Posts Announce Watermarks in All Claude Outputs
Researchers share and discuss claims about watermarks in Claude model outputs.
Shannon Shen posted about the value of detecting model distillation via generation patterns. Aran Nayebi retweeted a post from ns123abc stating that Claude models will now embed invisible watermarks in every text output along with comprehensive metadata on files. The packet also contains two generated headlines describing the same update from Anthropic. No first-party statement from Anthropic appears in the supplied lines. The discussion centers on the technical possibility of verifying whether other models have used Claude outputs for distillation.
🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
Combined views
Reactions from ranked influencers
2 postsIsn’t it clear to people that the primary reason is to have some statistically provable way to tell if others have distilled their models? (I.e., when your model distills the outputs, you can tell based on its generation distribution?
🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
