Claude’s new Scarlet Letter watermark is invisible—for now (arstechnica.com)

Anthropic has announced the implementation of invisible, machine-readable watermarks for all content processed by its AI models, in compliance with the European Union’s AI Act. This groundbreaking move mandates that AI-generated or manipulated outputs—spanning text, audio, images, and video—include identifiable watermarks, effective from August 2 for new models and applying a grace period until December 2026 for older versions. Anthropic's commitment extends beyond the EU, with plans to watermark content globally from the outset for all new models. Text outputs will feature embedded watermarks, while other generated files will carry digitally signed provenance metadata where applicable. This initiative is significant for the AI/ML community as it addresses growing concerns about content authenticity and manipulation. By watermarking processed content—even when the EU’s guidelines do not necessitate it for minor edits—Anthropic adopts a comprehensive "nuke it from orbit" strategy. However, the technical ramifications remain to be fully understood, particularly in distinguishing between substantial and trivial alterations, which raises questions about the watermark's effectiveness. Anthropic plans to provide detection tools to facilitate compliance with the EU law, although certain platforms may not support these watermarks. The shift emphasizes a proactive approach to content verification amidst increasing regulatory scrutiny in the AI landscape.