Anthropic is implementing invisible watermarks in Claude AI's output, adapted from Google DeepMind technology, to comply with the EU AI Act. The company acknowledges limitations, including potential removal through editing and the possibility of misidentifying human text edited by Claude.
This development addresses the growing need for transparency in AI-generated content, driven by regulations like the EU AI Act, while highlighting the technical challenges and user concerns surrounding watermarking effectiveness and potential misidentification.
Anthropic has begun embedding invisible watermarks into text generated by its AI chatbot, Claude, to comply with the EU AI Act's Transparency Code. The company detailed its approach, adapted from Google DeepMind's SynthID Text technology, explaining that the watermarks are undetectable to readers but verifiable with a key. Anthropic acknowledged that while light editing might not remove the watermark, a complete rewrite would, and that heavily edited human text could be misidentified. The company also noted that watermarking will have a negligible effect on generated code, with potential marks appearing in comments. Other major AI developers are expected to implement similar watermarking measures, though some users have expressed concerns and canceled subscriptions over the move.