> They're an artifact of a training process, not code that was written by someone.

If that were relevant to the licensing discussion, then you'd have to consider every "generated" parts (interfaces, dataclasses, etc) of every open source project artefacts. Historically, that was never the case. The license doesn't care if a hardcoded value was written by a person or "tuned" via a process. It's still source code if it's the preferred way of modifying said code. And it is. You can totally edit them by hand. It would not work as well (or at all), but you could do it.

There is actually a gray area about what code "counts" as source code to the point where you would consider it "open source" if it were licensed as such. I think if you had a repository consisting of only generated code and not the code used to generate it, it would definitely raise the question of whether it should be considered "source code" or "open source", and I think you could make arguments both ways.

On the other hand, I don't really think that argument then extends to model weights, which are not just some number of steps removed from source code, but just simply not really related to source code.