Anthropic Embeds Invisible Watermark in New Claude Model to Identify AI-Generated Text

On August 11, Anthropic announced that its new Claude AI model will embed a subtle, hard-to-detect watermark into generated text. The watermark does not affect the meaning or readability of the text and may persist even after the text is copied, pasted, or partially edited. Anthropic stated this move responds to transparency requirements under the European Union’s Artificial Intelligence Act, aiming to verify content provenance and prevent unauthorized copying or misuse. The watermarking feature will be enabled by default in new versions of the Claude model released on or after August 2. Anthropic is also adding this capability to older model versions. The technology applies globally to all content generated by Claude, including text produced via cloud service providers. Anthropic plans to offer third parties tools to detect these watermarks, helping publishers, schools, and universities identify AI-generated content. Anthropic also noted that the watermark may become undetectable if the Claude-generated text undergoes substantial modifications, rewriting, translation, or is mixed with other content. Furthermore, detecting a watermark does not definitively prove that Claude was the original author of the text, as using Claude solely for proofreading or translation could also leave watermark traces. Anthropic is the second major AI lab to introduce text watermarking technology, following Google DeepMind.