How Claude's Text Watermarking Works

This article explains how Claude's text watermarking works, based on the author's reading of Anthropic's released materials. It describes the technique as using secret keys and previous tokens to influence random sampling during generation, creating statistically detectable patterns. The author discusses how watermarking can be removed via editing or rephrasing with another LLM, though this may degrade text quality. The article also addresses the EU AI Act requirement for global watermarking, with a counter-argument about liability in cross-border scenarios. The content is technical and relevant to AI/IT, focusing on model behavior and regulation.