The idea that's been floating around in my head for the last few years is something like "it's being trained by the data produced by people, it's going to have many human flaws as a result"