An Open-Source Tool Just Went After Every Major AI Watermark — But Even Its Own Docs Admit It Can't Prove It Works

watermarks-remover installs as an agent skill and strips AI provenance marks across Claude, Gemini/SynthID and OpenAI. Layer A removes invisible Unicode and bidi characters. Layer B rewrites text to break statistical watermarks. File cleaning covers C2PA, EXIF and XMP in PNG, JPEG, SVG, PDF, DOCX, HTML and Markdown. Core scripts are Python 3.10+ stdlib. Soft binding and audio watermarks stay out of scope. 6.6k stars, MIT.
#AIGovernance #C2PA #ContentProvenance #AIAgents

An open-source project called watermarks-remover packages itself as an installable agent skill that strips AI-content provenance signals across Claude, Gemini/SynthID, and OpenAI outputs. It works in two text layers — deterministic removal of invisible Unicode/bidi characters, plus an optional AI-assisted rewrite meant to disrupt statistical (token-sampling) watermarks — alongside file-level cleaners that strip C2PA, EXIF, and XMP metadata from PNG, JPEG, WebP, SVG, PDF, DOCX, ODT, HTML, and Markdown. The core is Python 3.10+ stdlib with no dependencies; heavier optional add-ons for pixel-level image watermark removal rely on separately licensed third-party code. The project's own documentation is notably candid about limits: the rewrite layer is "best-effort" only, degrades writing quality, and — since no vendor has published a public detector — the maintainers say no tool can currently prove it defeats an official check. It's MIT-licensed and has crossed 10k GitHub stars.