Having spent a lot of time with PicBreeder, I suspect that the "ancestors look nothing like their descendants" thing is more an artifact of how our brains interpret images, rather than implying some deeper insight into how objectives affect the process.
There are AI projects that attempt to approximate a "natural image manifold" – essentially a blob in the space of possible images where you find things that look like photographs to humans. I think the PicBreeder thing has more to do with that; the set of images that people want to create occupies a relatively confined space within the set of all possible PicBreeder outputs. So working off of any previously built image is basically a "shortcut to the manifold", if you will.
Or to put it another way: when you consider the entire space of PicBreeder outputs, any existing image is much closer to whatever you want to make than starting from scratch.
There are AI projects that attempt to approximate a "natural image manifold" – essentially a blob in the space of possible images where you find things that look like photographs to humans. I think the PicBreeder thing has more to do with that; the set of images that people want to create occupies a relatively confined space within the set of all possible PicBreeder outputs. So working off of any previously built image is basically a "shortcut to the manifold", if you will.
Or to put it another way: when you consider the entire space of PicBreeder outputs, any existing image is much closer to whatever you want to make than starting from scratch.