Overview
- Anthropic said new Claude models launched on or after August 2, 2026 will embed imperceptible, machine-readable watermarks in generated text and attach signed provenance metadata to supported files.
- Text marks are woven into token selection so they are invisible to readers, travel with copied text, and can survive some edits while remaining readable.
- Files such as .svg, .png and .jpg will carry signed provenance following the C2PA standard so a valid label can show processing by Claude and flag tampering.
- Anthropic will publish detection tools and technical guidance but warns marks are probabilistic and can be lost by heavy editing, short excerpts, format conversions, screenshots, or use of older unmarked models.
- The move targets EU Article 50 transparency obligations and could give publishers, schools and platforms a new way to check origins of content while raising questions about detection limits and how provenance is preserved in common workflows.