It doesn't have to be doing gradient descent on a neural network to be machine learning. You could easily generate small perturbations to generated routes if they didn't add much time and took the driver down an under-monitored road.