> I can imagine good enough AI being able to spot truth even better than what humans do

The thing that makes me question that is the data that is used to train those models to begin with. To disect truth on the internet, can you use the internet as a source of truth to train it?