This gets into interesting weeds. It might get to a point where to actually "upskill" it might have to start sounding less and less human, "less and less agreeable to humans" being a waypoint to that ...

... to where, at some point, it might even begin to construct semantic loadings (heh) that are completely ininteligible to us while still superficially sounding like something we'd recognize.-

This is basically what I believe, too. Its outputs are generally syntactically correct human-language sentences, but semantically, and especially at higher levels of abstraction, it is no longer correct to think of them as human language.