Anthropic will embed invisible watermarks in text and add signed metadata to images generated by its Claude AI models. This move is to comply with the transparency requirements of the European Union's AI Act, which took effect on August 2, 2026, but the company is applying the policy globally.
The text watermark works by subtly influencing word choices to create a detectable statistical pattern, which will persist through copy-pasting and some editing. However, the watermark's presence only indicates that Claude was likely involved in processing the text—not that it authored it—and significant rewrites can remove the mark. Anthropic plans to release tools to allow users to detect the watermarks.