Anthropic is rolling out invisible machine-readable watermarks for all content processed by Claude models. This move is part of the company’s effort to comply with the European Union AI Act, which requires providers to mark AI-generated or manipulated media. While the law grants a grace period for older models, Anthropic is applying these markers to all new models effective immediately on a global basis.
The implementation strategy is broad. These watermarks are not limited to text generated from scratch. They also appear on content where Claude performs minor assistive tasks like grammar correction or simple text editing. Because the system cannot distinguish between wholesale generation and minor tweaks, human-authored text processed by the model may receive the same designation as fully AI-generated work.
Technical experts note that these marks remain invisible to the average user and travel with the text when copied or pasted. However, there are significant limitations to their reliability. The marks can be stripped through metadata editing, screenshotting, or by pasting the text into other AI tools that perform their own modifications. Furthermore, the presence of a mark confirms only that Claude touched the file, not that the output is entirely synthetic.
Anthropic acknowledges that these labels may occasionally overreach by marking content that falls outside the scope of EU transparency mandates. The company intends to release a detection API in the future to help users identify these signals. For now, the rollout remains a primary test of how AI companies balance legal compliance with the practical realities of standard user workflows.