I didn't expect Charles Hoskinson to pivot from blockchain governance disputes to AI watermark sabotage. But the forensic evidence is on GitHub: a three-layer tool called Anthropies, released on August 16, 2026, designed to strip Anthropic's invisible watermarks from Claude outputs. The timing is precise—six weeks after the EU AI Act's transparency mandates took effect. The codebase has four stars. The claim is bold. The technical reality is more nuanced than the headlines suggest.
Context: The Watermark War and a Founder's Sidestep
Anthropic's watermark, deployed in July 2026, uses a key-guided tournament sampling mechanism. It doesn't embed a hidden string; it injects a detectable statistical pattern into the token probability distribution at generation time. This is the gold standard for AI provenance—no post-hoc metadata, no fragile hashes. The EU AI Act (Article 50) requires AI output to be machine-detectable, and Anthropic's solution is the most technically mature implementation.
Hoskinson, fresh off a year-long argument about whether Ethereum copied Cardano's ledger design, launched Anthropies as a free open-source tool. The name is a pun on Anthropic and “anthropies” (sufferings). The tool's GitHub README frames it as a warning, not a utility. But the code tells a different story: it's a deliberate engineering response to a specific cryptographic compliance measure.
Core: The Technical Teardown – Three Layers, One Weak Link
The architecture is deceptively simple. Layer 1 strips git trailer co-authorship metadata—a deterministic operation with zero text alteration. Layer 2 removes C2PA image credentials by re-encoding the file. These are trivial bypasses for metadata, not content. The real challenge is Layer 3: prose watermark removal.
Here's the bottleneck wasn't a lack of algorithms—it was the constraint that rewriting text inside Claude or Bard re-applies the watermark. Hoskinson's solution is a “non-origin rewrite”: the tool detects the host model and refuses to execute on any model that applies watermarks. Instead, it routes the text through an external LLM (e.g., a local model or another API) to generate a new distribution that dilutes the original statistical signal.
Flash loans don't care about text semantics, but watermark removal does. The tool's effectiveness on prose is unknown. The codebase is pre-alpha: no performance metrics, no stress tests, no independent audit. The sole demonstration is on code—a structurally rigid medium where watermark signals are inherently weak. Code has minimal synonym replacement space and low statistical variance. Hoskinson chose the easiest target to make the tool look effective. This is a classic engineering deception: optimize for the demo, not the use case.
The Legal Trojan Horse
Anthropies' most dangerous component isn't the code—it's the legal argument embedded in the README. Hoskinson points to Anthropic's Terms of Service: “We assign to you all our rights, title, and interest in and to the Output... subject to your compliance with our Terms.” The phrase “subject to your compliance” is a condition precedent. If the user violates any term (e.g., by removing a watermark), the ownership assignment never executes. The user never actually owns the output.
This is a textbook contractual interpretation. It's not a settled legal principle, but it's plausible. If enough lawyers and academics adopt this reading, it forces AI companies to rewrite their ToS with explicit language like “ownership vests immediately regardless of compliance.” The result would weaken the legal basis for watermark enforcement, because the company would be admitting that the current language is ambiguous.
You don't need a courtroom to win this battle—you need to make the counterparty's drafting look sloppy. Hoskinson understands this. The Apache 2.0 license on Anthropies is a deliberate legal shield: it grants patent rights and prevents Anthropic from suing over derivative works. The tool becomes a public good that cannot be killed by a single lawsuit.
Contrarian: What the Bulls Got Right
Critics will dismiss Anthropies as a stunt—four stars, one day old, no real users. But the contrarian angle is that the tool's technical weakness (prose removal) is irrelevant to its strategic value. The real target is the legal and regulatory framework. Hoskinson is not trying to help writers evade detection; he's trying to expose the inconsistency in AI company contracts and the EU's reliance on a fragile watermark regime.
The bulls also have a point about the code's honest self-limitation. The “orchestrate” mode that refuses to rewrite on watermarked models is technically correct. It's a rare admission of defeat in open-source activism. The tool acknowledges its own boundaries—unlike many crypto projects that promise infinite scalability.
Furthermore, the timing is strategic. Anthropic is preparing for an IPO at a $2 trillion valuation. Any legal uncertainty about output ownership creates risk for institutional investors. Hoskinson's move is a form of adversarial due diligence—the kind of scrutiny that gets ignored in bull markets but matters when the SEC starts asking questions.
Takeaway: The Arms Race Has a New Front
Anthropies won't kill AI watermarks. But it will force the industry to evolve—both technically and legally. Expect a new generation of human-in-the-loop watermark detectors that analyze writing style rather than token statistics. Expect AI companies to renegotiate their ToS with greater precision. Expect the EU AI Act's enforcement guidelines to include anti-circumvention clauses.
The real question is not whether Hoskinson's tool works—it's whether the market will price in the legal risk that his argument exposes. The answer is probably no, until a lawsuit or a regulatory action proves otherwise. Until then, the four stars on GitHub are a canary in the coal mine of AI governance. I didn't expect to write that sentence about a tool that strips metadata from git trailers. But that's the nature of forensic analysis: you follow the data, not the hype.
