Anthropic adds hidden watermark to Claude output to flag AI-written text
Anthropic announced that new Claude models will embed an invisible watermark in generated text, aiming to improve transparency under the EU AI Act.
Anthropic revealed that its latest Claude models will automatically embed an imperceptible watermark into any AI-generated prose, a step intended to meet transparency obligations of the European Union AI Act. The watermark does not alter meaning or readability and can persist through copying, pasting, and some editing, though heavy rewriting or translation may erase it. Models launched on or after August 2 will include the feature, with retrofitting of earlier versions underway, and the company will also supply external tools for detecting the marks.
This development arrives as the publishing sector faces controversy over alleged AI-authored books, such as the crime novel "Call Me, I'll Hide the Body" and the horror title "Shy Girl." Anthropic joins Google DeepMind, which earlier this year began watermarking Gemini-generated text and video using its SynthID technology.
Why it matters
A built-in watermark gives educators and publishers a way to verify whether text was produced by AI, addressing growing concerns about undisclosed machine-written content.
In this story