Most self driving systems today are using neural networks, I don't think expert systems would be able to manage the complex issues that pop up with driving.

As a result, I don't think we could really understand all the nuances that these systems are capable of knowing once they have been trained for so long to drive. It's easy to create a network that learns to do a task relative to learning how the completed system is thinking.

These aren't brute force approaches anymore, there is a system that is learning as time passes, and getting better at what it does as time continues forward.

As such, these networks have one job, and do that one job very well. In this case that job is to prevent crashes.

Fiddle with that network to impose artificial limitations and you impose on a system optimized to do something, and more crashes will result in the long run. Although I'm sure there are cases where things go wrong with the program, or things need tweaked, these aren't the same as directly interfering with the car when it decides to take a course of action that could lead to hitting a schoolbus vs a normal car. It may well be that hitting the schoolbus causes less total harm for some reason, and we be sure that we understand the reasoning of the machine before we decide to mess with it.