When we need something that works 99.99999% of the time, it is written in ADA or similar safe language and tested exhaustively. You know how long processing will take to the microsecond, you use a special realtime OS, etc.
AI will have its own failings, some of them totally unpredictable, so yes, boiling it down to 'just needs to be trained properly' is terrifying when peoples lives are at stake
Look at this attack, and tell me how is this goonf to be avoided with 'proper training' https://www.theregister.com/2018/01/11/ai_adversarial_attack...
Same can be said about humans.
None of what you listed matters in the real world. Self-driving systems just have to be better than average humans, on average, to be useful.
> boiling it down to 'just needs to be trained properly' is terrifying when peoples lives are at stake
You have no problem with humans being trained to drive. 38.6 thousand people died in car crashes in the US alone last year.
So just like all miracle solutions, .i.e. communism, every time we tried it, we just weren't doing it right.