is a pattern of hidden characters planted in the text, which can be found, removed and verified. The other is a statistical signature in the choice of words, which no character-level tool can touch. StripShot does the first in your browser and is straight with you about the second.

No uploadNothing sent to a modelNever rewritesFree

Last updated September 2026

Two kinds of mark, two very different answers

Mark	Where it lives	Examples	Removable here?
Character carriers	In the bytes of the text: invisible or look-alike code points	Zero-width spaces, word joiners, soft hyphens, directional marks, tag characters, narrow no-break spaces, Cyrillic look-alikes	Yes. Deterministic, listed, verified by re-scan.
File provenance	In the container, not the words	C2PA Content Credentials, XMP, EXIF, document properties naming the tool	Yes, on images, PDF, DOCX and other files. See the format pages.
Statistical token watermarks	In which words were chosen, at every position	SynthID-Text (Google), the model-level marks Anthropic describes for Claude, Kirchenbauer-style green-list schemes on open models	No. Only rewriting changes them, and StripShot does not rewrite.
Stylistic tells	In phrasing and rhythm	Delve into, rich tapestry, a testament to, uniform sentence length, dense em dashes	Reported, not changed. You edit; the tool points.

Why StripShot will not paraphrase your text

The product promise is deterministic cleaning on your device. Rewriting text to shake off a statistical watermark means sending it to a model, accepting whatever that model changes, and still having no way to verify the result without the vendor's key. Research on current token-level schemes shows paraphrase weakens them, and also shows that a rewrite by another marked model can stamp the text again. That is not a service we can offer honestly, so we do not offer it.

What each vendor has said

Anthropic (Claude). Its help center describes imperceptible text watermarks applied at the model level for models launched on or after August 2, 2026, worldwide, with older models in transition, plus C2PA Content Credentials on supported file types. The algorithm is unpublished and third-party detection was described as forthcoming. Anthropic notes that a detected mark means the content may have been processed by Claude, that no mark does not mean human-only origin, and that proofreading or translating human text can stamp it.
Google (Gemini). SynthID-Text, published in Nature in October 2024, adjusts token sampling so a keyed scoring function can detect the pattern later. Google keeps the production key private and has changed which surfaces carry the mark over time. It has also published open research code, which is how independent tools verify the technique in controlled tests.
OpenAI (ChatGPT). OpenAI has said it has a text watermarking method and has not deployed it publicly. Its image outputs carry C2PA. Occasional hidden characters in text output were described by OpenAI in 2025 as a training side effect rather than a watermark.
Open models. Self-hosted engines can add green-list or keyed sampling watermarks (Kirchenbauer and successors). Detection needs the key; a negative result from any tool that does not hold it proves nothing.

What StripShot does on this page

Finds carriers. Every invisible or look-alike code point is listed with its name, code point and count, and highlighted in place so you can see exactly where it sits.
Removes them without collateral damage. Joiners that hold emoji, flags and complex scripts together are recognized by context and kept. Free-floating ones go.
Reports honestly. The re-scan shows what is gone. It does not say the text is now human or undetectable, because character cleaning cannot establish either.
Shows the tells. Phrase markers, sentence-length variation, lexical diversity and em dash density are scored locally, with no model involved. Use it as an editing aid.

Use it for what it is

Cleaning your own drafts, stripping fingerprints from documents you are about to share, removing hidden text that a chat interface or a web page slipped into your clipboard, and understanding what provenance signals your files carry are all legitimate reasons to use this tool. Misrepresenting AI assistance where disclosure is required is not something a character cleaner enables, and it is not something StripShot supports. The disclaimer spells out the boundary.

How to check text for AI marks and remove the ones that can be removed

1
Paste and scan
Paste the text above and click Scan text. Every hidden character is listed by code point and shown in place.
2
Read the tells
The writing-tells panel lists phrases and rhythm patterns that read as generated. It is a style read for you to act on, not a verdict.
3
Strip the carriers
Click Strip. Invisible characters are removed, spaces normalized, and, if the options are on, look-alike letters and punctuation are fixed. The output is re-scanned.
4
Know what is left
If the text came from a model that marks at the token level, that mark is still in the words. Only rewriting changes it, and StripShot does not rewrite.

Frequently asked questions

Does Claude watermark text?

Anthropic's help center describes embedded text watermarks applied at the model level for models launched on or after August 2, 2026, plus C2PA Content Credentials on supported files such as PNG, JPEG and SVG. Anthropic has not published the text algorithm, and says a detected mark means the content may have been processed by Claude, not that Claude wrote all of it. A text watermark of this kind lives in word choice. StripShot cannot remove it and does not claim to; it removes character-level carriers and the C2PA on files.

Does ChatGPT watermark text?

OpenAI has said it built a text watermarking method and has not deployed it publicly. ChatGPT images carry C2PA Content Credentials, which StripShot's photo tool removes. Pasted ChatGPT text often carries typographic and occasionally hidden characters, which this tool cleans.

Does Gemini watermark text?

Google published SynthID-Text in Nature in 2024, a watermark applied while the model samples each token, and has used it in Gemini products. Google does not publish the detection key, and which Gemini surfaces mark text has changed over time. Because the mark is in the words, character cleaning does not touch it.

Can any tool remove a statistical text watermark?

Only by changing the words. Paraphrasing, translation and heavy editing weaken these marks, and research shows current token-level schemes are fragile under rewriting, but no tool can certify that a vendor detector will come back negative without the vendor's key. StripShot does not paraphrase and does not send your text to any model.

So what does this page actually remove?

Character-level carriers: zero-width spaces and joiners, word joiners, soft hyphens, directional marks, variation selectors, Unicode tag characters, non-breaking and typographic spaces, and look-alike letters. Every removal is listed and the re-scan proves it. On files, StripShot also removes C2PA and metadata.

Is removing these marks legal and ethical?

For content you own or are authorized to process, cleaning invisible characters and metadata is ordinary hygiene and privacy work. Where disclosure of AI involvement is required by law, a platform or an institution, removing a mark does not remove the obligation. StripShot will not help you claim cleaned text is human-written.

Is my text sent anywhere?

No. Scanning and cleaning run in your browser tab. Nothing is uploaded and nothing is sent to a language model.

Clean 3 images free. Then go unlimited.

No card needed to start. Upgrade anytime.

Free

2 images/day free

2 images/day free
1 video/day free
AI detection (70+ signatures)
100% private, no uploads

Scan text free →

POPULAR

Pro

$9/mo

Unlimited photos

Unlimited image cleanings
10 videos/day included
Batch processing
Smart Strip & hex viewer
Edit & write EXIF tags

Start for $1 →

Full Access

$19/mo

Unlimited everything

Unlimited image cleanings
Unlimited video cleanings
Strip Sora, Runway, Kling
MP4 + MOV support
New generators added regularly

Get Full Access - $19/mo

Cancel anytime · No commitment · Files never leave your device