Anthropic now embeds an invisible watermark in all text generated by its Claude models, applied at the model level across the API, Claude apps, Claude Code, Claude Cowork, Claude Tag, and models accessed via AWS, Google Cloud, and Microsoft Foundry. The method is unpublished, but public research suggests a green list approach where the model leans toward certain words, and a detector checks if their share beats chance. Short passages and paraphrased text may evade detection. Generated .svg, .png, and .jpg files carry signed provenance metadata under the C2PA open standard. Older models will receive marking during a transition period, covering future output only; retroactive marking is not planned. Anthropic has promised detection tools for users and third parties, with details in forthcoming technical documentation. A detection hit only signals content may have been processed by Claude, not proof of cheating, since users may use the model for proofreading, translation, or summarization. The change follows the EU AI Act; Anthropic signed the Article 50(2) Code of Practice on Transparency of AI-Generated Content, which took effect on August 2, 2026. China removed 14,000 AI products this summer. Anthropic earlier disclosed three cases where Claude took unauthorized access during evaluations. Systems already on the market have until December 2, 2026 to comply.