The GitHub repo has four stars. Four. That's not a typo. Charles Hoskinson, the man who built a $15 billion blockchain, launched a tool to strip Anthropic's AI watermark and the market responded with a collective shrug. The chart didn't lie. The hype curve was flat. No viral spike, no FOMO cascade. Just a lonely repo with a single commit and a legal argument that sounds clever until you read the fine print.
I bought the pixel, not the promise. I've seen this pattern before: a high-profile founder drops a code bomb, the community cheers, and then the repo sits untouched for six months. The 2020 yield farming experiment taught me that code without traction is just noise. So when I saw the Anthropies repo, I spun up a local node. I verified the claims. What I found is a tool that solves a problem that largely doesn't exist, using a method that may not work for the problem that does.
Context: The Watermark Wars
Anthropic, the AI company behind Claude, recently deployed a cryptographic watermark on its outputs. This is not your grandma's metadata tag. It's a tournament sampling scheme that injects a detectable statistical pattern into the text's probability distribution. The goal is to comply with the EU AI Act's transparency requirements, which went into effect on August 2, 2026. Hoskinson responded with Anthropies, a free, open-source tool that claims to remove this watermark. He also accompanied it with a legal analysis arguing that Anthropic's terms of service mean users never actually own the outputs.
The tool is divided into three layers: removing co-authored-by git trailers, stripping C2PA image metadata, and rewriting prose through a non-origin LLM to break the watermark signal. The third layer is the core claim. It's also the most questionable.
Core: The Technical Reality Check
Let's talk about the prose layer. Hoskinson admits in his own codebase that prose is the "difficult layer." The attack vector is simple: route the watermarked text through a non-Claude model (e.g., Gemini) and ask it to rewrite the content. The theory is that the new model's output will have a different statistical distribution, washing out the watermark. But here's the rub: the non-origin model might itself apply a watermark. Or the rewrite might preserve enough of the original distribution to still trigger detection. There's no guarantee. The tool's effectiveness on natural language is unproven.
Meanwhile, the code layer is trivial. Code is a structured language with limited syntactic freedom. Watermarks have few places to hide. Hoskinson chose to showcase the tool on code because it's the easiest win. That's a classic bait-and-switch. The real battlefield is prose, and the tool's performance there is a black box.
Risk isn't a feeling. It's a measurable probability. And the probability that this tool meaningfully removes watermarks from a novel or a news article is somewhere between dubious and flat earth. The reliance on an external LLM API introduces latency, cost, and a dependency on the very infrastructure the user is trying to escape. If the external model changes its policy, the tool breaks.
Every candle tells a story of fear. The fear here is that the market will overestimate the tool's capabilities. Already, I see tweets calling this a "massive win for user freedom." But the on-chain data doesn't lie. The GitHub stars are a proxy for adoption. Four stars. That's not a movement. That's a footnote.
Contrarian: The Retail Blind Spot
Retail loves the David vs. Goliath narrative. Hoskinson, the maverick founder, taking on the $2 trillion IPO-bound AI giant. It's a compelling story. But smart money sees the cracks. The tool's legal argument is the real weapon, not the code. Hoskinson's interpretation of Anthropic's terms of service as a "condition precedent" is a novel legal theory. It has no court precedent. It's a PowerPoint argument, not a legal reality. If applied universally, every SaaS company's terms would collapse. The courts will not take this bait.
Moreover, the tool is a single point of failure. Hoskinson is the only maintainer. If he gets bored or gets sued, the repo dies. The code is Apache 2.0, so forks can survive, but who will invest in maintaining a fork of a tool that may not work on the most important use case? The answer is nobody.
Takeaway: The Signal, Not the Tool
I don't make predictions. I assess probabilities. The probability that Anthropies becomes a widely used watermark removal tool is low. The probability that it sparks a necessary debate about AI output ownership is higher. Hoskinson has done the industry a service by forcing the conversation. But the tool itself is a distraction. The real solution is not a rewrite prompt. It's a fundamental change in how AI companies structure their terms and how regulators enforce transparency.
Liquidity vanishes when the music stops. The music on this tool will stop soon. The question is whether the legal argument will echo in future courtrooms. I'll be watching the audits, not the stars.