SynthID-Text Playground

Method

Sentence being built

The finding was …

Token trail

first 3 words are prompt context (✓), not scored

  1. ✓The
  2. ✓finding
  3. ✓was
0.00cumulative evidence — not a probability

No text yet

Tokens scored: 0

1. The model predicts what could come next

Before any watermark exists, the language model ranks plausible next words with ordinary probabilities.

Model probability affects which words get sampled. The secret-key score decides the winner among those sampled words.

  • “clear”model 32%
  • “strong”model 27%
  • “convincing”model 23%
  • “mixed”model 18%