Theorizing: If the network acquires a pattern of understanding that characterizes the correct response (in other word, if the network "groks" a topic), it will express its knowledge internally as "a correct model (small, efficient)" and "various mistakes that people often make (big, uncompressed)". In this case, the correct pattern should get by far the most reinforcement (because it's a smaller target), and should usually be selected to generate the answer unless the network has a particular reason to expect itself to be playing somebody who's mistaken.

This is also how humans learn, although we can do this process purely internally.