Anthropic to watermark all Claude AI output globally
All new Claude AI models will embed invisible watermarks by August 2026...
Anthropic will add machine-readable labels to content from new Claude models launched in the EU on or after August 2, 2026, as part of its EU AI Act transparency commitments. The labels—invisible watermarks in text and C2PA provenance metadata for supported image files—will apply worldwide, but only where the model and feature support them. Anthropic says it's developing detection tools, but notes that a detected mark only suggests Claude may have processed the content—not that it was entirely AI-generated—and marks can be missed after heavy editing, very short passages, or on older models.
- All new Claude models (post-August 2, 2026) will include machine-readable watermarks for EU compliance and global rollout
- Invisible text watermarks persist through copying/editing; C2PA metadata added for images to detect tampering
- Watermarks indicate Claude *processed* content—not full AI origin—and may be removed via heavy editing or file conversion
Why It Matters
Sets a new transparency standard for AI-generated content, helping users detect AI assistance in outputs.