Anthropic details invisible watermark system for Claude AI text

Anthropic said its Claude language model will incorporate an invisible watermark into each piece of AI-generated text to comply with European Union transparency rules. The watermark relies on subtle word-choice patterns that are indistinguishable to readers but can be identified by parties possessing a secret key. Instead of selecting words with a standard random generator, the model will use the key together with preceding words to guide each token choice, creating a detectable signature without altering the content's meaning. The approach leverages low-impact lexical decisions—such as choosing between synonyms like “overcast” or “grey”—which occur frequently in generated text, allowing the watermark to be embedded through many small, random-like selections.