Take the sentence “The climate at this time was chilly and…”. The following phrase may be very unlikely to be “sugary.” However it’s fairly more likely to be “overcast” or “gray.” Beneath most circumstances, it doesn’t matter a lot to the reader which of those latter two phrases the mannequin in the end chooses—the which means of the sentence is essentially the identical both means. In instances like this, the selection is settled by a random quantity.
Watermarking makes use of low-stakes selections like these—which happen many occasions over a bit of generated textual content—to go away a sample in Claude’s responses. That sample is undetectable to the reader, however is detectable to anybody who has a key that encodes it. When watermarking is used, selections are nonetheless made at random, however the supply of the randomness is totally different. As a substitute of utilizing an arbitrary random quantity generator to select the subsequent phrase, watermarking makes use of the important thing and some phrases that come earlier than to settle what phrase the mannequin ought to choose.
