Anthropic publishes the watermark spec: it is SynthID-Text

A couple of days ago, I posted a step-by-step of how Claude's watermark probably works.

Anthropic has now published the actual spec.

Instead of splitting the vocabulary into green/red lists, and nudging the odds toward green (KGW), Anthropic is using Google's SynthID-Text, which was the other approach the infographic pointed to.

Normally, a standard computer random number generator picks which word wins based on those odds. SynthID-Text replaces that generic randomness with a deterministic "pseudo-random" choice driven by a secret key and the previous word(s).

The model's probability distribution stays exactly as computed, meaning the output stays statistically identical to unwatermarked Claude.

The full confirmed mechanism is in the infographic below.