Anthropic adds statistical watermarks to Claude text
Anthropic is introducing a watermarking system for Claude-generated text to meet requirements of the EU AI Act. The method, based on Google’s SynthID-Text, changes the source of randomness used when selecting words, producing a statistically detectable pattern without visible marks or hidden characters. Anthropic says the system does not affect content, creativity, or readability and is being deployed globally because it cannot be limited by region.
Blogger John Gruber disputes that assessment. He argues that synonyms are not fully interchangeable, so choosing between words according to a secret watermark key rather than exact meaning can reduce quality. He also says the quality measures in Google DeepMind’s SynthID research do not adequately capture such differences. The watermark can be removed with paraphrasing tools including Declaude, whose developer James Padolsey has called the underlying EU regulation arbitrary.
An analysis of legal use found few immediate concerns, but noted possible complications where clients prohibit AI, judges distrust it, or clients seek fee reductions. Markings may persist in templates and overlap when several language models contribute to one document. Anthropic says watermarking is less dense in fact-heavy passages, where fewer word choices are available. Claude models released after August 2 support the system; older models are expected to receive it later.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.