Anthropic adds Claude watermark that travels with copied text
Anthropic embeds invisible watermarks in Claude's output to comply with EU AI Act, making AI-generated text detectable globally.Summary
Anthropic will embed invisible statistical watermarks in all text generated by its Claude model to comply with Article 50(2) of the EU AI Act, which requires AI outputs to be detectable as machine-generated. The watermark is applied globally, survives copying and some editing, but may be removed by heavy editing. Other major model providers are expected to follow with similar measures.
For legal AI platforms built on frontier models, this raises practical questions about whether AI-assisted documents such as contracts, memos, and court filings will carry detectable markers after editing. Experts note compliance obligations may extend through the chain to platforms integrating models, requiring close cooperation between model providers and integrators. Watermarking could make AI use visible to clients and opposing counsel, potentially affecting billing transparency and requiring firms to document AI use policies. Consumption-based pricing trends may also be justified by disclosed AI usage.
(Source：Complete Ai Training)