We didn’t just hunt alpha; we rewired the game. But this time, the game isn’t on-chain—it’s in the invisible statistical fingerprints Anthropic embeds into every Claude output. When Charles Hoskinson dropped his free, open-source tool Anthropies on August 16, 2026, the crypto-native crowd shrugged. Another personal project from a founder known for picking fights? Yet as someone who’s spent nights in Jakarta’s co-working spaces auditing smart contracts and dissecting Terra’s collapse, I saw something else: a carefully crafted legal–technical hybrid that could reshape how we think about AI output ownership.
From core dev trenches to community heartbeat, I’ve learned that the most dangerous moves aren’t the ones that break code—they’re the ones that break assumptions. Hoskinson’s Anthropies does exactly that. It doesn’t just remove a watermark; it questions the very foundation of “you own your AI output.”
The Hook: A 4-Star GitHub Repo That Shook an AI Giant
On August 16, 2026, Charles Hoskinson—Cardano’s co-founder and perennial provocateur—dropped a GitHub repo called Anthropies (a portmanteau of “Anthropic” and “antithesis” or “anthropology”). The repo had 4 stars. It was a day old. It claimed to strip the invisible watermark Anthropic has been embedding into Claude’s outputs since early 2025. The tool is free, Apache 2.0 licensed, and comes with a legal memo that reads like a contract law exam.
Here’s the kicker: Hoskinson isn’t selling anything. No token, no paid tier, no “ecosystem fund.” He’s simply saying: “Anthropic’s watermark is a leash. I’m giving you the scissors.” And the industry is still trying to figure out if that leash was ever legally attached.
Context: The Watermark That Wasn’t Supposed to Be Removable
Anthropic’s watermark—officially called “tournament sampling with a key-guided bias”—isn’t a visible logo or a hidden string. It’s a statistical pattern injected into the word choice during generation. When Claude picks between two equally plausible tokens, it nudges the selection toward a cryptographically seeded “winner,” creating a detectable distribution that only Anthropic’s server can verify. The EU’s AI Act, which came into effect August 2, 2026, mandates such machine-detectable markers for AI-generated content. Anthropic complied.
But compliance creates a new tension: if the watermark is permanent, who really owns the output? Can Anthropic still track a text after it’s been rewritten? And if a user removes the watermark, are they violating the terms of service—or exercising their property rights?
Hoskinson’s tool, Anthropies, is his answer to those questions. It’s a three-layer decomposition:
- Layer 1: Co-Authored-By – Removes the git trailer that some AI tools append to code. Simple, deterministic, works.
- Layer 2: C2PA Image Metadata – Strips cryptographic credentials embedded in images. Re-encoding eliminates them.
- Layer 3: Prose – The hard one. Anonymizes the statistical watermark by routing the text through a non-origin LLM (i.e., not Claude, not Bard) to rewrite it, effectively breaking the pattern.
Core: Why the Technical Design Matters More Than the Code
Education is the new mining rig for the mind. And Anthropies, despite its early alpha state, teaches us three things about the state of AI watermarking.
First, the “non-origin rewrite” hack is a design confession. Hoskinson’s tool explicitly refuses to run on top of Claude or other watermarked models. It detects the host model and rejects the request if it’s an Anthropic endpoint. Why? Because rewriting inside a watermarked model just re-applies the watermark. That’s a brutally honest admission: you can’t remove the watermark from within the system that created it. You need an external, unmarked LLM to dilute the signal.
Second, code is the easy win, prose is the battlefield. The tool’s demo showcases code—because code has almost no word-choice redundancy. The syntax is fixed, the tokens are deterministic, and watermarks have nowhere to hide. Hoskinson’s choice to highlight code first is a smart move for a proof-of-concept, but it also masks a critical weakness: for natural language, the success rate is unknown. The watermark signal is distributed across synonyms, sentence structures, and stylistic choices. A third-party rewrite can scramble the pattern, but it also scrambles the text’s character. The trade-off between removal and fidelity is real.
