Anthropies: A Four-Star Warning on AI Watermarking and Ownership

A GitHub repository with four stars. A single commit. A founder's name. That is all it took to ignite a debate on AI watermarking and ownership. On August 16, 2026, Charles Hoskinson, the founder of Cardano, released Anthropies—an open-source tool designed to strip watermarks from Anthropic's Claude outputs. The codebase is sparse. The claims are ambitious. The data, however, tells a more nuanced story. Data does not lie; it only reveals hidden patterns. The pattern here is a mismatch between narrative heat and technical substance. Hoskinson's tool targets a specific watermarking scheme—Anthropic's tournament sampling—which injects statistical biases into token selection. The tool decomposes the problem into three layers: removing git trailers, stripping C2PA metadata, and rewriting prose through a non-origin LLM. The first two layers are deterministic. The third is the crux. From my 2017 audit of ERC-20 token contracts, I learned that claims of simplicity often hide complexity. The prose layer requires rerouting text through an external LLM to break the watermark signal. This introduces a fidelity trade-off. The tool's own documentation admits that prose is the 'hard layer.' The code snippet demonstrates that the tool actively avoids rewriting within the same model ecosystem—a design choice that acknowledges the limits of its own method. Context: Anthropic's watermarking is a response to the EU AI Act's transparency requirements. The tool's legal argument is more potent than its technical one. Hoskinson reads Anthropic's terms of service—'subject to your compliance with our Terms'—as a condition precedent. If a user violates the terms by stripping watermarks, ownership never transfers. This is a contractual interpretation, not a court ruling. But it exposes a structural ambiguity in how AI companies assign ownership of outputs. Core analysis: The adoption metric is the critical signal. Four stars on GitHub after one day. Not zero, but negligible. The repository has no forks, no issues, no pull requests. The download count is unavailable. The tool's functionality is unverified by independent parties. My experience with the 2020 Uniswap V2 liquidity mapping taught me that early adoption curves are unreliable. But here, the curve is flat. Contrarian angle: The tool's effectiveness is inversely correlated with its utility. It works best on code—where watermarks are least needed. It works least on prose—where watermarks matter most. This is not a bug; it is a feature of the tool's design. Hoskinson chose code as the demo because it is the easiest target. The prose layer remains unproven. Correlation does not equal causation: the tool's legal argument does not make it technically effective. The legal argument itself is untested. No court has ruled on the condition precedent interpretation. The narrative of 'David vs Goliath' obscures the data: a single developer against a company preparing for a $2 trillion IPO. The odds are not in favor of the tool's longevity. Takeaway: The next-week signal is not the tool's adoption. It is Anthropic's response—or silence. If Anthropic ignores the tool, the narrative collapses. If they respond, the debate escalates. The data suggests silence is the rational choice. The tool's impact is likely to be a footnote in the larger AI governance debate. The real takeaway is a lesson in data literacy: not every four-star repository is a revolution. Some are just warnings. And warnings, if ignored, become patterns. Based on my audit of the 2024 Bitcoin ETF inflows, institutional behavior is slow to change. The same applies here. The legal and technical infrastructure for AI watermarking will evolve, but not because of a single tool. The data does not lie: the tool's current state is a proof of concept, not a product. The hidden pattern is that Hoskinson is repositioning himself from blockchain founder to AI governance critic. This tool is his entry point. The question is whether the market will follow him.