The Anthropies Paradox: Hoskinson’s Weapon Against AI Watermarks and the War Over Ownership

We assume that watermarking AI outputs is a technical solution to a regulatory problem. But beneath the surface of that assumption lies a legal and philosophical quagmire that no algorithm can solve. On August 16, 2026, Cardano founder Charles Hoskinson released a free, open-source tool called Anthropies—a name that reads like a taunt aimed at Anthropic. The tool claims to strip the invisible watermark that Anthropic embeds in outputs from its Claude model. Within hours, the GitHub repository had four stars. The market yawned. Yet this unassuming code release might be the most consequential signal in the AI-crypto intersection this year. We are hunting for truth in a mirror maze of hype. The context is essential. Anthropic’s watermarking scheme, known as “tournament sampling,” was developed in response to the EU AI Act, which came into effect on August 2, 2026. The regulation requires that AI-generated content be machine-detectable. Anthropic’s approach is elegant: instead of appending a hidden string after generation, it biases the token selection process during sampling—a cryptographic key guides the model to break ties between equally plausible words in a way that leaves a statistical fingerprint. This is not a sticker on the output; it is woven into the fabric of the text. Hoskinson, a veteran of the crypto wars—he co-founded Ethereum and later created Cardano—saw an opening. His tool, released under the Apache 2.0 license, decomposes the watermark problem into three layers: removing co-authored-by git trailers, stripping C2PA metadata from images, and, most critically, rewriting prose through a non-origin large language model (LLM) to erase the statistical signal. The core of the analysis lies in the mechanics. Hoskinson’s approach is not a novel cryptographic breakthrough; it is a clever engineering composition. The so-called “orchestrate” mode is the key insight: if you attempt to rewrite a Claude output inside Claude itself, you simply re-apply the watermark. The tool detects the host model and refuses to execute on the same platform. Instead, it routes the text to a third-party LLM—say, a local model or a competing API—to regenerate the content, thus breaking the statistical pattern. This is technically sound, but it carries a hidden cost. The ledger remembers what the heart forgets. The tool is most effective on code, because code has minimal syntactic variation and thus little room for watermarking. Hoskinson chose code as his primary demo, which is revealing. Based on my experience auditing code repositories during the 2017 ICO mania, I have seen how easy it is to overstate the effectiveness of a tool when the test case is cherry-picked. The prose layer—the domain where watermarking matters most—remains the hardest. The tool’s reliance on an external, potentially watermarked model for rewriting introduces a chain of dependencies that undermines its claim of universal removal. The tool’s effectiveness is inversely proportional to the value of the content it seeks to protect. But the contrarian angle is more subtle. The real significance of Anthropies is not technical but legal. Hoskinson has woven a legal argument into the codebase: he points to the fine print in Anthropic’s terms of service, which state that output ownership is transferred “subject to your compliance with our Terms.” He interprets this as a condition precedent—meaning that if a user violates the terms (for example, by stripping a watermark), the ownership never actually transfers. This is a lawyer’s trap, not a hacker’s tool. The contrarian view is that the tool is a Trojan horse for a larger narrative about control. It is not designed to help the average user remove watermarks; it is designed to frame the debate around who truly owns AI-generated content. The tool’s very existence serves as a demonstration that the watermark can be removed, thereby proving that the user’s “ownership” is conditional and fragile. This is a powerful rhetorical move, but it also invites a backlash. If Anthropic or other firms respond by tightening their terms or by using more robust, non-removable watermarking techniques, the end result could be a more locked-down ecosystem, not a freer one. The tool might accelerate the very centralization it claims to oppose. The ledger remembers what the heart forgets. The takeaway is that Hoskinson’s Anthropies is a mirror, not a solution. It reflects the deep tension between the promise of AI as a tool for individual empowerment and the reality of corporate control over the means of generation. The narrative will evolve not through code, but through the courtroom and the regulatory body. The key question is not whether the tool works—it is whether the legal argument will be taken seriously by courts. If it is, we may see a wave of challenges to AI service terms. If it is not, the tool will fade into the history of crypto curiosities. Either way, the story has just begun. And we are still hunting for truth in a mirror maze of hype.