Blog
How AI watermarking works, who does it, and how to handle it honestly.
How AI watermarking works
The four places a mark can hide: file metadata and C2PA, invisible Unicode, statistical token signals, and pixel-level marks. What each one means if you need to detect or remove it.
Who watermarks what
Claude, Gemini, ChatGPT, Meta, xAI, and open-weight models: what each provider has said publicly about marks in text, images, and audio.
Removing AI watermarks: what actually works
Metadata and invisible Unicode come off cleanly. Statistical text watermarks only yield to heavy rewriting, and that costs quality. What removal can and cannot do.
Thank you to the researchers behind this project
This site strips AI provenance marks — and it only works because of a long line of research and open-source work on watermarking, detection, and removal. A thank-you to every team and author involved.