Overview
- Anthropic said on August 2 that Claude models released on or after that date will embed an imperceptible, machine‑readable watermark in generated text and that some generated files will carry digitally signed provenance metadata.
- The watermark is statistical in nature, favoring certain token choices so it does not visibly change meaning or readability and it is designed to travel with copied text, but Anthropic has provided few technical details so far.
- Anthropic plans to retrofit older Claude versions and to supply tools for third parties to detect the markings, and EU rules give vendors a transition period until December 2, 2026 to update preexisting models.
- The company and independent commentators warn the watermark can be erased or obscured by heavy edits, translations, mixing with other text or short passages, and finding a watermark does not prove Claude authored underlying ideas.
- Publishers, educators and platforms may use the mark to investigate suspected AI use, yet experts caution it could produce false positives and create misplaced trust, and other labs such as Google/DeepMind are pursuing similar approaches under the EU transparency code.