Driving is a social act. Watch anyone drive for a day and you will see countless social interactions they have with other drivers. These myriad social interactions create a set of localized driving conventions which are not strictly described by the broader rules of the road. Sometimes the conventions even technically violate the rules of the road, but police do not act on the the violations (unless they are looking for an excuse to arrest someone) because they are accepted norms and allow smooth functioning of the road system.
Autonomous driving will likely initially consist of some sort of model AI for each model of vehicle. This AI driver model will be updated to the latest software version routinely, and that’s great. However, this AI driver model will likely lack regional variations, at least initially. Think of it as the standardized national accent used for broadcast television, instead of a bunch of regional accents.
When these nationally standardized AI drivers are introduced to the roads they will not know all the behaviors that local drivers share, some of which optimize use of local roadways under regional conditions. For example, in Los Angeles it is conventional for at least 2 cars to turn left when the light turns red at an intersection without a left turn signal. Even though the timing of the light cycle does not technically allow for both cars to turn left, there are so many situations in which left turns need to happen that the communal behavior is to allow for this so as to make things move more smoothly for everyone, even if only marginally better (this is Los Angeles after all).
Whenever autonomous driving software becomes widely available, it will presumably be required to follow all the driving laws laid out in the state’s driving rules. They will be locked into by-the-books driving methods, and for good reason! But this means that AI drivers will sometimes have no way to follow a local driving pattern with dubious connection to the formal rules of the road. In a situation like the Los Angeles multi-left-turners on red convention, autonomous cars will end up being more conservative than the local traffic pattern. And they will use up extra resources (e.g. extra stoplight cycles) in these certain situations.
This shouldn’t be a problem that lasts forever; the situation should resolve itself as more fine-grained AI models are released. Local driving patterns can be mimicked, but it takes extra work being done for that specific locality and that won’t happen for every region initially.
So when fully autonomous vehicles are initially introduced, the general driving public could see even more congestion in certain parts of their commute. For example, say they get stuck behind an entire line of AI drivers each taking an extra stoplight cycle to turn left.