Anthropic adds invisible watermarks to Claude AI output, sparking new concerns
Anthropic has begun embedding a hidden watermark in text generated by its latest Claude models, a move intended to label AI-created content but that raises practical and privacy issues.
Anthropic announced that its newest Claude models automatically tag generated text with an invisible, machine-readable watermark. This mark survives when users paste the output into various formats such as documents, webpages or code repositories. The initiative seeks to address the growing flood of AI-produced text, images and audio by providing a way to trace origin.
However, the company acknowledges that extensive re-phrasing can erase the watermark, limiting its reliability. Moreover, the rollout applies globally, covering all Claude services—including Claude Code—rather than being confined to markets that mandated such labeling. Critics argue the solution adds complexity without guaranteeing proof of authorship.
Why it matters
The watermark could affect how AI-generated content is identified, shared and regulated across the internet.
In this story
