Anthropic shares more details about how Claude’s new watermarks will work

- Anthropic finally explains how Claude's watermarks work, and surprise — it's bureaucracy with extra steps. The idea is every chunk of AI-generated text gets a cryptographic stamp baked in at the token level, so you can tell machine output from some meat wallet pretending to write. Can you strip it with light editing? Mostly no, unless you rewrite enough that you basically authored it yourself. And code? Watermarks survive in comments and structure but break under heavy refactoring — so your copy-pasted smart contract might still snitch on you. The real question nobody's asking: who holds the verification keys, and where's my cut? Watermarking is fine in theory, but if it's gated behind one company's API, it's just a toll booth with better PR.