eicker@lemmy.world to Technology@lemmy.worldEnglish · 26 days agoHow Claude marks AI-generated content: (1) watermarks embedded in text, and (2) signed provenance metadata attached to files.support.claude.comexternal-linkmessage-square10linkfedilinkarrow-up134arrow-down13 cross-posted to: hackernews@lemmy.bestiver.se
arrow-up131arrow-down1external-linkHow Claude marks AI-generated content: (1) watermarks embedded in text, and (2) signed provenance metadata attached to files.support.claude.comeicker@lemmy.world to Technology@lemmy.worldEnglish · 26 days agomessage-square10linkfedilink cross-posted to: hackernews@lemmy.bestiver.se
minus-squarebrucethemoose@lemmy.worldlinkfedilinkEnglisharrow-up5·edit-226 days agoNah, token biasing can be really subtle. Enough pairs of one word being slightly more probable than the other, and it’s imperceptible, and below the “noise floor” of the sampling itself.
Nah, token biasing can be really subtle. Enough pairs of one word being slightly more probable than the other, and it’s imperceptible, and below the “noise floor” of the sampling itself.