Nothing is added to your writing. The word choices themselves are the mark. Here’s what it detects, what it can’t, and the honest version of removing it.
Anthropic announced in August that Claude models launched from August 2, 2026 on write with an invisible watermark, and older models get it over the coming months. Nothing gets inserted into the text. The word choices themselves are the mark.
Most people came here for the remover tool, so it’s first. The part worth understanding comes right after.
It’s an open-source project (MIT licensed) by Guillaume Meyer. It runs as an agent skill or as a small local server that needs only Python 3.10+.
Open the watermark remover on GitHub →
It works on three layers, and the difference between them is the whole story.
01
Invisible characters (lossless)
Zero-width spaces, bidirectional marks, exotic spaces and tag characters. Your visible words don’t change at all. Handy for anyone who’s pasted text and had strange artifacts ride along.
02
File metadata (lossless)
Strips C2PA content credentials, EXIF, XMP and document properties from images, PDFs, Word and Office files, HTML, Markdown and common audio and video files. The content itself stays the same.
03
Statistical rewriting (best effort only)
This is the layer aimed at the actual watermark. The only way to touch the pattern is to change a big share of the words, so the README says plainly that rewording can flatten your tone and that no tool can certify a vendor’s detector will come back clean.
The project says it’s for content you own. The EU marking rule falls on AI providers, not on you as a writer. Passing off AI work as your own where you agreed not to is the same problem it always was, tool or no tool.
A model writes one word at a time. Take "the weather today was cold and." The next word could be overcast or grey. Both work and the meaning stays the same. Normally a plain random number settles that pick.
Watermarking keeps the same word options and changes the source of the randomness. A secret key plus the few words before it decide which option wins. The output still reads naturally, but whoever holds the key can scan a long passage and check whether the choices match the key’s pattern. Anthropic says it’s a version of Google DeepMind’s SynthID-Text method.
Per Anthropic, it can’t be traced to you, your account or your company. It costs zero extra tokens, and readers can’t tell watermarked text apart. Translations Claude produces stay watermarked too. Images and files Claude makes get a C2PA content credential in their metadata instead.
Claude isn’t alone here. The EU now requires AI providers to mark generated content, and Anthropic is one of around 190 signatories to the EU’s Code of Practice on it. Anthropic applies it everywhere because it has no reliable way to limit it by region yet.
Know this before someone waves a result at you.
01
It can suggest Claude was involved
Given enough text, a key holder can estimate how likely it is that Claude touched it. Longer passage, higher confidence.
02
It can’t tell writing from editing
It can’t separate "Claude wrote this essay" from text Claude heavily edited, and it can’t prove a human wrote something. Light proofreading may leave too little to detect.
03
It’s weak on short text, facts and code
Fewer word choices means less signal, so a caption gives it almost nothing. Where only one right answer exists, there’s no free choice to hide a pattern in. After "Newton’s most famous work was called Principia," the next word has to be "Mathematica."
Anthropic’s detection API is now in private preview for regulators, media, fact-checkers, researchers, schools and companies with compliance needs (interest form). There’s still no public checker for text. If a school or employer treats a result as proof of authorship, that’s a policy problem more than a tech one.
This prompt does the half of the job that’s safe. It cleans what can be cleaned without changing your words, tells you what it can’t see, and makes you answer two questions before you degrade your own writing chasing a pattern that may barely be there.
To install, go to Customize (now being renamed Plugins), then Skills, click +, then Create skill, then Upload a skill, and pick the ZIP without unzipping it. (How skills work)
Marking machine-written content is reasonable. The gap is that a mark shows involvement, not authorship. The best move is keeping more of your own thinking in the work. Less to strip, and better writing anyway.