Anthropic has rolled out a new statistical text watermarking system for Claude models. Enabled globally across models released after August 2, 2026, it aims to meet EU AI Act transparency requirements. This watermark is purely statistical, adding nothing to the text itself. Claude applies a mathematical bias during token generation, nudging word selection probabilities towards a "green" list.
Detection requires Anthropic's private key and can only estimate Claude's involvement, not prove human authorship. Traditional cleaning methods, like Unicode sanitization or simple synonym replacement, fail because there are no physical characters to strip. The only reliable removal method is a complete, meaning-preserving reconstruction of the text by an independent system that re-samples every word.