Origin: Anthropic Implements Watermarking in Claude AI

Anthropic began embedding invisible watermarks in all text Claude writes and adding signed metadata to images it creates, to meet a European rule that AI output be detectable. The marks apply across Claude apps and the API, so they travel into third-party products, and the rollout is global. New models launched on or after August 2, 2026 include it at launch, and older models will be updated over the coming months.

The text mark is a statistical signature woven into the words Claude picks, guided by a secret key. It adds no characters, survives copy and paste, and carries no information about the user or organization. There is no setting to switch it off.

The mark only shows Claude was likely involved and cannot separate authorship from assistance. A translation created by Claude is fully marked, but heavy rewrites, paraphrases, or translations done elsewhere can erase the signal. A text-detection API is planned; for images, a signed note in the file’s metadata lets compatible tools confirm Claude’s involvement and spot tampering.

AI-synthesized from multiple reported sources