How AI text watermarking works
Summary
This article explains how AI text watermarking works, detailing how a secret key subtly biases word choices to embed a detectable mark without changing meaning. It covers detection, the impact of edits, and how different watermarking schemes operate in practice, including SynthID and Claude’s approach, with emphasis on the probabilistic nature of detection and the dependence on document length. It also notes practical considerations, limitations, and references to primary research.