Every piece of technology has a variety of failure modes, some more likely than others. FSD is not likely to take aim at a pedestrian and floor it, just like your brakes aren't likely to explode, and neither of you are irresponsible for making those assumptions

No single human fully understands these AI models, and they can change daily. Yet Tesla is putting them in control of multi-ton vehicles and leaving all liability on humans with worse reaction time and little to no understanding of how it works or tends to go wrong.

I think you are making scenarios that no reasonable person would assume. There is a difference between 'getting confused at an intersection and turning wrong' and 'actively trying to kill the occupant by accelerating at max speed while turning the steering'.

FSD is a blackbox. Even Tesla appears to be unable to prevent regressions. One of the most sophisticated companies in the world can't prevent regression in a safety critical software they frequently update. Let that sink in.