A good way to look at AI as psychology. The idea that to understand how a natural intelligence works, one tries building a working replica. Once we have several going, we can ask which of them (if any) is most likely to be embodied in the brain, and how. This helps suggest psychological and neurophysiological experiments.
Probably most successful so far in low-level vision, where we have fairly tight constraints on the task, the hardware available, and the structure of the information available (geometry of 3-d space, physics of light, rigidity of most objects, etc).