Is it free software?

I don't think anyone ever contested that it is derivative of the training data, just to which extent it is derivative of any particular piece in the training data and if that meets the threshold for copyright to become an issue.

Technically speaking you could e.g. make a model that is a single bit and reflects if the input has an odd or an even number of characters and produces output with the same characteristics. Most people would agree that that would not violate copyright.

On the other end of the spectrum you could just make a "model" that is a filesystem folder of all the input data and the algorithm just selects a random section from the input verbatim. Most people would agree that that should violate copyright.

The question is where in between those two extremes do we draw the line where copyright no longer applies.