GitHub stars: 4. Narrative impact: undefined.
You just paid for an AI output. You think you own it. But what if you don't? What if, buried in a 10,000-word terms-of-service document, there's a clause that says, 'Subject to your compliance with our Terms' — and breaking that clause means the ownership never actually transferred?
That's the bomb Charles Hoskinson just dropped. Not with a new Cardano upgrade. Not with a DeFi protocol. With a free, open-source tool called Anthropies — a name that sounds like a joke but carries a legal shrapnel that could hit Anthropic's $2 trillion IPO dreams.
Chasing the green candle that never sleeps, but this time the candle is a legal argument.
Context: Why Now, Why This
Anthropic's Claude deployment started watermarking outputs in August 2026, responding to the EU AI Act's transparency mandates. The watermark is invisible to the human eye — a statistical pattern embedded through a tournament sampling process that biases word choices toward a hidden key. It's elegant, robust, and designed to survive simple rewrites.
But Hoskinson saw an opening. Not in the watermark itself, but in the legal scaffolding around it. He released a tool that strips the watermark — but more importantly, he published a legal argument that questions whether Claude users even own their outputs in the first place.
Bear market tip: When liquidity dries up, attention shifts to existential debates. Ownership of AI output is the new frontier. And this is a frontier skirmish.
Core: The Three Layers, the Legal Trap, and the Code Bias
Let's get into the guts. I've been in this space since 2017, from the ICO madness to the DeFi summer to the NFT circus. I've read more whitepapers and terms of service than I care to admit. And I can tell you: Hoskinson's tool is a brilliant engineering hack, but the real alpha is in the contract language.
The Three Technical Layers
Hoskinson's codebase breaks the watermark removal into three layers:
- Layer 1: Co-Authored-By — Strips the git trailer that Claude sometimes appends. This is trivial. A simple regex gets it done.
- Layer 2: C2PA Image — Removes embedded metadata from images. Again, straightforward. Re-encode the image, and the C2PA credential disappears.
- Layer 3: Prose — The hard one. This is where the watermark lives in the statistical distribution of words. The tool's approach: route the text through a non-Claude LLM (like GPT or Gemini) to rewrite it, breaking the original statistical pattern.
Here's the kicker: The tool actively refuses to run on Claude or Gemini directly. It calls this "orchestrate mode" — it detects the host model and bails out. Why? Because rewriting inside the same model that applied the watermark would just re-apply it. This is honest engineering. It admits the tool's limitation.
But I've seen this pattern before. During the DeFi summer, every yield aggregator promised "zero risk" — but the fine print always had a caveat. The orchestrate mode is the caveat. It means the tool can't work in a closed loop. It needs an external, non-watermarked model to do the heavy lifting.
The Code Elephant in the Room
Code is the easiest layer to strip. The watermark has few places to hide in the rigid syntax of programming languages. But code is also the least watermarked content. Who cares if a for loop is flagged as AI-generated? The real battlefield is natural language — blog posts, political commentary, academic papers. And for prose, the tool's effectiveness is unproven. No independent verification. No benchmarks. Just a claim.
I've audited enough smart contracts to know that a demo on the easiest case doesn't prove the tool works on the hard case. Hoskinson is a savvy marketer. He chose code first because it's a guaranteed win. The prose case? That's a question mark.
The Legal Mousetrap: You Don't Own Your Words
This is the heart of the story. Hoskinson's legal argument is deceptively simple:
Anthropic's terms say: "Subject to your compliance with our Terms, we assign to you all right, title, and interest in and to the Output."
Hoskinson reads that as a condition precedent — a legal mechanism where the ownership transfer only happens if you comply with all terms. If you use the watermark removal tool, you've violated the terms. Therefore, the ownership never transferred. You never owned the output. You were just leasing it.
This is a legal landmine. If this interpretation holds even a fraction of water, it means every piece of content generated by Claude users who later strip the watermark is technically stolen property. The implications for copyright, defamation, and regulatory compliance are staggering.
I've seen this game before. During the 2017 ICO boom, every token sale had a "subject to" clause that the SEC eventually used to classify tokens as securities. The fine print matters. And Hoskinson just found the fine print that could upend the entire AI-as-a-service model.
The Apache 2.0 Shield
Hoskinson didn't just release the code. He licensed it under Apache 2.0, which includes an explicit patent grant. This means Anthropic can't sue him for patent infringement on any watermarking technology they might have patented. The tool is now a public good that can be forked, modified, and redistributed. Even if Hoskinson gets hit with a DMCA takedown, the code lives on in a thousand forks.
Speed is the only currency that matters here, and Hoskinson moved fast. He published the code, the legal analysis, and the licensing in one shot. It's a three-pronged attack: technical, legal, and social.
Contrarian: The Tool Is a Distraction, the Persona Is the Play
Everyone is focusing on whether the tool works. Is it effective? Can it strip prose watermarks? That's the wrong question.
In the jungle of alerts, silence is gold. And the silence from Anthropic is deafening. They're preparing for a $2 trillion IPO. They don't have time to debate a 4-star GitHub repo. But their silence is a gift to Hoskinson. It allows him to frame the narrative: "They didn't respond because they can't."
Here's the contrarian take: This tool is not about watermark removal. It's about Hoskinson rebranding himself from "Cardano founder" to "AI governance critic." He's been in a multi-front war all year, arguing about Ethereum copying Cardano's ledger design. This move expands his battlefield. It's a power play for attention and relevance.
And the tool itself? It's a symbolic protest. The orchestrate mode means it's not a standalone solution. The reliance on non-Claude models means it's essentially replacing one AI with another. The user still doesn't own the output — they just swapped one watermark for another. The tool doesn't solve the ownership problem; it just shifts it.
In a bear market, survival matters more than gains. And this tool might actually increase risk for users. If you use it, you're violating Anthropic's terms. If Anthropic ever decides to enforce, you could lose your account, your content, or worse.
Collecting moments, not just tokens, in the chaos — but this moment might come with a legal bill.
Takeaway: The Sprint Ends, but the Ledger Remains Open
The real impact of this event won't be measured in GitHub stars or ADA price. It will be measured in how AI companies rewrite their terms of service in the next six months. Hoskinson has exposed a structural weakness in the "you own your output" promise. Lawyers will be busy.
For traders: Don't buy ADA on this news. The event is a narrative blip, not a fundamental shift. But watch Anthropic's response. If they change their terms, it's a confirmation that the argument has teeth. If they stay silent, the narrative will fade.
For builders: This is a blueprint. The combination of open-source code, legal argument, and Apache 2.0 licensing is a new playbook for challenging centralized AI power. Expect more of these "assault tools" to appear.
For the rest of us: The question remains — who really owns the words the machine writes for you? The answer might be: no one. And that's a problem that no watermarking tool can solve.
The sprint ends, but the ledger remains open. I'll be watching the fine print.