How Claude's text watermarking works

Future versions of Claude will embed a digital watermark in their generated text. This watermark serves as a marker to assess whether Claude was involved in creating the text, supporting transparency and compliance with the EU AI Act. The watermarking technique involves subtle modifications to the model’s word choices, based on a secret key, which are undetectable to readers but can be verified with the key.

Large language models like Claude generate text by selecting words sequentially, with each choice influenced by the preceding context. Watermarking leverages this process by subtly biasing the model's choices in a way that leaves a detectable pattern, without affecting the overall meaning or quality of the output. This pattern is only identifiable with a specific key, enabling verification after the text is produced.

Importantly, watermarking does not alter the quality, creativity, or readability of Claude’s responses. Internal testing and human evaluations have shown no significant difference between watermarked and unwatermarked outputs. The technique is designed to be sparse, only influencing word choices in situations where multiple options are equally plausible, such as in general descriptive text.

The watermarking method is based on a concept introduced by Google DeepMind, which uses a sequence of digits (like pi) to determine word choices, making the process akin to a game with a hidden pattern. This approach allows for post-hoc verification of whether a text was likely generated by Claude, without impacting the user experience.

However, the watermarking has limitations. It can only estimate the probability that Claude contributed to the text, not confirm human authorship or identify other AI-generated content. Its effectiveness increases with longer texts and is less reliable in short passages or when the text is heavily edited or factual, where word choices are constrained. Additionally, the watermark is less applicable to exact outputs like code or specific factual statements, where word choices are limited.

Overall, Claude’s watermarking aims to enhance transparency and accountability in AI-generated content, aligning with regulatory requirements while maintaining output quality.