Anthropic to embed invisible watermarks in all Claude outputs to meet EU AI Act
Anthropic announced that its Claude models will automatically embed machine-readable watermarks in any text they process, aiming to comply with the EU AI Act, though the marks may also appear on lightly edited human-written content.
Anthropic said it will begin embedding invisible, machine-readable watermarks in all text generated or processed by its Claude models worldwide, not only in the European Union. The move is intended to satisfy the EU AI Act, which requires AI providers to label AI-generated or manipulated outputs with digital provenance metadata. The company noted that the watermarks will be embedded at the model level and will travel with the text when copied, though they may be lost if the content is edited by another system or stripped from images and video.
"A detected mark provides a signal that content was processed by Claude, but is not fully conclusive," Anthropic explained in a support article.
The approach goes beyond the law’s minimum, which exempts assistive functions such as grammar correction. As a result, even minor edits performed by Claude could trigger the watermark, potentially labeling human-authored text as AI-processed. Critics warn that the marks are easy to remove and may be misunderstood by users and downstream audiences, since the presence of a watermark does not guarantee that the content was wholly generated by AI.
Anthropic said it will eventually publish detection tools required by the EU regulation, but the effectiveness of the system remains untested. The company also acknowledged that the watermarks will not work on platforms that lack support for the required metadata. Failure to comply with the AI Act could expose Anthropic to fines of up to €15 million or 3 % of global revenue.