Anthropic adds AI text watermarking to Claude models worldwide

Anthropic mentioned it would start embedding machine-readable watermarks into textual content generated by new Claude AI fashions worldwide. This fulfills commitments the corporate made beneath the European Union’s AI Act transparency code.

EU guidelines, world rollout. The brand new necessities apply to Claude fashions launched within the EU on or after Aug. 2. Anthropic mentioned it would allow the watermarking system all over the place Claude is out there, not simply in Europe.

• Watermarking will cowl output from supported fashions throughout Claude, the Claude API, Claude Code, Claude Cowork and Claude Tag.
• Anthropic can also be working so as to add watermarking to fashions launched earlier than Aug. 2 throughout the AI Act’s transition interval.

Invisible marks in textual content. When a supported Claude mannequin generates textual content, it embeds an imperceptible, machine-readable watermark into the response. Anthropic mentioned the watermark doesn’t change the which means, high quality or readability of the textual content.

• As a result of the watermark is embedded within the textual content itself, it stays with the content material when it’s copied and pasted. It might additionally survive some modifying.
• Watermarking occurs on the mannequin stage, so it’s included no matter which Claude product generated the output.
• Supported PNG, JPG and SVG information may even obtain signed C2PA provenance metadata.

Sure, however. Anthropic mentioned it would present instruments for customers and third events to detect embedded textual content watermarks and signed provenance metadata. Technical documentation can be launched later. The corporate additionally cautioned {that a} detected watermark solely exhibits that the content material could have been processed by Claude.

• It doesn’t show Claude created the unique concepts or wrote the unique textual content.
• Likewise, the absence of a detectable watermark doesn’t imply content material wasn’t AI-generated.
• Older fashions, closely edited textual content, and information with stripped metadata could not comprise detectable marks.

Why we care. Claude is broadly used to create, edit, and refine advertising content material. Anthropic mentioned the watermarks are meant to enhance transparency. If AI textual content watermarking turns into extra widespread, it might change how AI-generated content material is recognized and disclosed.

Claude’s assist web page. How Claude marks AI-generated content