How AI text watermarking works

At each step a language model produces a probability distribution over possible next tokens, and usually many are equally acceptable. A watermarking scheme uses a secret key to bias that choice in a reproducible pattern. The text reads normally, but across hundreds of tokens the pattern is statistically measurable by whoever holds the key.

Google DeepMind's SynthID-Text is the published implementation of this idea and the basis for Anthropic's approach. In DeepMind's own words, SynthID 'adjusts these probability scores' to generate a watermark, rather than adding anything to the output.

Three consequences follow. The mark cannot be seen by inspecting characters. It degrades gradually as text is edited rather than breaking cleanly. And detection requires the key, which is why no public text detector exists — including from Google, whose SynthID Detector portal accepts image, video and audio but not text.