ANTHROPIC has disclosed the mechanics of its text watermarking system in compliance with the EU AI Act. Unlike traditional watermarks, Claude uses a covert statistical algorithm, integrated within the text without any visible symbols or quality degradation. The algorithm selects words based on a cryptographic key rather than standard random number generators, allowing it to leave an undetectable statistical footprint.
However, the watermark can only identify Claude's involvement and cannot determine if it solely generated the text or edited human content. Its effectiveness is limited with short texts, factual statements, and coding, as these provide less flexibility for word substitution. Anthropic plans to release a detection API for authorized users and will apply watermarking to all Claude models post-August 2, 2026. The watermarking serves as a compliance tool but is not definitive proof against plagiarism or copyright issues.