Anthropic plans to add invisible, machine-readable watermarks to content processed by new Claude models worldwide, partly to meet requirements under the EU AI Act. The policy matters because it could label not only material generated by AI but also human-written text that Claude merely edits, potentially blurring the distinction between minor assistance and wholesale AI creation.

The EU rules apply to models released after 2 August, with older models given until December 2026 to comply. Anthropic says text marks may survive copying and some editing, while images and videos will use C2PA provenance metadata; however, both approaches can be removed or disrupted relatively easily. The company acknowledges a detected mark only suggests content may have been processed by Claude and is not conclusive evidence of AI generation.