Anthropic plans to offer a watermark detection API that lets third parties check whether text was written by Claude, according to The Decoder. The method builds on Google’s SynthID approach and changes word-selection randomness while Anthropic says text quality is not affected.
The feature is aimed at provenance: helping schools, platforms or organizations identify AI-generated writing from Claude. Unlike visible labels, statistical watermarks are embedded in generation patterns and require a detector to interpret.
The limits are important. The report notes that the approach is weaker for fact-heavy text, code and heavily rewritten output. Those are exactly the categories where many professional users may want reliable attribution.
A detection API could make Claude output easier to audit in controlled workflows, but it will not settle the broader AI-writing debate. Watermarks work best when users generate text through systems that preserve the signal and when organizations understand the detector’s error boundaries.