Four GitHub stars. That’s the tally for Anthropies, Charles Hoskinson’s open-source tool to strip Anthropic’s AI watermark. Four stars, twenty-four hours after launch. The market didn’t care. The hype machine didn’t ignite. And yet, buried in the code and the accompanying legal argument, there’s a signal louder than any star count. This isn’t a tool. It’s a legal grenade disguised as a Python script. And the fuse is a single condition in Anthropic’s terms of service.
Context
Let’s rewind. August 2026. The EU AI Act is in effect, demanding transparency labels on AI-generated content. Anthropic, preparing for a $2 trillion IPO, deploys a tournament sampling watermark on Claude outputs. It’s a clever statistical bias injected at generation time—no hidden strings, just a detectable pattern in the probability distribution. Hoskinson, ever the provocateur, releases Anthropies. The tool’s name is a pun: “Anthropic” + “antidote.” It promises to remove the watermark in three layers: git trailers, C2PA image metadata, and prose. The first two are trivial. The third is the battle.
Hoskinson isn’t just a developer. He’s the founder of Cardano, a blockchain with a $15 billion market cap. He’s been in a year-long feud with Ethereum’s community over technical credit. Now he’s aiming at AI. The move is classic Hoskinson: attack the dominant narrative, frame it as a fight for user rights, and wrap it in open-source license protection. But the data tells a different story. The tool’s GitHub repository has no forks, no issues, no pull requests. It’s a solo act. The code is minimal—a few hundred lines of Python. The real weight is in the README, where Hoskinson lays out his legal theory.
Core
Layer 1: Co-Authored-By. Git trailers are simple metadata. Stripping them is deterministic. No watermark there. Layer 2: C2PA. Image metadata embedded by AI tools. Re-encoding the image removes it. Again, straightforward. Layer 3: Prose. This is the hard part. Tournament sampling leaves a statistical fingerprint across the entire text. Replace a few synonyms? The watermark survives. Paraphrase? The pattern shifts but remains. Hoskinson’s solution: route the text through another LLM that doesn’t add a watermark. But here’s the catch—the router must detect the host model. If it’s Claude or Gemini, the tool refuses to rewrite. It calls this “orchestrate mode.” It’s a self-imposed limit, a technical honesty that reveals the tool’s weakness.
The code carries almost no watermark signal. Syntax has little substitution space. Hoskinson chose code as the demo because it’s the easiest case. Prose is the real test, and the tool’s effectiveness there is unproven. No independent audit. No benchmarks. Just a claim. The tool’s architecture assumes a non-watermarked LLM endpoint exists. That’s a fragile assumption. If all major providers adopt watermarks, the router becomes useless.
Every rug pull has a fingerprint; I just read it. The watermark is a fingerprint. Anthropies is a fingerprint eraser. But the eraser itself leaves marks. By routing through another model, you introduce new statistical patterns. The output is no longer the original. Who owns it? The legal question is more dangerous than the technical one.
The ledger remembers what the analysts forget. The tool’s code is open, but the analysis is missing. No one has tested it against a real Claude output. No one has measured the text fidelity loss. The tool is a concept, not a product. And concepts don’t kill watermarks.
They buried the truth in the gas fees of 2020. No, they buried it in the terms of service of 2024. Hoskinson’s real bombshell is legal. Anthropic’s ToS says: “All rights, title, and interest in the Output are transferred to you, subject to your compliance with our Terms.” Hoskinson interprets “subject to” as a condition precedent. If you violate the Terms—say, by stripping the watermark—the ownership never transferred. You were never the rightful owner. This is a radical reading. If upheld, it means every Claude user who ever used the output for commercial purposes without complying with the ToS (which includes not tampering with watermarks) is technically infringing. The tool itself is a violation. Hoskinson is daring Anthropic to sue.
He chose Apache 2.0 license for the tool. That grants patent rights, making it harder for Anthropic to claim infringement. It’s a legal shield. The tool is a proof of concept for a legal argument, not a piece of software.
Contrarian
The market is wrong. The four stars don’t matter. The tool’s technical limitations are irrelevant. The real value is in the legal precedent. If Hoskinson’s condition precedent argument gains traction, every AI company will need to rewrite their ToS. That’s a systemic impact. The tool is a test case. It’s a way to force the conversation.
But correlation is not causation. The tool’s low adoption rate doesn’t mean the argument is weak. It means the audience is distracted. The crypto community is obsessed with price action. The AI community is focused on training efficiency. Neither group is reading the fine print. Hoskinson is betting that the legal system will pay attention.
There’s a second contrarian angle: the tool may actually be a honeypot. By releasing it, Hoskinson exposes users who download it to potential legal liability. If Anthropic decides to enforce its ToS, anyone who ran Anthropies could be a target. The tool is a test of the legal waters. If nobody gets sued, the argument is weak. If someone gets sued, the case becomes a landmark.
Volatility is the noise; liquidity is the signal. The volatility in the AI watermark debate is the noise. The liquidity of legal arguments is the signal. Hoskinson is injecting liquidity into a dry legal space. The question is whether the courts will trade.
Takeaway
The next signal is not a GitHub star count. It’s a legal filing. Watch for Anthropic’s response. If they modify their ToS to explicitly state that ownership transfers regardless of compliance, Hoskinson wins the argument. If they ignore the tool, the argument festers. If they sue, the tool’s code becomes evidence. The data detective’s job is to follow the legal paper trail, not the code. The tool is a distraction. The terms of service are the real battlefield.
This is a bull market for legal theory. The markets are euphoric, but the technical flaws are hidden in plain sight. Hoskinson is using a code audit lens to see through the marketing. The tool is a magnifying glass, not a solution. Look at the contract, not the output. The ledger remembers, but the analysts forget. I won’t.