Anthropic's Claude AI output watermarking sparks user backlash amid EU AI Act compliance

The watermark is applied at the model level and is text-native, meaning it is woven directly into the generated words and can persist through copying, pasting, and some editing. Files also receive signed metadata under the C2PA standard. While Anthropic has kept the exact watermarking technique secret, researchers speculate it uses a statistical signature similar to Google's SynthID Text.

The implementation has sparked significant user backlash, with many fearing the watermarks will expose their use of Claude for professional or academic tasks, likening it to a 'digital tattoo.' Open-source projects aiming to remove these watermarks have already appeared on platforms like GitHub.

Conversely, many users support the watermarking as a necessary measure for transparency and identifying AI-generated content, arguing that there is no valid reason to oppose it unless intending to deceive. Anthropic has stated that the mark proves Claude's involvement but does not guarantee the text was entirely AI-generated, and that heavy editing can strip the watermark. The company has not yet announced when it will release detection tools.