1) Regarding the fact that at any given time, there is knowledge that people have that hasn't been encoded in a form usable by the optimizer: I think this can be addressed by generating a set of decent designs with machine learning (on the pareto front, if you will) and then providing user interface tools to allow the humans to pick the designs that best handle the knowledge that wasn't optimized against. These user interfaces will often have to be quite elaborate but that's a tractable problem.
2) When working with Genetic Algorithms, and any machine learning technology, really, my experience has been that the first results only show what is wrong with the problem setup. The designs that are returned by a genetic algorithm are invariably nonsense at first and the then it's a matter of playing whack-a-mole as the optimizer exploits inaccuracies in the problem statement, a fix is made, a new exploit is found, repeat. This is often viewed as a deficiency of genetic algorithms but in my opinion the fact that the metrics being used can be gamed is valuable information that needs to be addressed as early as possible.
Finally, I understand my original comment came across as somewhat naive. If we ever get to a point where we can design physical products, I am imagining starting with something along the lines of a mechanical pocket watch, or maybe a nice desk chair. Airplanes are gonna be last.