Cardano Founder Launches New Free Tool to Remove Anthropic’s AI Watermark

Yield wasn't the only thing being harvested that summer. As August 2026 unfolded, the crypto and AI worlds collided in a way that felt less like a merger and more like a territorial dispute. Charles Hoskinson, the Cardano founder who has spent the year sparring over technical credit and Ethereum’s ledger design, dropped a new open-source tool called Anthropies. Its stated purpose: strip the invisible watermark that Anthropic embeds into every Claude output. But the real payload wasn't in the code—it was in the legal logic that accompanied it. Context: This isn't just another GitHub repo. Anthropies arrives days after the EU AI Act’s transparency provisions took effect on August 2, requiring AI providers to make generated content machine-detectable. Anthropic had already deployed a "key-guided tournament sampling" watermark—a statistical fingerprint woven into the text’s probability distribution, not a hidden string. Hoskinson’s tool claims to remove it using a three-layer decomposition: stripping git trailers, re-encoding C2PA image metadata, and—most critically—rewriting prose via a third-party LLM router that consciously avoids re-watermarking. Core: The technical architecture is where Anthropies reveals its true character. The "non-origin rewrite" mode is the cleverest part: the tool detects if the host model is itself watermarked (like Claude or Bard) and refuses to execute the rewrite there, because doing so would simply re-apply the mark. Instead, it routes the text to an unwatermarked model—a design that acknowledges the limits of adversarial action. But the real signal is in the choice of demonstration. The code layer is almost immune to watermarking by design; there’s little syntactic variance to exploit. Hoskinson chose code as the showcase because it’s the easiest case. Prose, the layer where watermarking matters most, is explicitly labeled the "hard layer." The tool’s effectiveness on natural language remains unverified, and the trade-off between fidelity and watermark removal is sharp. This is a project at the concept stage—four stars on GitHub, one day old, no independent audit. The math of secrets is delicate here: the probability distribution shift introduced by a third-party rewrite can dilute the signal, but at the cost of semantic drift. The truth protocol is not yet written. Contrarian: The prevailing narrative frames this as a David vs. Goliath technical victory—a lone developer cracking the AI giant’s watermark. But the real story is legal. Hoskinson’s accompanying thread attacks Anthropic’s terms of service, arguing that the phrase "subject to your compliance with our Terms" converts the output ownership transfer into a condition precedent. If a user violates the terms (e.g., by using a watermark removal tool), the ownership never actually transfers. This is a clever reading, but it’s untested in court. The tool’s Apache 2.0 license is a deliberate legal shield, preventing Anthropic from using patents to block forks. Yet the technical substance is thin. The tool is more of a "warning"—Hoskinson’s own word—than a production utility. The risk is that the narrative overstates the tool’s capability, creating a false sense of security for users who need real watermark removal for sensitive content. The next pivot is already in motion: this is not a tool for the masses, but a signal to regulators and lawyers that the AI watermarking regime has a vulnerability—not in the math, but in the contract. Takeaway: Anthropies will not change the watermark arms race overnight. But it will force a conversation that the industry has been avoiding: if an AI company can embed a permanent identifier in your output, do you truly own what you create? The legal argument may outlast the code. The question is not whether the tool works, but whether the terms of service can survive the scrutiny of a generation that has been burned by fine print. Yield wasn’t the only thing being harvested—trust was, too.