Alpha dropped: Charles Hoskinson just released a tool that peels off Anthropic's AI watermark. Ledger update: Capital is fleeing from the narrative that users truly own AI-generated content.
On August 16, 2026, the Cardano founder dropped a GitHub repository named "Anthropies" โ a free, open-source tool designed to strip the invisible watermark Anthropic embeds in Claude's outputs. The move is a direct response to the EU AI Act's transparency mandate, which went into effect on August 2. But beneath the surface of a simple code release lies a sophisticated legal and technical attack on the entire AI service provider model.
Context: Why Now? Anthropic's watermark uses a technique called "key-guided tournament sampling" โ a statistical bias injected at generation time to create a detectable pattern across the entire text. Unlike traditional post-hoc watermarking, this method is robust: it's not simply a string you can delete. The EU AI Act requires that AI-generated content be machine-detectable, and Anthropic's implementation is among the most advanced. Hoskinson, never one to stay out of a fight, saw an opportunity to challenge the legal foundations of this ecosystem.
The tool is barely a day old, with only 4 GitHub stars at the time of writing. Its code is raw, pre-alpha. Yet the narrative is already bigger than the codebase. Hoskinson is not just releasing a technical bypass; he's releasing a legal argument wrapped in open-source licensing.
Core: The Technical and Legal Architecture Hoskinson's tool operates on three layers. First, it strips the "Co-Authored-By" git trailer โ a metadata tag that Claude sometimes appends to code. Second, it removes C2PA image metadata by re-encoding the file. Third, and most importantly, it attempts to remove the statistical watermark from prose by routing the text through a third-party LLM that does not itself apply a watermark. This "non-origin rewrite" is the key insight: the tool refuses to run on Anthropic, Google, or any model known to apply watermarks, because doing so would simply re-stamp the output.
This is technically honest but also a limitation. The tool's effectiveness is highest on code โ where there is almost no statistical room for watermarking โ and lowest on prose, where the rewrite can alter meaning. In my experience auditing over a dozen AI-crypto hybrids, this kind of selective demonstration often masks the tool's real weakness: it works best where it's least needed, and worst where it's most needed.
But the real core of the story is legal. Hoskinson's accompanying X thread argues that Anthropic's Terms of Service transfer ownership of the output "subject to your compliance with our Terms." He interprets this as a condition precedent: if you violate the terms (e.g., by stripping the watermark), the ownership never transfers. This means millions of Claude users may never have actually owned their outputs. The argument is creative, untested in court, but built on solid contract law principles. He then released the tool under Apache 2.0, which includes an explicit patent grant โ effectively blocking Anthropic from using patent lawsuits to shut down forks.
Contrarian: The Tool Is a Signal, Not a Solution The mainstream coverage will frame this as a David vs. Goliath story โ a lone developer poking a $2 trillion IPO candidate. But the contrarian view is that the tool is a narrative trap. Hoskinson is not solving a problem; he is exposing a paradox. The EU AI Act mandates watermarking, but the same watermarks may void the user's ownership rights. The tool is a rhetorical device, not a scalable utility.
Three blind spots emerge. First, the tool's reliance on a third-party LLM for rewriting creates a new dependency. If the only available non-watermarking model is a smaller, less capable one, the output quality degrades. Second, the legal argument cuts both ways. If a court accepts Hoskinson's interpretation, it could lead to a cascade of retroactive ownership disputes โ but it could also prompt AI companies to rewrite their Terms with ironclad language, closing the loophole permanently. Third, the tool could be weaponized for fraud. By stripping the AI-generated identifier, bad actors can pass off synthetic content as human-written, undermining the very transparency the EU AI Act seeks to enforce. Hoskinson has added no usage restrictions, leaving the tool open to abuse.
Ledger update: Capital is fleeing from the notion that AI output ownership is settled. The market has not priced this risk. Anthropic remains silent, likely because any response would legitimize the debate. But if the legal argument gains traction among academics or regulators, the entire service model of "you own your output" will require renegotiation. The next move is not technical โ it's legal. Watch for AI companies to quietly update their Terms of Service to sever the link between watermark compliance and ownership transfer.
Takeaway: The Battle Has Shifted from Code to Contract The real impact of Anthropies is not in its ability to strip watermarks, but in its ability to frame the conversation around terms of service. Hoskinson has turned a technical compliance issue into a property rights dispute. The question now is not whether the watermark can be removed, but whether the output was ever yours to begin with. Alpha dropped: Follow the money. The next wave of AI regulation will be written in contracts, not code.