The thing is, the term bots might be the more accurate of the two in all cases.

I guess it depends on your definition of AI. How much do they need to actually think for themselves before we can genuinely call them AI instead of bots?

A player named Sophie driving Gran Turismo Sport is the only real example of AI in racing that I'm aware of.

However, she is not in the game, she has been playing the game. And playing like a real-life player, practising and learning from mistakes, exploring the track limits, and trying to find other potential exploits.

She shows what's possible with enough computing power. And that brings us to the racing game developer's dilemma.

Do they create a program which makes the AI (we're using it as a common slang term now) use all the same in-game physics details as the player?

This makes for a better experience overall because track conditions (weather, track temperature and more) and vehicle condition (like tyre wear, fuel load and damage) will affect their performance throughout the session. It can be rather taxing on processing power though (not just CPU performance, but also memory speed and other hardware parameters), therefore reducing the potential player base to those with a more powerful, and therefore more expensive, PC.

Do they simplify them a bit so a higher-end PC isn't needed, making the game playable to a wider potential customer base? Or are they restricted by the processing power of a particular platform and need to make them so simple that they're just mobile chicanes unaffected by even crashing into them?

Of course, Sophie needs a lot of processing power so it must be simplified to some degree so that your machine (or device) can drive all the opponents. But how much do they simplify it?

One criticism that Project Cars 2 drew (before it was de-listed in September 2022) was that the AI cheated, especially in the wet.

Assetto Corsa Competizione (ACC) however, draws criticism for using so much processing power for its AI (and the physics in general) that even in mid-range PCs neither the CPU or GPU appear to be running at full capacity during offline races (suggesting that memory speed or something else in the hardware is bottlenecking the system).

So maybe we can categorise AI into multiple levels. One would be true AI who think for themselves. In this case Sophie, who is on another machine elsewhere in the world which makes her another online opponent (a spectacularly good one at that), which doesn't help us in offline play (not until that much processing power is available cheaply).

Then we have sort-of AI, which use the game's physics, but work within sensible preset parameters. This is how the moddable racing sims operate. The racing line is created for them when the track is created. In Assetto Corsa (AC) for instance, it's a recorded lap from a real person. The AI's track limits are also set by the track creator (which the AI will try to stay inside of). Hints can be given in AC with a specific hints file as well, usually to make them more cautious in trickier spots to avoid too many crashes. And then the end user can use a sliding scale to determine how strictly they hold to the racing line when they need to overtake (or are being overtaken).

IsiMotor2 has a similar development path, but for creators who bother there is a learning mode in the development stage where the AI can actually get better over time, using the racing line and track limits as a starting point. For the end user though, they're back to working within preset parameters, deviating from that line based on how aggressive you've made them with your settings.

In many games there's still the question of how much the changing conditions of track, tyres, fuel load and more affects them.

Below that we should probably call them advanced bots; opponents which can still deviate off the line to overtake or otherwise race side-by-side as needed. And then there's just bots; the mobile chicanes of old which don't seem to be affected by your presence at all.