I think it's more complicated than that. Training does a lot more than just make models imitate the highest quality training data, and even what high quality training data includes can be subjective depending on your goals and tastes. And a lot of effort does go into making sure the models don't have "bad personality" - I'm sure a lot goes into making sure they trend towards appropriate reading levels and various other things that aren't strictly about being a Mark Twain or JRR Tolkien level writer.