At least plane crashes are usually mechanical and not software, or it's straightforward software (hello MCAS), investigating a machine learning's system behavior ("what input lead it produce this output?") sounds like a more expensive problem.

Until you look you don't know if i was hardware, software, road conditions or user, unless you look its just a a pile of twisted metal.

Not currently. If we get to mixed 50/50 traffic...possibly. (Saying "but won't happen, because magic future technology" would be literal handwaving)

If we required full investigations now, while the numbers are small enough for it to be feasible, then perhaps by the time we get to 50/50 traffic, the technology will be that much better.