I am working in AI now and I could not agree more with Culzean. What AI is starting to do now is learn from humans by observing what we do in every situation. The problem is, with the variability of driving these days, the amount of observation and trial by error tp be done will take many years. The only way it could work, even given today's latest t technology, would be for us all to have it fitted and switched off, but in observation mode, recording real world scenarios and human decisions, for many years, before enough data exists for the AI to take over.
Only if every single car had AI driving fitted and they all followed absolutely identical rules and behaviours (none of this Volvo AI vs BMW AI nonsense), and they all communicated in real time with each other about where they are, what speed, direction, acceleration, vehicle condition, weather, what they have seen, all of that - then everything would be wonderful and safe, and so incredibly dull.
But imagine a Volvo AI car approaching an BMW AI car and then something unexpected happens. How can one type of AI anticipate the other types behaviour? Or if a collision is inevitable? Does the car carrying a pair of pensioners head for a tree to save the lives or a car carrying a young family?