Claude started watermarking its text, so I built a tool that removes AI watermarks
So Claude made the news recently for watermarking its output. I got curious and pasted some AI text into a hex viewer to see what was actually in there.
Turns out there are two completely different things going on:
Invisible Unicode characters — zero-width spaces and friends, literally hiding in your text. They survive copy-paste and end up in your Google Docs, emails, wherever.
Statistical watermarks — the SynthID-Text stuff from the news, where the model's word *choices* form the pattern. Way sneakier.
I couldn't find a tool that handled both, so I built one: (https://textwatermarkremover.org)
- Quick Clean — finds and strips 34+ types of invisible characters, instantly, entirely in your browser. Text never gets uploaded anywhere. Free, unlimited.
- Deep Rewrite — for the statistical kind, it runs your text through an LLM paraphrase so the word-choice pattern breaks.
Would love feedback — especially if you find text it *doesn't* catch.