Anthropic watermark removal tool is now available. Charles Hoskinson published it on Sunday.
He named it Anthropies. It is a blend of Anthropic and herpes. The tool removes Claude’s invisible markings.
How Anthropic watermark removal works
Anthropic switched on invisible marking five days ago. Hoskinson shipped his tool under an Apache 2.0 license. The watermark answers the EU AI Act transparency code.
The rollout took effect on August 2. The mark is not a hidden character string. Instead, a secret key steers tournament sampling. This process breaks ties between fitting words.
Hoskinson’s repository splits the problem into three layers. The “Co-Authored-By” git trailer strips out cleanly. C2PA credentials on images disappear after a re-encode.
Prose is the hard layer. The tool routes text through a model other than Claude. The repository calls this a non-origin rewrite.
The design carries a blunt limitation. Running the rewrite inside Claude would stamp the mark back on. The skill therefore refuses to do this.
The repository ships three modes. Clean strips trailers deterministically. Humanize rewrites prose elsewhere. Orchestrate blocks the job when the host model is Claude.
Code barely carries a signal at all. Syntax leaves almost no room for substitution. The watermark has nowhere to sit.
The legal argument behind Anthropic watermark removal
Hoskinson framed the release as a warning. He is not offering a simple utility. He said the watermark could cause issues later.
Roughly half the repository is not code. Hoskinson attacks the assignment clause. This clause assigns output ownership to users.
The phrase “subject to your compliance with our Terms” matters. He reads it as a condition precedent. Users must clear a hurdle before ownership passes. A breach could mean ownership never transferred.
The git trailer draws separate fire. A published co-authorship line could surface as evidence. It might suggest joint authorship later.
He also flags a risk of false positives. An editor could run human copy through Claude. They might just want a quick proofread. A detector could trip months later.
AI defaults now arrive unasked. Twitch showed this when it opted creators into training. Anthropic has stayed quiet so far. The company is preparing a listing above $2 trillion.
Hoskinson has spent 2026 picking fights. He claimed Ethereum copied Cardano’s ledger design. The repository counted four stars on Sunday morning.
Adoption matters less than the question it poses. Who owns a sentence once the model has signed it?
