Re: The Nasty Little Truth About Deep Learning
The thing is that humans are so bad at driving, an imperfect AI can still have a better average safety record than humans.
But it's going to fail in *different* ways, there will be scenarios where an average human driver would have been fine and the computer kills everyone, and there will be scenarios where the human would have killed people and the computer has no problem.
Of course, "computer controlled car goes for drive, no-one dies" is not (any longer) a news story, so much of the press will be of the form "computer screws up when human would have been fine".