Anthropic announced that all Claude models will implement text watermarking to comply with EU regulations. The company's initial announcement, titled "How Claude Marks AI-Generated Content," provided no actual explanation of the technical implementation. A later document, "How Claude's Text Watermark Works," revealed the use of semantic watermarking through steganography.

This technique involves sorting words into "green" and "red" lists at each token generation point, making the model slightly more likely to choose words from the green list. Detection requires a secret key held only by Anthropic, meaning other AI providers cannot detect Claude's watermarks and vice versa. The author criticizes this approach as adulterating the semantic quality of text generation, arguing that no two synonyms carry identical meaning and that word choice precision matters.

The watermarking works similarly to detecting a biased coin through statistical analysis - the more text generated, the more confident detection becomes. However, the author finds it unacceptable that AI tools should compromise text quality or meaning to embed hidden provenance markers, viewing it as a fundamental violation of user needs for clarity and coherence in generated content.