There are two things where human beings far outpace AI in competence. The first is the ability to recognize when inputs just simply don't make sense and thus the intended algorithm is going to return garbage results. A related, but important issue (particularly if you talk about networking cars together) is being able to recognize malicious inputs.

The second issue is that humans are good at higher-order planning. An example as it relates to driving is what we do when we can't tell the state of the traffic light (direct sunlight, say). Humans naturally look for ways to indirectly infer the state: the status of the traffic light in the cross direction, the current disposition of cars around the intersection.

Autonomous cars will only be better at handling these situations if the engineers are aware that these situations need to be handled and take measures to make the cars aware of them. And history suggests that is sufficiently unlikely to happen to make me fearful of the future.