Other case mentioned was with face recognition, imagine a coop trusts the AI that tells him that you are a dangerous criminal, you could get killed if you are unfortunate because the cop could assume the computer is right and you are very dangerous and never thinks about the false positives
I remember when car navigation systems were a new thing and there were frequent news reports of people who drove into rivers or lakes because the navigation system told them to cross the water, but not that the crossing happens by ferry (rather than via a bridge).