Anthropic Embeds AI Watermarks in Claude Text Globally

Anthropic announced it will begin embedding machine-readable watermarks into text generated by new Claude AI models globally, a move designed to enhance transparency and fulfill commitments made under the European Union’s AI Act transparency code. These new requirements are set to apply to Claude models launched in the EU on or after August 2. Anthropic stated that it will enable this watermarking system across all regions where Claude is available, extending beyond just Europe. The watermarking feature will encompass output from supported models across various Anthropic products, including Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag. Furthermore, Anthropic is actively working to integrate watermarking capabilities into models released prior to August 2, during the AI Act’s designated transition period. The watermarking process involves embedding an imperceptible, machine-readable mark directly into the text response generated by a supported Claude model. Anthropic asserts that this watermark does not alter the meaning, quality, or readability of the text. Crucially, because the watermark is an intrinsic part of the text itself, it remains with the content even when copied and pasted, and may persist through some forms of editing. The watermarking occurs at the model level, ensuring its inclusion regardless of which specific Claude product generates the output. In addition to text, supported PNG, JPG, and SVG image files will also receive signed C2PA provenance metadata. Anthropic plans to provide tools for both users and third parties to detect these embedded text watermarks and signed provenance metadata, with technical documentation to be released at a later date. The company also issued a cautionary note, explaining that the detection of a watermark signifies that the content may have been processed by Claude, but it does not definitively prove that Claude originated the core ideas or the initial text. Conversely, the absence of a detectable watermark does not preclude content from being AI-generated, as older models, heavily edited text, or files with stripped metadata might not contain detectable marks. This initiative is particularly relevant as Claude is widely utilized for creating, editing, and refining marketing content, with Anthropic emphasizing that the watermarks are intended to bolster transparency in AI-generated content.