Even if the cumulative amounts of fatal accidents is reduced, certain types of accidents happen less, if different types of accidents happen more often, say accelerating into a wall uncontrollably, that is not optimal. I worry that we will accept 'slightly better' as good enough, along with the consequences that come with it.

At some point, economics and actuaries come into play. It will be worrying when 'AI is safer than humans!' can be pointed to as a reason to not correct expensive, perhaps intractable, AI failures.