Where an AI Watermark Can Hide in Plain Text

Covered by 2 sources
Read full post
Anthropic announced it will watermark all Claude AI-generated content, using C2PA metadata for images and an imperceptible watermark embedded in text through word choice patterns rather than byte-level encoding. The text watermark persists even after copying and pasting, and its exact method remains undisclosed, but it likely resides in the model's selection of words rather than in the text's byte encoding.

Covered by 2 sources

More on this story


More in LLM & Text Generation

DeepSeek V4.1 Flash now available on AI Gateway

Covered by 2 sources

Harvey raises $550M more to develop AI tools for legal teams

SiliconANGLE