I've just done some testing for F3 AI calibration and then tried them in a race weekend.
The Spark F2 with the wing sort of improved. It was able to take the corner really gently. Quite slow. In testing they were stopping well before the corner and crawling around. In the race it was just slow and wide missing of the apex.
The other Spark seemed to just manage to keep away from the hay, but was still on the dirt. About as fast as the winged spark.
The Eve F2 didn't seem to learn anything. In the test days, the first time it just got worse and worse, going along the fence as far as the ambulance (that's why it's there, from the rF1 days and the CSGT 917 AI..). The second time i think they were just hitting the hay... I don't think any of there were failing to at least turn as far as the hay after the learning though. None seemed to hit the tree or go left of the tree and around the back, so maybe there was some learning.
My premises/reasoning/excuse for not bothering with separate AI paths for them:
Most cars work well enough.
The Spark and Eve F2s don't work out of the box, but some calibration might help them.
Neither the Spark or Eve F1s will make the corner either (calibration might help), though the BT20 does work ok.
It could take me many hours or a day or two before i decide/realise that a hacked AI path does or doesn't help.
If a hacked AI path does help, then at some point if there is a code-change to the AI or cars that fixes the problem, the cars that I've hacked an AI for will then be slow and broken.
So I'm going to right ahead and assume that tweaking the cars or the AI code is the best and most correct option, and neither of those are things that I have to do

Therefore I can avoid tearing my hair out over things i don't understand or have required access to, and get on with... you know... making the track.