Claude Begins Adding Text Watermarks to Comply with the EU AI Act

HarperSCarroll · x · 2026-08-12

To comply with the new EU AI Act, Anthropic is adding watermarks to the text outputs of its Claude models.

How it works: AI generates words based on probability distributions. The watermark slightly shifts the probability distribution of the output text away from its normal path. This watermark is invisible to the human eye, and currently, only Anthropic can detect it, though they are working to enable third-party detection. For images, Claude adds a signature to the metadata, which is easy to remove.

Industry Context: Google has already implemented watermarking via its SynthID technology. With the new EU regulations, AI watermarking is set to become an industry-wide standard.

Related event: Anthropic Adds Invisible Watermarks to Claude Outputs for EU Compliance(3 posts)→

Original post →

More from Models

Models channel →