TLDR:
- Anthropies targets Claude’s keyed text watermark by changing wording through an unmarked rewriting model.
- The project also removes C2PA credentials, Git attribution trailers and other Claude-related metadata.
- Anthropies says watermark detection indicates Claude contact but cannot independently establish who authored text.
- Hoskinson’s repository links AI watermarking with wider questions about ownership, attribution and digital provenance.
Charles Hoskinson has launched Anthropies, a tool targeting watermarking and attribution markers attached to Claude outputs. The project aims to remove Claude’s keyed text watermark through rewriting with models that do not carry the same mark.
Anthropies also targets C2PA image credentials and Claude-related Git attribution trailers. The launch frames AI provenance as a growing issue around ownership, authorship and digital content.
Anthropies Targets Claude AI Watermarks and Metadata
According to the Anthropies repository, Claude can apply three separate markers across text, images and software projects. The text watermark changes token selection without adding visible characters to generated content.
The repository says a secret key determines how Claude selects between similarly suitable words during generation. Anthropies describes this watermark as difficult to remove without changing the underlying wording.
The project also addresses C2PA credentials attached to supported image files. Those credentials contain signed metadata indicating that Claude processed an asset. The repository says re-encoding or stripping metadata can remove that credential from the file.
A third marker appears through Claude Code’s Git commit trailers.
The repository says Claude Code can add a “Co-Authored-By” entry identifying Claude in commit history. GitHub can then parse that trailer and display Claude as a contributor.
How Anthropies Removes Claude Attribution
Anthropies provides a command-line interface alongside an agent skill called /purge-anthropies. Its cleaning function removes attribution banners, Git trailers and invisible Unicode characters. Its humanization function instead changes prose structure through a separate rewriting model.
The repository specifically advises users against rewriting Claude output with Claude itself. It says doing so could preserve or reapply the same watermarking mechanism.
Instead, Anthropies recommends an unmarked model for the structure-changing rewrite process.
The project also presents detection limits for Claude’s text watermark. Its documentation reports stronger detection rates on longer passages, while shorter text remains harder to identify. It also stresses that detection indicates system contact rather than proving authorship.
Hoskinson’s repository connects the technical system to broader questions about AI ownership and attribution. However, those legal sections represent the project’s argument rather than an established court ruling on Anthropies. The repository itself states that its legal discussion constitutes argument rather than legal advice.
