What's interesting is the "trolley problem" of driving safely and following the laws vs. being predictable to other drivers.
Human drivers are bad, but they're bad in ways that are often predictable. They frequently break laws, but in ways that are predictable. Should AI-driven cars also break those laws to be predictable to human drivers? Or should they break the same laws that human drivers break in the same ways so that the human drivers aren't surprised?
You should actually break the law to follow the flow of traffic.
Speed limits that are set too low, for example, are known to increase traffic accidents. Driving at the rate of other drivers (minimizing the speed delta between other cars) reduces the risk of accidents.
Laws that don't respect real human behavior can be dangerous, and in the case of vehicles that danger manifests in destruction and death a meaningful amount of the time.