When you are an novice artist it's usually bad thing to rely on your mental model of reality. Children draw like that. Instead it's better to just observe patterns of light and darkness on the model and reproduce them. But at some point of that, when you start being creative with your art it really pays off to actually know how many fingers humans do have.
It doesn't mean neural networks are completely incapable of creating the model of reality, AlphaGo creates models of Go reality that's better than human players, they just can't do that through language alone.
Surely there's some jokes in the training database... enought to allow generating "new ones" by switching pieces
The question about: are humans doing differently is the one about creation. Let's say that maybe 90% a "just" simple pattern matching (same mechanic as already existing jokes) and 10% is new mechanic (original joke)
If not, how does it work exactly? Is there any difference between "a sophisticated pattern matching on a (very huge) corpus" and how a human would invent a joke?