A tool with 4 GitHub stars just challenged the legal foundation of a $2 trillion AI IPO.
On August 16, 2026, Charles Hoskinson released "Anthropies" — a free, open-source utility designed to strip Anthropic's invisible watermark from Claude outputs. The codebase is live. The GitHub stars are negligible. The implications are anything but.

This is not a story about adoption metrics. It is a story about how a single legal argument, wrapped in a technical package, can force the entire AI industry to rewrite its terms of service.
Context: The Watermark That Wasn't Supposed to Be Removable
Anthropic's watermark is not a hidden string injected after generation. It is a structural bias embedded during the sampling process — a tournament-sampling mechanism that tilts the probability distribution toward certain tokens without sacrificing quality. Approved by the EU AI Act for transparency, it was designed to be cryptographically robust.
Hoskinson's response: a three-layer decomposition that treats the watermark not as a security feature, but as a contractual liability.
Layer 1 strips git trailer co-authorship markers. Layer 2 re-encodes C2PA image metadata. Layer 3 — the critical one — routes text through a non-origin LLM to dilute the statistical signal. The orchestrate mode actively refuses to rewrite on watermarked models, because rewriting on Claude means re-applying the same watermark. This is technically honest, but it also reveals the tool's limitation: it cannot operate within the ecosystem it purports to defeat.
Core: The Architecture of an Argument
The tool's technical design is clever, but its real architecture is legal.
Hoskinson's key insight: Anthropic's terms of service state that output ownership is "subject to your compliance with our Terms." He interprets this as a condition precedent — a legal barrier that must be satisfied before ownership transfers. If a user violates the terms (e.g., by stripping the watermark), then ownership never legally transferred. The output was never truly theirs.
This turns the tool into a legal weapon. Use it, and you may be proving that you never owned the content in the first place. The tool itself becomes evidence of a contractual breach.
Narrative is the new liquidity. Hoskinson understands this better than most. He is not selling a tool; he is selling a story — one where the AI company's promise of ownership is a mirage, and the only way to prove ownership is to break the terms that supposedly guarantee it.
Contrarian: The Tool's Weakness Is Its Strength
Here is the counter-intuitive angle: the tool's technical weakness — its poor performance on prose — is actually its strategic advantage. By focusing on code, where the watermark signal is weakest, Hoskinson ensures a high success rate for the easiest use case. This makes the tool appear more effective than it is, which amplifies the legal narrative.
But the contrarian blind spot is this: if the tool gains traction, AI companies will simply modify their terms to include "regardless of compliance" clauses. The condition precedent argument will be patched. The legal window may close before the technical one opens.
Hype decays; utility endures. The tool's utility is not in its code — it is in the debate it forces. The 4 stars on GitHub are a feature, not a bug. They signal that this is a provocation, not a product. The real audience is not the developer who forks it, but the regulator who reads about it.
Takeaway: The Next Battle Is Over Ownership, Not Watermarks
The Anthropies incident is a prelude to a larger conflict. The AI watermark arms race will escalate, but the decisive battles will be fought in courtrooms and contract law, not in pull requests. The question is not whether watermarks can be removed, but whether the output was ever owned by the user at all.

Code talks, but stories sell. Hoskinson's story — that AI companies are trapping users in a web of contradictory promises — is more dangerous to Anthropic than any technical exploit. The next step is not a better tool. It is a class-action lawsuit citing the terms of service themselves.
And the 4-star GitHub repo? It will be Exhibit A.