We also have humans who vary widely in skills who are writing software. The same people who allow acceleration and braking to be linked to the entertainment system, and by extension, controllable over the internet. You seem to be assuming the worst possible case for autonomous car software will necessarily be better than the best human driver, but that's not an assumption backed up by reality. More likely (to me) is that, while both humans and autonomous cars might be safe in most situations, the fail states for autonomous cars are likely to be less predictable.
