We actually don't understand the inner workings of most neural networks, this is a well-known problem. We can evolve them to solve a problem but we won't necessarily know how they are solving it.