To comply with the EU AI Act's Article 50(2) Code of Practice, Anthropic is implementing machine-readable transparency measures for Claude-generated content. For all models launched on or after August 2, 2026, Anthropic is integrating two primary marking techniques: imperceptible embedded watermarks in text and C2PA-compliant Content Credentials for generated files like PNG or JPEG. Text watermarks are applied at the model level, ensuring they persist across platforms—including the Claude API, Claude Code, and cloud partner services like AWS, Google Cloud, and Microsoft Foundry—even when content is copied and pasted. For legacy models released before the August deadline, Anthropic aims to have watermarking support fully implemented by December 2, 2026. While text watermarks are being rolled out globally, file-based Content Credentials depend on the platform's ability to support file generation. To facilitate verification, Anthropic has launched a watermark detection API currently in private preview for regulated entities, such as law enforcement, researchers, and fact-checkers, with plans for broader enterprise access. However, the company notes technical limitations: marks indicate Claude's involvement but do not confirm original authorship, as text may be modified, heavily edited, or translated, potentially obscuring the signal.