How Claude's text watermarking works

Anthropic has published a technical explanation of how Claude's text watermarking feature works. The watermark is an invisible pattern embedded in the generated text that allows Anthropic to detect whether a piece of text was produced by Claude. It works by making subtle, pseudorandom choices between plausible word alternatives during generation, guided by a cryptographic key. This creates a statistical signature that can be recognized later, while remaining unnoticeable to readers. The system is designed to help mitigate misuse by enabling traceability of AI-generated content. Anthropic emphasizes that the watermark is robust and does not compromise the quality or fluency of the text.