I've never understood techno-optimists on this point.

If Uber has such amazing technology that avoids accidents, they could deploy them as accident-avoidance for regular drivers. They could give normal driving control to the humans, and take over only if there is a potential of accident.

If you argue that these driver assistance technologies already exist, then you better not be comparing your magical AI on supersafe cars with sunny climate on well marked highways to the average driver on the average vehicle under average climate on average roads.

That is a good point, especially when considering that modern cars have semi-autonomous controls which are likely to actually be effective in preventing ‘human’ failures, while still maintaining the UX of a non-autonomous vehicle.

Another point to consider, is that the actual experience of autonomous driving may increase the likelihood for ‘manual intervention’ failures due to the system being able to react correctly most of the time!

Sure, but do they? Can we really know with such different sample sizes?

Regarding sample sizes.

Say, you claim your product works, on the average, for 80 years before breaking.

I believe you, buy it, and it promptly breaks within 2 years of operation.

Maybe I was unlucky. But high more likely, you were bullshitting me.

That's what we have here.

Human death rate: 1 death / 80 million miles

Uber's death rate: 1 death after 2 million miles

So yes, we can make judgement calls.

All evidence points to nope.

Citation needed. This is my argument. Unless you are working for Uber/Tesla/Waymo where is the data? I'd even be satisfied with seeing incidents where the system successfully avoided collisions, at least as a baseline. How do regression lines trend in comparison to non-autonomous cars considering scale?

>Citation needed.

Sure!

Traffic fatalities in the US, human drivers: 1.25 deaths per 100 million vehicle miles [1]https://en.wikipedia.org/wiki/Transportation_safety_in_the_U...

Uber's self-driving car fatality rate: 1 death per 2 million miles[2]

That is, Uber's state-of-the-art is an order of magnitude worse than human drivers.

I know, you'll complain about small sample size. The thing is, assuming Uber's cars match humans, the chances of a collision happening only after 2 million miles are very small, like pulling an ace of spades out of a shuffled deck of cards.

The point is, we do have the data, and the data does not support optimistic beliefs.

[1]https://en.wikipedia.org/wiki/Transportation_safety_in_the_U...

[2]https://arstechnica.com/cars/2018/10/waymo-has-driven-10-mil...

I'm pretty sure if they were better than humans by any real statistical margin they would be trumpeting that in press releases. The fact that they work so hard to maintain such secrecy- only revealing data that they are legally mandated to- suggests that they, the people must familiar with the system performance, think it is inadequate.

In essence, because the people who control the data have an interest in revealing data that is superior to human performance, the lack of conclusive data on this leads to an inference that it isn't.