Future Claude models will include a text watermark to identify the likelihood of Claude's involvement in writing the text.

Watermarking involves using low-stakes word choices to embed a pattern in the text, detectable only with a specific key. This method doesn't affect the text's quality or meaning, as the random word selection still aligns with what Claude would naturally choose. The watermarking complies with the EU AI Act and allows for probabilistic verification of Claude's text generation without altering output quality or introducing unusual word choices.

Source: anthropic.com