Anthropic to start embedding invisible watermarks

Archive.is link

Does anyone have a link to the method? Does it need Unicode or is it ASCII spacing? Word frequency bias? Can a human inadvertently trigger a detector for it?

ICYMI: https://news.ycombinator.com/item?id=49303350

Basic idea is that when picking a next token from the probability distribution from the model, they use a pseudorandom generator (ie deterministic) whose signature they than can pick up.

Not sure how text (ascii) can be watermarked ...

https://arxiv.org/abs/2301.10226

Just to be clear, as I understand it, that's not the method.

Ascii can't but I guess most of the text on the internet is unicode which easily can be.

biased word choice