SynthID-Text draws several candidate tokens from the model's unmodified distribution and compares them through keyed tournament rounds. Green-list decisions select winners without applying one fixed bias directly to every preferred token logit.
Detection repeats the keyed verdict process across tokens and rounds, then aggregates the evidence. The design aims to preserve response quality, but detectability still depends on output length, token entropy, key access, and whether later editing or paraphrasing disrupts the signal.
ELI5
SynthID-Text is Google's keyed statistical method for placing a hidden pattern in language-model output. It samples several likely token candidates and uses secret-keyed tournament choices to select among them without directly adding one fixed bias to every token.
For example, an authorized detector can repeat the keyed decisions across a long passage and measure whether the expected pattern appears. Short text, low token variety, editing, or paraphrasing can weaken the signal.
How does SynthID-Text differ from fixed logit-bias watermarking?
It selects among honest samples with keyed tournaments rather than adding one constant bias directly to preferred token logits.
Does SynthID-Text make an AI watermark universally visible?
No. Detection requires compatible keyed analysis, and short, low-entropy, or paraphrased output can weaken the signal.

