The point is that it's not, but making it so is a design goal of the paper.
Example: Mario immediately jumping into a pit at level 2. According to the learned utility function of the system, it's a good idea. According to ours, it's not.
Just as with optimizing smiling faces. But while that one was purely a thought experiment, this paper presents a practical, experimentally testable benchmark for utility function learning, and, by the way, shows a not-yet-perfect but working solution for it. (After all, Mario's Flying Goomba Kick of High Munchkinry definitely satisfies our utility functions.)
"Pretty simple" algorithm playing games quite impressively.
http://www.youtube.com/watch?v=xOCurBYI_gY
First, this is awesome - enjoy!
Paper here http://www.cs.cmu.edu/~tom7/mario/mario.pdf
One interesting observation made by Tom Murphy is that the AI found and exploited playable bugs in the game not (commonly) known to human players. I think it's a good example to have available suggesting what a really smart AI might look for to win.