I've been thinking this for a while. Attribution is sorely needed, I believe there was a recent paper on this so maybe it will happen.

In the meantime the only way to really sort this is to have models that have only been trained on some particular kind of license - there is quite a big corpus of GPL'd code out there so a GPL based model could potentially be one of the first, and of course the output could only be GPL.