Claude Will Watermark All AI Text by Default—Here's What That Means for You

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Starting August 2, 2026, all new Claude models launched in the EU will embed machine-readable watermarks in generated text and attach signed provenance metadata to supported file types. This applies globally, across every surface where Claude runs—API, Claude Code, Claude Cowork, and even through cloud partners like AWS and Google Cloud.

Watermarks are imperceptible and woven directly into the text, surviving copy-paste and light editing. For files like .svg, .png, and .jpg, Claude will attach C2PA-compliant signed metadata that can detect tampering. Anthropic also plans to provide detection tools for third parties to verify whether content carries a Claude mark.

But here's the catch: a detected mark doesn't prove Claude authored the content—it only signals that Claude processed it. And the absence of a mark doesn't mean the content wasn't AI-generated. Heavy editing, short passages, or metadata stripping can all erase the signal.

If you build on Claude, Anthropic's guidance is to assess your own obligations under Article 50. They promise technical documentation soon, but for now, the specifics of detection remain vague.