On August 16, 2026, Charles Hoskinson uploaded a GitHub repository. It had 4 stars. That number is a camouflage. Behind it lies a surgical strike on the legal architecture of AI-generated content ownership.
The repository is called Anthropies—a free, open-source tool designed to strip the invisible watermark Anthropic embeds in every Claude output. The European Union’s AI Act, effective August 2, 2026, mandates that AI-generated content be detectable. Anthropic complied. Hoskinson responded with a tool that claims to reverse that compliance.
But the tool is not the story. The story is the legal argument Hoskinson has woven into its code and its license. He is not just building a technical bypass. He is building a case that the very act of watermarking may void the ownership rights of every user who has ever clicked “Accept” on Anthropic’s terms of service.
Context: The Anatomy of the Attack
Hoskinson’s Anthropies is a three-layer tool. Layer one strips git trailer co-author metadata—a deterministic, trivial operation. Layer two removes C2PA image credentials by re-encoding. Layer three targets the prose itself: the statistical watermark that Anthropic calls “tournament sampling.” This is the hard layer. Hoskinson acknowledges it as the “difficulty layer” in his documentation.
The method for the prose layer is non-origin rewrite. The tool routes the watermarked text through a third-party LLM—not Claude, not Gemini—to generate a new version that carries a different statistical fingerprint. The design insight is counterintuitive: rewriting inside the same model ecosystem re-applies the watermark. So the tool actively detects the host model and refuses to execute if it detects a watermarked source. This is honest engineering. It admits the tool cannot operate inside the walled garden it is attacking.
But the real weight of the project lies not in the code but in the license. Hoskinson chose Apache 2.0. That license grants explicit patent rights and allows anyone to fork, modify, and sell the tool. It is a legal shield. If Anthropic tries to sue the tool out of existence, the code can survive as a thousand forks. Hoskinson has seen this play before. In 2017, I watched ICO teams bury poison pills in their whitepapers. This is the same instinct—pre-empt the legal counterattack before it lands.
Core: The Condition Precedent Trap
Here is the insight that separates this event from a simple open-source spat. Hoskinson’s X post lays out a contractual argument. Anthropic’s terms of service state: “To the extent permitted by applicable law, we assign to you all rights, title, and interest in and to the output, subject to your compliance with our Terms.” The phrase “subject to your compliance” is the trap.
Hoskinson interprets this as a condition precedent. In contract law, a condition precedent is an event that must occur before a duty to perform arises. If the user fails to comply with any term—including, potentially, the prohibition against removing watermarks—then the ownership transfer never occurs. The output never becomes the user’s property.
This is not a hypothetical. The watermarks are designed to be detectable. If Anthropic ever detects that a user has stripped the watermark, it could argue that the user violated the terms and therefore never owned the output. The user could be liable for copyright infringement of their own work. The tool itself becomes evidence of the violation.
Hoskinson’s argument is not settled law. No court has ruled on this interpretation. But the argument is plausible enough to worry Anthropic’s legal team. And in a world where the company is reportedly preparing for an IPO at a $2 trillion valuation, legal uncertainty is a liability they cannot afford.
During the 2020 DeFi summer, I learned that the most profitable trades are not based on technical innovation alone. They are based on structural inefficiencies in the rules of the game. Hoskinson has found a structural inefficiency in the contract between AI companies and their users. The watermarks are not just a compliance tool. They are a surveillance mechanism that can retroactively void the user’s ownership. The tool is a mirror that reflects that vulnerability back to the company.
Contrarian: The Tool’s Real Value Is Not Technical
The market will interpret Anthropies as a technical tool. It will be compared to GPTZero or other detection systems. That comparison is a category error. The tool’s technical effectiveness is secondary. Its primary function is to shift the narrative from “AI watermarking is a benevolent transparency measure” to “AI watermarking is a contractual trap that allows companies to retain control over your work.”
This is the contrarian view that the market is not pricing. The 4-star GitHub repository is a Trojan horse. The code is the shell. The legal argument is the payload. Hoskinson is not building a product. He is building a precedent. He is testing whether the open-source community can weaponize contract law against the very companies that provide the infrastructure.
There is a hidden vulnerability in Hoskinson’s own strategy. He chose code as the primary demonstration because code carries almost no watermark signal. The tool’s effectiveness on prose is unproven. If the prose layer fails, the entire narrative collapses. But that failure does not weaken the legal argument. The legal argument stands on the service terms alone. The tool is just a prop. The real weapon is the condition precedent interpretation.
I have seen this pattern before. In 2022, during the Terra collapse, I had to execute a market sell order at a 60% loss to preserve capital. The rule was: when the protocol’s foundation cracks, do not wait for consensus. Hoskinson is applying the same rule to a legal framework. He is not waiting for a court to rule. He is publishing the argument now, knowing that the market will react to the uncertainty faster than the legal system can resolve it.
Takeaway: The Watermark Is the Canary
Anthropies is not a tool you will use. It is a signal you should read. The next six months will determine whether the condition precedent argument becomes a standard clause in all AI service terms or a footnote in a settlement. The EU AI Act’s enforcement agencies are watching. If they decide that watermark stripping is a violation of the transparency mandate, Hoskinson’s tool becomes illegal. But if they decide that the watermarks themselves create a contractual unfairness, the entire compliance framework could shift.
The code is law, but the governance vote can kill it. In this case, the governance vote is not a DAO. It is a court. And the court does not care about GitHub stars. It cares about the words on the page. Hoskinson has written those words. Now we wait to see if the judgment matches the narrative.
Ledgers don’t lie. But contracts do. The question is: which side of the contract are you claiming ownership from?