thedaian's answer is a good one since it provides examples.
I think the main problem seen with these approaches is one of control. Once you start dealing with heuristics, which these ultimately rely on, it becomes very hard to evolve a system that does what you want it to do. Quite frankly, it's a matter of project cost and risk. If you have a game idea that you can implement without using these, or using some more direct procedural approach, the better for you.
The problem with even the most elementary AI approaches (take A* pathfinding for example) is that they take some time to first understand, and considerably more time to master. I think this keeps us closer to simpler, more static approaches. It doesn't feed innovation, but a quick return on investment is a hard thing to ignore.
Having said that, I'm definitely on the side of "If Nobody Tries It We'll Never Know". A GP project has been brewing in my mind for some time.
(Another approach you might want to look into is neural networks, it's often grouped with these as being a machine process that mimics natural forms of improvement -- in this case learning -- through elimination.)