I think that the main hitch is that the generation of content isn't strict enough in terms of following certain rules. It's all fuzzy logic, and that's why weird glitchy faces are generated sometimes. What needs to happen is that the fuzzy logic is strictly constrained by some explicit rules, like the flow of causality. Otherwise there will be way too many plot holes and plots that break causality.

Basically I think AI needs the capacity of reality testing.