[CydeWeys]

Imagine a situation in which Uber cars are better than human drivers, but not by a lot. Then imagine some hastily thrown-together patch being pushed fleet-wide that ends up causing crashes in some unanticipated situations. You could imagine quite a lot of carnage happening out there until they're all yanked off the roads and rolled back.

Statistics doesn't tell the whole story because it's backwards-looking, not future-looking, and the Uber crashes aren't going to be independent events.

[notahacker]

Statistics of accidents with human drivers also include a lot of humans who are banned from the road and sometimes even imprisoned for not driving safely enough; the average human could and should be driving a lot better than those statistics. And in many jurisdictions, there's an expectation that a commercial driving service is much better than the average person permitted to use a car. If the service aims to reduce transport costs to the extent that vastly more journeys are taken before - which Uber certainly does aspire to - then its autonomous vehicles can significantly increase the numbers dying on the roads even if they actually do have a persistent and statistically significant per-mile safety advantage over human-driven cars.

And if I'm a regulator tasked with reducing road deaths, the question I'd be asking wouldn't be "is this car's autonomous mode marginally better than the average driver so Uber can cease paying drivers and hit profitability?", it's "is the human behind the wheel of a car operating in [semi]-autonomous mode no longer preventing a statistically significant number of crashes?" (or "is telemetry suggesting their interventions are actually responsible for more accidents than they prevent?"). Just because a car is a lot safer with a computer behind the wheel than a car with the average person and no computer behind the wheel doesn't mean that permitting autonomous vehicle operators to rely on technology alone is a safety improvement.