Anthropic to embed invisible watermark in AI-generated text

Anthropic announced that all Claude models released on or after August 2 will embed an imperceptible, machine-readable signal in the text they generate. The watermark is designed to travel with the content when it is copied, pasted, or lightly edited, though extensive rewriting or translation may remove it. The same technology will be applied to images produced by Claude, indicating whether the file was processed by the model and if it has been altered.

The initiative aligns with the European Union's AI Act, which requires generative-AI providers to make synthetic output detectable and machine-readable. While other firms have focused on watermarking images, Anthropic says text watermarking has been more challenging because of frequent copying, paraphrasing, and translation. The company notes the mark only confirms Claude’s involvement, not that the entire piece was generated by the model.

Anthropic’s move comes amid broader industry efforts to address low-quality AI-generated content on platforms. Substack has introduced a reader-triggered AI scanner, and YouTube has tightened its policy on “inauthentic content,” restricting monetization for channels that rely heavily on generic AI-produced material. Critics warn that a simple AI label could conflate malicious large-scale fake content with legitimate uses such as editing or translation.

Anthropic acknowledges that watermarking alone will not eliminate all misuse, as determined actors can still alter or erase the signal. Nonetheless, the company views the technology as a step toward giving consumers and publishers a clearer way to differentiate AI-generated text from human writing.