Comment by friendzis

> Is something that in no way or shape resembles a copyrighted work (i.e. a model) actually distributing that work?

Does a digitally encoded version resemble a copyrighted work in some shape or form? </snark>

Where is this hangup on models being something entirely different than an encoding coming from? Given enough prodding they can reproduce training data verbatim or close to that. Okay, given enough prodding notepad can do that too, so uncertainty is understandable.

This is one of the big reasons companies are putting effort into the so called "safety": when the legal battles are eventually fought, they would have an argument that they made their best so that the amount of prodding required to extract any information potentially putting them under liability is too great to matter.