Deep learning is only applicable to systems that almost never change. In the paper "Deep Learning: A Critical Appraisal", Marcus describes how deep learning was able to learn how to play Breakout to an acceptable degree. But when a small wall was added to the middle of the screen, the deep learning AI completely broke down. Now imagine this same type of AI being employed on real life roads, in which the environment changes often and unexpectedly.

Deep learning AI is flat out unacceptable for driving. "Unsafe at any speed" well describes it.