The video begins with the hard red-list method proposed by Kirchenbauer and colleagues. A secret key and the previous token seed a deterministic function that divides the vocabulary into green and red lists. The model is forced to sample only green tokens, allowing a detector with the same key to replay the assignments, but the restriction can damage quality when the correct continuation falls into the red list.
Soft watermarking avoids banning red tokens. It adds a fixed bias to green-token logits, so high-probability answers such as Paris can still win while ambiguous, high-entropy choices lean toward the green list. Detection then becomes statistical: the detector calculates whether a passage contains more green tokens than chance would predict, commonly using a one-sided z-test.
Google's SynthID-Text replaces the fixed logit bias with tournament sampling. It draws many honest samples from the model's original probability distribution, pairs them in repeated rounds, and uses keyed green-list decisions to choose each winner. Google reported no statistically significant quality difference in an analysis of roughly 20 million rated watermarked and unwatermarked responses, while the detector aggregates the keyed verdicts across tokens and rounds.
The method has practical limits. Short answers and low-entropy tasks may not offer enough interchangeable token choices to encode a detectable pattern, while paraphrasing with another model can weaken or remove it. Detection also requires access to the secret watermarking key, so a provider can identify its own outputs or share a detector without making the watermark universally visible.
Watch on YouTube
