https://news.ycombinator.com/item?id=35292445
But like, to repeat some of it in a different way for this slightly different context: for that argument to work, Facebook would have to be actually doing something as the author that wasn't just automated in this codebase.
Like, if you use GIMP to work on an image, and then I download GIMP and merely run it... it doesn't do the same thing, right? You--the artist--were actually important in that story, because you provided the expressive intent that led to the resulting image.
But, in this case, the model is merely the result of running that code, not someone using that code as part of their own work: the model is a reasonably-deterministic output of code licensed under the GPL being run on data notably owned by people other than Facebook.
Imagine if, instead, you wrote a program that used GIMP to automatically create a really fancy image. You put a lot of work into the script to generate that image... and then you released that code under the GPL. I think you would be hard-pressed to argue that the GPL wouldn't have something to say about this output.
(I could maybe see an attempt at an argument that the GPL is an awkward license to apply to things that aren't programs; but, a model is in fact a program designed to execute in an interpreter on still yet other data as input, which makes this whole thing feel like a program which algorithmically generates the code for another program, which is actually quite a common use case for the GPL.)
To be clear: I think Facebook doesn't own the copyright, not that the GPL infected it; but, if Facebook DID own the copyright, AFAIK their only expressive input comes in the form of this GPL codebase. (See above linked comment for more exposition of the possibilities. Also see that comment for a more extensive "IANAL" disclaimer, but: I am not a lawyer, no matter how much I focus on copyright issues.)