People don’t like what they can’t understand. Most of the accidents caused by autonomous vehicles are weird and probably would have been avoided by human drivers. However, this isn’t the whole picture, because those same vehicles also avoided a bunch of accidents that humans would have caused.
My interpretation of the data is that autonomous vehicles are already statistically safer than human drivers.
Problem is, humans are emotional beings who aren’t always capable of pure rationale. If you tell someone they’re statistically safer in a Tesla, but it might accelerate into a wall at 60mph, many people won’t take that risk.
There hasn't been a ton of data especially in areas where the risk is important.
You’re joking right?
The human driver didn’t handle the “unexpected information” of a human being on the road when the light was green.
The human driver didn’t handle the “unexpected information” of running over a human being and drove off instead.
Humans are awful at unexpected information. This is exactly the moment that emotions take over and emotions are what makes humans really bad at tasks like driving.
The only way we have gotten any semblance of safety is "slam on breaks if unrecognized" which happens quite frequently.
At least half the people out there are terrible at this. We have expression like "deer in the headlights" to describe how they behave.
A jay walker in Arizona got ran over because "people are only in crosswalks" (I don't remember the company offhand)
Autonomous driving isn't automatically safer. Companies keep focusing on the potential for improved safety without technically claiming they are that much safer.
The potential is great but that means nothing. What matters is what we make.
https://westseattleblog.com/2023/03/crash-blocks-westbound-w...
Friend of mine just witnessed a crash the other day where a driver at a high rate of speed lost it on a wet corner and the passenger got ejected.
Americans drove 3.2 million million miles (aka trillion).
Comparing things that have managed in total barely 1/100,000th of a years usage means anecdotes don't matter.
Because there is at minimum one unavoidable aspect here: we don't know what accidents an AI could prevent and way way more importantly we don't have enough data on what accidents an AI would cause at highway speeds.
Insufficient data is available for how effective AI is at preventing accidents.
The Arizona one I gave for instance, the company was found not liable since a person might have done the same but could they have slammed on the breaks turning a fatality into an injury? We will never know and can never no.
The best we can do is measure real world performance and pay close attention.
As terrible as humans are we do have a very important trait. Being able to handle unexpected information efficiently.
All the current AI models do have ways of slightly generalizing but only slightly. You still need to effectively train "when there are cones you need to switch lanes" and "when there are barriers you can drive on the line" kinds of things.
And while yes everyone mostly talks about low stakes examples that is mostly because the high stakes ones haven't been handled by AI yet.
E.g. a Tesla driving under a semi because it thought it was a billboard.
I look forward to any replacement to human drivers (although eliminating cars from Metro areas would be much more efficient I digress..) but autonomous vehicles need to be better before they can be that.