[Claude-generated text will now include 'invisible watermarks,' and minor edits may not remove them] According to monitoring by Beating, Anthropic has started embedding machine-recognizable 'invisible watermarks' in text generated by Claude. This applies to Claude, Claude Code, Cowork, API, as well as supported models on AWS, Google Cloud, and Microsoft Foundry. New models will support this feature upon launch, and existing models are gradually being updated to include it. This change is primarily aimed at complying with the transparency requirements of the EU's AI Act, but Anthropic will enforce it globally. The watermark is embedded directly into the text generated by the model, rather than being added as metadata to the file, and it remains intact even after copying and pasting. Anthropic has not disclosed the algorithm yet, but the official description suggests it is more akin to Google's SynthID statistical watermark. It likely does not involve hidden zero-width characters but instead subtly alters the token selection probabilities during generation, leaving detectable statistical signals in the text. The official statement notes that if the text is too short, there may not be enough signal, and significant rewriting, translation, or mixing with other content may also render the watermark undetectable—this aligns with the characteristics of statistical watermarks. Therefore, simply changing a few words may not be effective; substantial rewriting is more likely to disrupt the watermark signal. Anthropic also cautions that detecting a watermark only indicates that the text may have been generated by Claude, but it does not prove that the entire piece was written by Claude. [Original link]