Anthropic's AI Watermark Sparks Tools to Remove It

Anthropic recently announced that some Claude models will add invisible watermarks to text to improve transparency. The watermark relies on statistical patterns in word choice and survives copy-paste. This move aims to meet EU AI Act commitments but has raised concerns, as even light edits like proofreading may leave marks. Some users canceled subscriptions, and US searches for "AI watermark remover" rose 60% month-over-month.

Developers reacted quickly. Paris founder Guillaume Meyer released open-source tool Watermarks Remover, which strips hidden characters and rewrites text to break the pattern, gaining over 14k stars on GitHub. Tokyo dev Ansh Aneja launched MarkScrub, growing to 8,500 users in a day. A researcher at ETH Zurich noted there's always a way to remove watermarks, like full rewrites.

These tools sit in a legal gray area. The EU AI Act requires persistent marks but doesn't ban third-party removal. If a tool uses AI to rewrite, it might need to add new marks. Anthropic's policy bans impersonating humans, but nothing stops removal tools. With a detection API coming, the battle between watermarks and counter-tools continues.