Anthropic announced that all Claude models will soon incorporate watermarking in their generated text to comply with an EU regulation.

Instead of using invisible Unicode characters, Anthropic will apply steganography by embedding subtle word-choice patterns as fingerprints. Their support document claims these watermarks are imperceptible and do not alter meaning, quality, or readability. This approach suggests that the text will maintain its clarity while embedding these hidden markers, although the exact method remains speculative.

Source: daringfireball.net