> It’s unlikely that model weights can be copyrighted, as they’re the result of an automatic process.
If they can’t for that reason alone, then the model is a mechanical copy of the training set, which may be subject to a (compilation) copyright, and a mechanical copy of a copyright-protected work is still subject to the copyright of the thing of which it is a copy.
OTOH, the choices made beyond the training set and algorithm in any particular training may be sufficient creative input to make it a distinct work with its own copyright, or there may be some other basis for them not being copyright protected. But the mechanical process one alone just moves the point of copyright on the outcome, it doesn’t eliminate it.
That isn't necessarily true. You're saying that the model weights would be a derivative work. Derivative work is insufficiently transformative to fully distinguish the output from the input.
In this instance, it's very likely that the process is sufficiently transformative. A set of model weights look nothing like the Mona Lisa, nor can they be directly transformed back into it. What it is NOT is the product of a creative process on the part of a human, and is thus ineligible for copyright.
It is as though we are able to distill meaning using an automatic process. Copyright doesn't protect meaning, only expression, and it only protects expressions that were generated by humans.
The network itself might be patentable if it isn't obvious.
If a network was copyrighted and it was found that the function of the network was inseparable from its expression, it would no longer be eligible for copyright on that basis.
I think people often misunderstand the application of copyright to GPL code in this way.
Now, will any of this stop people from CLAIMING copyright? No. It'll have to be fought out in courts.
If they can’t for that reason alone, then the model is a mechanical copy of the training set, which may be subject to a (compilation) copyright, and a mechanical copy of a copyright-protected work is still subject to the copyright of the thing of which it is a copy.
OTOH, the choices made beyond the training set and algorithm in any particular training may be sufficient creative input to make it a distinct work with its own copyright, or there may be some other basis for them not being copyright protected. But the mechanical process one alone just moves the point of copyright on the outcome, it doesn’t eliminate it.