if we ask what the reasonable expectation is after an advertisement tells a customer that the capability "exceeds human safety" I would say that the average customer thinks of a fully automated system.

This couldn't be further from the truth as automated vehicles still suffer from edge cases (remember the lethal accident involving a concrete pillar) where humans can easily make better decisions.

A system that is advertised as superior to human judgement ought to strictly improve on human performance. Nobody expects that to mean that the car drives perfect 95% of the time but accidentally kills you in a weird situation. This 'idiot savant' characteristic of ML algorithms is what makes them still dangerous in day-to-day situations.