Overview
- Anthropic announced that Claude models released on or after August 2 include an invisible, machine-readable watermark embedded directly in generated text to mark content that the service produced or processed.
- For supported file types such as PNG, JPG and SVG Anthropic will append digitally signed provenance metadata using open standards like C2PA and said it will publish tools for third parties to detect the marks.
- Anthropic warns the watermark does not prove authorship because Claude can leave marks when it only edits, translates or reformats a human text, and heavy edits, paraphrasing or mixing with other material can weaken or remove the signal.
- The company is working to retrofit older Claude models and to ensure the marking works across the API, consumer apps and cloud partners, and regulators may still enforce compliance under the EU AI Act and a voluntary transparency code.
- Publishers, educators and platforms may use the marks to investigate suspected AI use, but critics warn the system can create false positives and a false sense of certainty so human review and policy updates will remain necessary.