Claude Text Watermark Remover
Claude's watermark is a statistical pattern in which words the model chose. There is no character to delete, so the only thing that touches it is changing those words. This tool does that mechanically — and reports honestly how much it changed, because that is the number that matters.
Statistics appear once you paste some text.
Why deleting characters does nothing
Anthropic's watermark is applied inside the model. At each step the model has several plausible next words; a secret key sorts them into two groups and nudges sampling toward one of them. One word tells you nothing, but across hundreds of words the distribution is measurably skewed. The mark is not hidden in the text — it is made of the text. Strip every invisible unicode character and the word choices, and therefore the signal, are untouched.
What actually degrades the signal
Anthropic's own documentation lists the cases that leave no detectable mark: text that has been heavily edited, paraphrased, translated or mixed into other writing, and passages too short to carry a reliable signal. Every one of those works the same way — it replaces the model's word choices with someone else's. That is the mechanism this tool automates: substitute synonyms, swap the connectives models overuse, and toggle contractions, so a measurable share of the original tokens are gone.
How much rewriting is enough?
Nobody outside Anthropic can answer that, because the detection threshold and key are not public. What is known from the published research this class of watermark is built on is the shape of the curve: detection confidence falls as the proportion of substituted tokens rises, and long passages tolerate more editing than short ones before the signal breaks. So treat the substitution rate as the honest metric and assume more is better — this is disruption, not a guaranteed erase.
Read the changes before you use the text
Mechanical substitution has no understanding of your sentence. It will occasionally produce something awkward — "help you find" becoming "aid you locate" is the classic failure. The Review changes panel lists every substitution so you can catch them. Text you have actually reread and fixed by hand is both better writing and, incidentally, further from the original token sequence.
What a detected mark would have proved anyway
Worth knowing before you assume you need this. Anthropic is explicit that a mark indicates content may have been processed by Claude — not that Claude wrote it. Proofreading, translating or summarising your own writing produces a marked output. If you are facing an accusation based on a detector result, the limitations section of Anthropic's own help centre article is more useful to you than any rewriting tool.
Frequently asked questions
+ − Can you actually remove Claude's text watermark?
Not with certainty, and be sceptical of anything claiming otherwise. The watermark is a statistical property of the model's word choices, so it degrades as those choices are replaced rather than switching off. Heavy rewriting, paraphrasing and translation are what Anthropic itself lists as leaving no detectable mark.
+ − Why can't a normal watermark remover do this?
Because character cleaners delete invisible unicode, and Claude's watermark is not made of invisible unicode. There is nothing for them to find. A cleaner is still worth running for other reasons, but it has no effect on the embedded mark.
+ − Does rewriting hurt the quality of my text?
At moderate strength the substitutions are near-exact synonyms and read naturally. At maximum strength quality drops noticeably. Use the Review changes panel and fix anything awkward — hand editing improves the text and moves it further from the original tokens at the same time.
+ − Is my text sent to a server?
No. The dictionary ships with the page and the rewriting runs in your browser. There is no upload, no API call and no logging — it works offline.
+ − What about translation as a method?
Round-tripping through another language is one of the most effective disruptions, because it replaces essentially every token. It cannot be done offline in the browser, so it is not built in here, but translating out and back with any translation service achieves it.
More free tools
ChatGPT Watermark Remover
Detect and strip invisible unicode, zero-width spaces and hidden markers from ChatGPT text. Free and entirely in your browser.
Em Dash Remover — Strip Em Dashes From AI Text
Remove em dashes, en dashes and two-em dashes from AI text, or swap them for plain hyphens. Free, instant, no upload.
Invisible Character Detector
Reveal invisible unicode in any text. Lists every hidden code point with its name and count. Free and client-side.
Zero Width Space Remover
Remove zero width spaces, joiners, non-joiners and byte order marks from your text instantly. Free, runs in your browser.
AI Text Cleaner
Strip invisible unicode, em dashes, smart quotes and messy whitespace from AI output in one pass. Client-side, no upload.
Free Text-Cleaning MCP Server
A free hosted MCP server giving your agent text-cleaning tools. Install in Claude Code, Cursor or VS Code in one line.
Stop AI Using Em Dashes — Skills & Prompts
Install a Claude Skill or paste custom instructions to stop ChatGPT, Claude and Gemini using em dashes and AI filler.
C2PA Metadata Remover
Remove C2PA provenance metadata, EXIF and XMP from AI-generated images. Re-encodes in your browser; files never uploaded.
Related guides
Em Dash vs En Dash vs Hyphen
The difference between the em dash, en dash and hyphen, when to use each, and how to type them on Mac and Windows.
How to Type an Em Dash (—)
Every em dash shortcut: Mac (Option+Shift+-), Windows (Alt+0151), Word, Google Docs, iPhone, Linux and HTML.
Why Does AI Use So Many Em Dashes?
ChatGPT and Claude overuse the em dash badly. The real reason it happens, why it signals AI writing, and how to fix it.
Does ChatGPT Watermark Its Text?
OpenAI built a ChatGPT text watermark and never shipped it. What is confirmed, what is myth, and how to check your text.