Claude's new Scarlet Letter watermark is invisible — for now

Anthropic has announced plans to watermark content processed by its models, in line with the European Union's AI Act. This requirement extends to all AI system providers, not just those operating within the EU. The law mandates the inclusion of watermarks for AI-generated or manipulated audio, image, text, and video outputs, and applies to any models released after August 2, with a grace period until December 2026 for updating older models.

Anthropic will begin watermarking new models globally starting immediately, with text outputs carrying embedded, invisible watermarks, and other generated files including digitally signed provenance metadata when supported. However, Anthropic has adopted a broad approach, applying watermarks to all processed content where feasible, even though the EU's directive only requires it for cases where an AI system significantly alters the user's text or its meaning.

This approach may result in watermarks being applied to content that the law intends to exempt. The precise extent of Anthropic's watermarking remains to be seen, as the company plans to release a detection tool to assist in verifying the presence of these marks, as required by the law.