There is an interesting exchange going on between Salvador Cordova of Uncommon Descent and Dave Thomas of Panda's Thumb.

First, Dave posts about his genetic algorithm for finding Steiner trees ("the shortest networks of straight-line segments connecting a given collection of fixed points"). He claims his algorithm (explained in detail in an earlier post) is not targeted because it does not contain the optimal solution, but selects for incrementally better ones randomly. He challenges ID'ers to come up with a solution to the 6-node Steiner Tree problem. I think his aim was to prove how much better his genetic algorithm was than the intelligent attempts of design proponents.

Next, Salvador responds with a genetic algorithm he designed, which correctly finds the sum of the first 1,000 integers. The algorithm does not have a direct target (i.e., it is not explicitly searching for the answer), but instead uses Gauss's algorithm - obviously targeted in a secondary sort of way, which he implies is akin to Dave's Steiner tree algorithm.

Finally, Dave insists his algorithm is not targeted and is in no way analagous to Salvador's, which is blatantly targeted, and Salvador replies that of course his algorithm is targeted, and that the smoke and mirrors in his algorithm are indeed similar to Dave's.

The contentious issue here is what, exactly, constitutes a "target". Salvador's point is that if your cost function is intentionally designed to approach the solution you are looking for, the absense of the actual optimum value is a technicality. Dave's algorithm is clearly only good at finding Steiner trees because he designed the cost function to reward better Steiner trees. As johnnyb points out in a comment to Salvador's latest reply:

The problem is that the fitness function cannot be directly related to the problem being solved. Natural selection simply says the animal must survive or be better at reproduction, but somehow guides processes such as eye formation. Therefore, in a fitting scenario the selection algorithm should not directly correspond with what you are searching for, in fact there should be a large disconnect.

 This is exactly right, and so succinct I wish I'd come up with it myself. There may or may not be a genetic algorithm for constructing a structure like the human eye (depending on the number and structure of interdependencies between its parts). If one exists, then its cost function will be specifically designed for producing an object like the eye, whether the "target" is explicit or not. There is no evidence that a vague cost function like "survive" is capable of doing anything interesting.