There are many design decisions to make before even beginning to develop the systems that determine AI behavior.
First you need to ask what type of behavior would be appropriate. If your driving game is more arcade like and you wish to form an experience that is more fantasy than reality, the AI will be shaped to that vision in mind, all freedoms and limitations included.
I'm going to assume that you are referring to driving simulators by the content of your post. You're working with actual cars(or cars that could possibly exist) and physics honed with math and physics models that describe the forces of reality. So, ignoring the idea of 'fun' at the moment, let's consider the level of abstraction to which we simulate the actual 'driver' of the vehicle. Do we consider it to be an actual physical entity bearing the qualities of the human form and mind? Or are we simply going to emulate the decision making process of a human being while describing behavior that will allow for errors to occur in an expected manner?
The key difference being that the former is a living, breathing person. He sees using his eyes, drives by manually turning the wheel and reaching to shift, feels the swing of inertia and the pressures of acceleration. His knowledge is limited, his sense are imperfect, his distractions are infinitely variable and unpredictable based on any given factor moment to moment and day to day. He will grow tired, let instinct overtake training, lose battles of focus when his emotions run high from failure and success alike. His reactions are limited by the speed of signals in his nervous system, and his decisions are affected by a series of judgments that follow a line of reasoning dependent on potentially faulty information, or faulty interpretation. He can either fail or succeed in the same exact situation(or different situations that are similar enough, which are likely to happen even in the same race on different laps) in various ways.When racing at a professional level, success is determined by inches and milliseconds, and catastrophe can come from a perfect alignment of many factors occurring at once that would not even be noticed if they happened individually.
That would be a small selection variables to consider to even begin to replicate a human beings behavior while driving. Of course, that wasn't exactly your question, but all of those things are important for the aspect of realism, and emulating realistic qualities of the human is one way to model AI that provides a competitive experience that is less transparent. As you stated, AI is costly, and doing all of that(even if everything I mentioned is even possible to model given current science regardless of processing power) is no guarantee of success in realism or, most importantly, providing an enjoyable experience. The balance between accuracy of the simulation and the joy of the experience is a question of design that will always be completely subjective. It's also the relationship that provides the real issue behind game AI; as already stated multiple times in this thread, the real issue lies in handicapping the opponents to provide challenge without letting the player feel they were allowed to win.
Considering publishers actually want to make money off of the game, a half decade of development just to improve one aspect of the offline experience that has terribly high diminishing returns is not a possibility. So we model the driver more abstractly. He has an actual model you can view, his weight is probably considered, he may even have to actually move his hands to shift and his feet to accelerate/break, but all the finer sufferings of the human experience are ignored. With that in mind, he's little more than a very basic finite state machine that's working with a very small amount of variables. If the car loses traction at this speed range in this direction with/without obstacles, apply breaks/let off acceleration until friction coefficient = whatever it needs to be based on a formula that the 'driver' has computed before the next frame is done rendering. That's how most decisions are made in games, adopted to your genre. I'm sure there's some pretty complex stuff going on and i'm being pretty reductive, and i'm no authority on AI or math or physics, but that's pretty much it from every talk i've seen. Granted, "pretty much it" is one or more programmers who could do whatever they want in the general field of computing working 50+ hours of hard fucking labor with software packages written in house by other people, constantly adjusting to feedback from an army of game testers who are not required to have any technical knowledge beyond what is need to be known in order to describe a situation clearly.
Long, rambling and half off topic answer, but the short of it is: this shit is hard, so hard that no one, in any field, has really done better than a hyped up state machine. The field has barely changed in any practical manner for decades, except for:
If only we had neural net procesers.. Some form of learning computers.
Reading headlines about neural networks trained to drive cars and trade stocks may lead you to believe they are an end to any solution regarding automation, but the truth is that they are limited by individual context and are at least as complex to approach from a design perspective alone. We want thing to do stuff: What are we actually asking it to do? Can we provide information that will allow it do that? Do we know how to tell it
how to learn with what we can provide? Then, applied to the context of video games: can we even get it to do what we want in a way that's adjustable in difficulty? Is what it accomplishes appropriate for what we want out of the game? Because most behavior resulting from neural nets produce solutions that are completely incomprehensible to anything mankind has ever made or would have likely thought to make. Given the rules of Forza, and the goal we want, is this thing going to fling itself from the dirt to the track in order to drift down a straight segment, somehow winning against all static models?
Neural nets are terrifying because they show us how constricted the human perspective is, which is probably one of the reasons we have difficulty recreating what we do single moment of conscious existence.
Another huge question regarding this entire subject with respect to design is: Are we modelling the actions of a human driving a car, or are we modelling the actions of a human playing a game that simulates the act of driving a car? I'm assuming the latter will provide a better playable experience, but a less believable reality.
Yet F.E.A.R achieved the perfect AI. It can be done.
F.E.A.R. used a surprisingly simple model that was focused on taking individual agents and giving them set goals, while providing a general tool set to allow those goals to be accomplished.
So instead of: If see player, then shoot, else if see player and near cover, then shoot while moving to cover;
it was: Ultimate goal: Reach player in order to shoot him, but do so in a manner that will give him the least opportunities to shoot you, using cover, only running while provided with cover fire(that is communicated instead of known by all agents as if by telepathy) by teammates that are able(also know when the option for cover fire is available), so on and so forth. The results were a more sandbox like feel, more varied due to the limited knowledge enemies had of both their teammates and the player, which required far more caution and allowed a more gradual experience to play out. It also allowed for some emergent situations where you would be flanked, as the enemies would act in manners with the intention of distracting the player from an enemy that is advancing with the knowledge that it's teammate is currently taking fire, and knows how to avoid line of sight. Also, even on the hardest difficulties, they were prone to miss, so no sniping headshots, no grenades lobbed into your shirt pocket from 50 yards, and that part was simply done well from an artistic perspective. It's not an exact science.
The problem is adapting those principals to a game where you're racing vehicles. Completely different arena, and the reasons F.E.A.R. seemed so far beyond had a lot to do with the unpredictable nature of a group of enemies actually working together under the constraints of reality. Before that(and too commonly, since then), agents in shooters just worried about shooting you as soon as you were seen, maybe protecting themselves, maybe taking advantage of where you were looking. Sometimes they would even team up in an aggressive way that at least forced you to think on your toes(Halo comes to mind), but they were usually all knowing, and that was where difficulties become cheap and the magic was revealed to be machine like.
AI has even less of a chance to improve in the coming years due to the popularity of adding multiplayer, regardless of whether it's appropriate or not. Until neural nets become so advanced that they can be formed to any occasion with any goals and data provided, we'll be facing the same predictable meat puppets and four wheeled bumper carts we have been since doom and mario cart.
