Hook: The 4-Star Anomaly
At timestamp 2026-08-16, a GitHub repository named Anthropies appeared with exactly four stars. The ledger shows no prior commits, no forks, no pull requests. The codebase is a single-person project, uploaded by a user with the handle IOHK_Charles. The repository description reads: "A tool to strip Anthropic’s invisible watermark from Claude outputs." Four stars is not a signal of adoption; it is a signal of intent. The anomaly is not the tool itself—it is the legal bomb it carries inside the code.
Context: The Man, the AI, and the Act
Charles Hoskinson, co-founder of Cardano and perpetual provocateur, has spent 2026 in a multi-front war over technical credit. Claiming Ethereum copied Cardano’s ledger design, suing former colleagues, and now launching a tool to break Anthropic’s AI watermark. The target: Anthropic, the $2 trillion valuation AI startup preparing for an IPO. The weapon: an open-source utility released under Apache 2.0. The battlefield: the EU AI Act, which came into effect on August 2, 2026, mandating that AI-generated content be detectable. Anthropic responded with a tournament-sampling watermark—a subtle, statistical fingerprint embedded at generation time. Hoskinson’s answer: a three-layer ripper that claims to erase that fingerprint.
But the code is not the story. The story is the legal argument Hoskinson embedded in the README: a clause-by-clause deconstruction of Anthropic’s Terms of Service, arguing that the very act of using the tool proves the user never owned the output in the first place. The ledger never lies, it only waits to be read—and this time, the ledger is a contract.
Core: The Three-Layer Decomposition and the Non-Origin Rewrite Trap
Let’s audit the technical architecture. The tool, named after the Greek word for "human" (anthropos), uses a three-layer decomposition:
- Layer 1: Co-Authored-By Trailer Removal – A deterministic regex that strips git-style trailer lines from Claude outputs. Anthropic’s API sometimes appends a "Co-authored-by" footer when the model is used in code collaboration. This is trivial, but it signals Hoskinson’s attention to detail.
- Layer 2: C2PA Image Metadata Stripping – C2PA (Coalition for Content Provenance and Authenticity) is a standard for embedding cryptographic provenance into images. The tool re-encodes the image, wiping the metadata. This is robust against simple metadata removal but not against steganographic watermarks embedded in pixel data. The tool does not claim to handle that.
- Layer 3: Prose Rewriting – This is the core. The tool takes the text and routes it through a non-Origin LLM (e.g., another model that does not apply a watermark) to rewrite the content. The critical design insight: performing a rewrite inside Claude or Gemini would re-apply a watermark. So the tool actively detects the host model and refuses to run in "orchestrate" mode on watermarked engines. This is technically honest—it acknowledges the tool cannot operate within the watermarked ecosystem.
From my experience auditing smart contracts, I see the same pattern: the code is the most watermarked-free element. Hoskinson chose to demonstrate the tool on code because code has almost no syntactic substitutability—the watermark signal is minimal. The README proudly shows a diff where the output is identical after stripping. For prose, the tool relies on a third-party LLM to rewrite, which introduces semantic drift. The effectiveness is unknown. The repository has no test suite, no benchmark, no evaluation of text fidelity. The ledger is silent.
The Legal Architecture
Hoskinson’s real innovation is not the code—it’s the legal framing. He points to a line in Anthropic’s Terms: "We assign to you all our rights… to the Output, subject to your compliance with our Terms." He argues this is a condition precedent, not a condition subsequent. In contract law, a condition precedent means ownership never transfers until the condition is satisfied. If the user violates the terms (e.g., by stripping watermarks), the ownership never vested. Therefore, the user never had the right to use the output. The tool is designed to expose the contradiction: if you own the output, why can’t you remove a watermark? If you remove it, did you ever own it?
This is a legal Trojan horse. The code is a tool, but the README is a lawsuit waiting to happen. The Apache 2.0 license ensures that even if Anthropic takes down the repository, forks survive. The legal argument is open-source too.
Contrarian: The Watermark Is Not the Problem – The Contract Is
Here is the counter-intuitive angle: the tool’s technical effectiveness is almost irrelevant. The narrative that Hoskinson is a David fighting Goliath is seductive, but the data tells a different story. The GitHub repository has four stars. The tool has not been independently audited. The rewrite layer depends on a third-party LLM that may itself apply watermarks in the future. The prose layer’s success rate is unmeasured. The entire project is a pre-alpha signal, not a production solution.
But the contrarian view is that the tool’s real value is as a strike against the legal foundation of AI watermarks. If Hoskinson’s interpretation of the Terms of Service is even partially correct, it changes the game for every AI company. Suddenly, the watermark is not a transparency tool—it’s a leash. The user is never truly free. The EU AI Act assumes watermarks are a technical fix, but Hoskinson is arguing they are a contractual trap. The silence from Anthropic (they are busy preparing for a $2 trillion IPO) is telling. They have not responded to the legal argument. Silence in the logs is louder than noise.
Another blind spot: the tool is designed to demonstrate the problem, not solve it. Hoskinson himself called it a "warning," not a utility. The tool’s existence is a rhetorical device. The code is a prop. The real product is the legal argument. The four stars are irrelevant—the number of law professors who download the README is what matters.
Takeaway: The Next Signal
Watch the legal dockets, not the GitHub stars. If Hoskinson’s interpretation gains traction in academic papers or practitioner blogs, AI companies will be forced to rewrite their Terms of Service with explicit ownership language that closes the condition precedent loophole. That would be a win for clarity but a loss for the argument. The tool itself will likely be abandoned within months—it is a one-person project with no sustainability. But the question it raises will outlive the code: who owns the output of an AI that is never truly free? The ledger never lies, but the contracts always do. The next signal is a court ruling, not a commit.