NLP has been looking at and using MGC for decades and has done fairly well in certain cases (these being topic-specific, data-driven, heavily trained, etc.).
The major change is the size, scope of models, and the availability (or easier implementation) of scripts that access those models.
Where before you had to know the data structures, and/or linguistics, or implement systems that you had to train, etc., now you can just connect to an API with parameters.
Google doesn’t really care if it’s MGC, human-made, hybrid, or produced by blind mice singing psalms — so long as it’s not garbage, they’ll take it 😀