Four GitHub stars. That's the current traction for a tool designed to dismantle Anthropic's AI watermark. Charles Hoskinson's Anthropies isn't about code—it's about a legal loophole that could redefine who owns AI output. But the market's quiet. The hype is loud. The data? Almost silent.
Context: The Watermark That Binds
Anthropic's invisible watermark isn't a post-hoc stamp. It's a statistical fingerprint injected during generation—tournament sampling that biases token selection. The EU AI Act, effective August 2, 2026, mandates machine-detectable AI content. Anthropic complied. Hoskinson responded with a free, open-source tool that claims to strip that watermark. The tool is a day old. The code is straightforward. The implications are not.
Core: The Three-Layer Deception
Anthropies breaks the problem into three layers. Layer 1 removes git trailers—deterministic, trivial. Layer 2 strips C2PA image metadata—re-encoding, effective. Layer 3 attacks the prose: the watermark's core. This is where the tool's true architecture reveals itself.
Hoskinson's key insight: rewriting watermarked text within the same model family re-applies the watermark. So Anthropies uses a non-origin rewrite—it routes text through a third-party LLM (non-Claude, non-Anthropic) to break the statistical pattern. The tool actively detects the host model and refuses to rewrite if it detects a watermarked output source. This is the "orchestrate" mode—a self-imposed limitation that admits defeat in the native ecosystem.
But the choice of demo matters. The tool's GitHub repository showcases code. Code carries almost no watermark signal—syntax is rigid, substitution space is tiny. The prose layer remains unvalidated. Based on my experience auditing smart contracts, I've seen this pattern before: a vulnerability in the demo, not the real battlefield. The code demo is a controlled environment. The real world is natural language.
The legal argument is the real payload. Hoskinson focuses on one clause in Anthropic's Terms of Service: "Anthropic hereby assigns to you all its rights, title, and interest in and to the Output, subject to your compliance with our Terms." He reads "subject to your compliance" as a condition precedent—meaning ownership never transfers if the user violates any term. Watermark removal? That's a violation. So the user never owned the output. The legal implications are massive: every Claude user who stripped a watermark may have been working with unowned content.
Chain doesn't lie. The code is open. The legal argument is public. The market hasn't priced this yet.
Contrarian: The Tool Is a Trojan Horse, Not a Silver Bullet
The hype suggests Anthropies is a magic eraser. It's not. The prose layer is the hardest, and the tool's effectiveness there is unknown. The demo uses code—a best-case scenario. The "orchestrate" mode's self-limitation means the tool can't work inside the ecosystem it targets. That's a technical honesty, but also a confession.
The real value is the legal argument. Hoskinson is not trying to build a sustainable tool. He's planting a flag. The Apache 2.0 license ensures the code can't be easily killed by a DMCA takedown. But the legal argument is fragile: courts have not ruled on whether "subject to your compliance" is a condition precedent or a promissory condition. If it's the latter, ownership transfers immediately, and violation is a separate breach. The argument is an interpretation, not a precedent.
Leverage kills. Hoskinson is leveraging his reputation as a Cardano founder to amplify a legal attack on an AI giant. The tool is the lever. The legal argument is the fulcrum. The market is the object being moved.
Takeaway: The Next Signal
Watch Anthropic's response. Silence is a signal—it may imply the legal argument has merit. A swift term update would confirm the vulnerability. Either way, the narrative has shifted from "can the tool remove watermarks?" to "who owns the output?" That's a question no algorithm can answer. The next week's data: GitHub fork count, Anthropic's blog, and any legal academic citations. If the legal argument gets cited, the tool's purpose is fulfilled. If not, it's a four-star footnote.
Follow the exit liquidity. The liquidity here is legal credibility, not capital. Whales are circling—but they're lawyers, not traders.
