Overview
- Anthropic updated its policy to embed imperceptible, machine‑readable watermarks in text and attach C2PA‑signed provenance to supported files for Claude models released on or after August 2, 2026.
- Text marks are implemented at the model level by biasing token choices so the watermark travels with copied text, while images and other files receive digitally signed metadata using the C2PA standard.
- The company says the marking will apply worldwide across Claude products and when accessed through cloud partners, and it plans to retrofit older models during the EU Act’s transition period.
- Anthropic warns the marks are a signal not conclusive proof of authorship because heavy editing, intense paraphrasing, short excerpts, or file conversion can remove or weaken them, and it has not yet published technical specs or public detection tools.
- The announcement prompted immediate user criticism over the inability to opt out and concern that human writing edited by Claude could be labeled, raising practical and policy questions for publishers, developers, and platforms.