Anthropic to Watermark All Claude-Processed Content, Including Edits

Anthropic has revealed plans to incorporate invisible watermarks into content processed by its Claude AI models, extending beyond just generated material. This initiative aims to align with the European Union's AI Act, which requires AI system providers to watermark outputs that are either AI-generated or altered. The regulation affects models launched after August 2, with a transition period until December 2026 for those already in use. From the outset, Anthropic will implement watermarks on all new models worldwide, encompassing text outputs and other files with digitally signed provenance metadata. Interestingly, Anthropic's comprehensive approach means watermarking all processed content where feasible, even in situations where the EU does not mandate it, such as minor grammar corrections. The full scope of watermarking will be clarified once Anthropic introduces a detection tool to fulfill the EU's technical support requirements.