The Watermark Arms Race Has a New Contender — and It's Open Source

5K stars in less than 24 hours… and growing. 🤯
Well, that didn’t take long.
AI labs are starting to add more provenance signals to AI-generated content.
Open source responded immediately:
watermarks-remover.
A new repo designed to strip multiple types of AI provenance marks across ecosystems including Claude, Gemini and OpenAI.
It can tackle:
→ Invisible Unicode characters
→ C2PA / EXIF / XMP metadata
→ Document properties
→ Statistical text watermarks through rewriting
And it works across PNG, JPEG, PDF, DOCX, HTML, Markdown and more.
The cat-and-mouse game between AI provenance and watermark removal has officially started.
And open source moves ridiculously fast.

Fact-check results — the core claim holds up, with some nuance:

Real repo, real growth. guillaumemeyer/watermarks-remover is a real, actively maintained MIT-licensed GitHub project by developer Guillaume Meyer. It's now at roughly 10,000 stars (as of today), so the "5K in <24 hours" claim about its initial spike is plausible — Bleeping Computer reported it had passed 4,500 stars just a day after Anthropic's watermarking announcement, and it's roughly doubled since.

What it actually does — mostly accurate, one exaggeration:

• ✅ Strips invisible Unicode characters (zero-width chars, bidi controls, etc.) — this is verifiable and works cleanly.
• ✅ Strips C2PA/EXIF/XMP metadata and document properties from PNG, JPEG, WebP, SVG, PDF, DOCX, ODT, HTML, and Markdown — also verifiable.
• ⚠️ "Statistical text watermarks through rewriting" is overstated as a capability. The tool's own README is explicit that this layer is "best-effort," not a guarantee — it works by heavily rewriting text to disrupt token-choice patterns, which necessarily degrades tone/voice/precision. The README itself says nobody can currently verify this works against Claude's actual watermark, because Anthropic hasn't published a public detector.
• The repo also explicitly states pixel-based image watermarks and "training backdoors" are out of scope for its core tool (optional external add-ons exist for some image watermarks).

Context the post leaves out: The repo emerged shortly after Anthropic disclosed it had turned on invisible provenance marks across Claude's text output (an EU AI Act compliance move), and similar tools/services (StealthGPT, etc.) popped up around the same time. Coverage from outlets like Bleeping Computer noted that none of the "defeats the watermark" claims can currently be independently verified, since none of the major labs have released public detectors to test against.

So: real project, real star count trajectory, genuine capability on metadata/Unicode — but the "cracks AI watermarks" framing overstates what's actually been proven for the statistical text-watermark layer.