Third, the Apache 2.0 license is a legal shield, not just a nod to open source. Hoskinson didn’t pick MIT. He picked Apache 2.0, which includes an explicit patent grant. That means anyone can fork, modify, sell, or bundle Anthropies, and Anthropic cannot use its patents to block those forks. This is a classic open-source defense against a single-company lawsuit. Hoskinson is signaling: you can’t kill this project by suing the original author; the code is now a public good.
The Legal Gambit: “Subject to” Is the New “Smart Contract”
Here’s where Anthropies stops being a tool and starts being a legal grenade.
Hoskinson’s accompanying X post didn’t just say “download this tool.” He argued that Anthropic’s own terms of service create a catch-22. The terms state: “Output is owned by you, subject to your compliance with our Terms.” In contract law, the phrase “subject to” can be interpreted as a condition precedent—meaning the ownership transfer never happens if the condition is violated. If you remove the watermark, you’ve violated the terms (assuming the terms prohibit removal, which they arguably do by implication). Therefore, you never owned the output in the first place.
This is a devastatingly clever argument. It turns Anthropic’s “you own your output” promise into a legal illusion. And it doesn’t stop at Anthropic. OpenAI, Google, Meta—all have similar clauses. If Hoskinson’s interpretation gains traction, it could trigger a wave of scrutiny over AI company terms of service.
But is it legally sound? Not yet. There’s a distinction between a “condition precedent” and a “promissory condition.” Courts haven’t ruled on this specific language. Hoskinson is picking the interpretation that favors his argument. Still, the mere existence of the argument forces Anthropic to either clarify its terms or face a reputational hit. In a world where the company is preparing for a $2 trillion IPO, silence is costly.
Contrarian: The Tool Is a Warning, Not a Weapon
Hoskinson himself called Anthropies a “warning, not a practical tool.” That’s a rare moment of self-awareness from a founder who usually doubles down. And he’s right—for now.
Anthropies has 4 stars on GitHub. No independent audit. No user reports. The prose layer relies on calling an external LLM API, which introduces latency, cost, and the risk that the external model also applies its own watermark. The tool’s entire premise assumes there exists an unwatermarked LLM accessible via API. That assumption may not hold for long as AI companies race to comply with the EU AI Act.
*The real contrarian angle is this: Anthropies might actually help Anthropic.* By proving that watermark removal is possible via a third-party rewrite, Hoskinson is giving Anthropic a blueprint for hardening their watermark—making it robust against exactly this kind of attack. The cat-and-mouse game accelerates. The tool becomes a stress test, not a solution.
Moreover, the tool’s existence legitimizes the very concept of watermarks. Hoskinson implicitly accepts that Anthropic’s watermark is detectable and removable. But if the watermark is so easy to remove, why should regulators care? The EU may respond by requiring even stronger, unremovable watermarks—like embedding signals in the semantic embedding space rather than token selection. The tool could backfire by strengthening the very system it opposes.
Takeaway: The Architect Who Wakes Up When the Market Sleeps
When the market sleeps, the architects wake up. Hoskinson isn’t building a product; he’s building a precedent. Anthropies is a political statement wrapped in code and licensed under Apache 2.0. It says: AI output ownership is not a given; it’s a legal battle waiting to happen.
For the crypto community, the lesson is deeper. The same philosophy that drives decentralization—trustless verification, user sovereignty, resistance to censorship—applies to AI. We’ve spent years arguing about who owns the blockchain. Now we need to argue about who owns the AI-generated text. Hoskinson is forcing that conversation.
Will Anthropies become a widely used tool? Probably not. Its technical limitations and early stage suggest it will remain a niche artifact for developers and activists. But as a narrative anchor, it’s powerful. It reframes the watermark debate from a technical curiosity to a civil rights issue.
The question that keeps me up at night: If Anthropic’s watermark is a leash, and we cut it, what happens to the dog? The dog is the user—now free, but also unidentifiable. In a world of deepfakes, misinformation, and AI-generated propaganda, is anonymity always a good thing? Hoskinson’s tool gives us the power to choose. But with power comes responsibility. The architects who wake up when the market sleeps must also think about what happens after the alarm rings.
Art is the interface; blockchain is the canvas. But the watermark is the signature. And sometimes, removing the signature means the painting becomes a forgery, even if it’s yours.