How Claude marks AI-generated content

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, covering both generative AI models and systems. The company is working to include machine-readable marks in content Claude generates, and will update this article with more detailed technical guidance as it becomes available.

Claude uses two complementary marking techniques. First, embedded watermarks are woven imperceptibly into generated text at the model level, so they are present across all Claude products and may persist through copy-paste and some editing without changing meaning, quality, or readability. Second, signed provenance metadata is attached to supported file types such as .svg, .png, and .jpg, following the C2PA open standard, which records provenance information and enables tamper detection.

Anthropic is also working on detection mechanisms so users and third parties can check whether text or files carry a supported Claude mark. However, machine-readable marks have limitations across all content types. For developers deploying Claude in their own products, Anthropic notes they must independently assess what Article 50 requires and says it will share technical guidance on marking and detection approaches to support compliance.