Thats a fundemental misunderstanding of AI. Even AI is configured to make decisions within a defined set of parameters and thats not "go drive a car" it will be broken down into a number of logical components that get joined together.It's not being 'programmed for', that's what makes this AI rather than what we've had so far (which has far more programmed behavior). It's been trained with real world experience, just like how human drivers learn.
There's a set building a picture of the environment, whats around it, what the classification is of those objects etc. Thats where AI can play a part becuase it can take images and compare them to others to work out whats a car, whats a road and where its edges are, whats a pedestrian, whats a traffic light, what colour is the traffic light etc. Once it has the environmental map ot can then try and apply a mixture of rules based skills (maximum speed), where to position itself in lane etc, After that you can layer on top navigation and lane skills to determine whether it wants to change lane etc.
They've previously talked about a model that was purely designed to determine if a car in an adjacent lane might enter the lane you are in, trying to pick up of the nuances of behaviour of a driver in a differnet lane, reflecting that cars drift towards the inside of the lane on a curve, but what about a driver drifting because of a cross wind, a large puddle they've spotted and want to avoid, a car in a further lane over that the car in the adjacent lane decides to give a slightly wider birth to, none of which are likely to result in them coming into your lane. Its the level of finesse here on just one aspect of driving, judging what the car next to you might do that has already got monumentally complex and the data you'd need to take in to train it, and that specifric AI model won't pick up on many of those aspects unless the AI is fed with all the potential inputs including observed behaviour of that driver from the first time it appeared (have yuou never spotted somebody driving erratically from a distance and they've stood out to the extent that you've paid extra special attention to them?). What about the things you've never seen before but you instinctively know what to expect, an accident ont he opposite carriage way, a chunk of wood in the road, a fishtailing trailer up front of the car next to you and not you, This is the local minima issue - you assume that your inputs and freedom of response caters for everything but what about the stuff that never been seen or happens so infrequently its not learnt how to deal with it?
Its far to easy to think you can just stick all the data into one big model and think its going to pop out an answer telling the car what to do, but it really not that simple.