I see people distracted and not paying attention every time I drive, and they don’t have anything in the car helping them. The only reason they don’t hit me is because I moved out of the way when they come in my lane.
If these systems actually did the hard part of driving, avoid crashes with super-human reaction times - then they would be extremely valuable. But that's not the case and they are essentially worthless. Might as well hire a driver for you of you feel driving isn't for you. Save you and the whole planet the trouble until these systems actually do the stuff that matters.
This is a wrong way to look at it. recently a Tesla just ran over a motorcyclist. These people would have been alive if it was a normal human driving. These are fault modes that did not exist before. Also, we have adapted to drive along with other human drivers. Bikers are taught to make sure they are seen on the road and make eye contact with drivers etc.. How do you make sure a Tesla's ML model recognized you? What does that recognition even mean?
That's the wrong way to look at it. 109 people were killed on the road today. How many would've been alive if Autopilot was active?
Assuming your goal was to reduce fatalities.
Tesla does not disclose the information needed to make a proper apples to apples comparison to the public or researchers. They do not disclose the number of miles or the number of accidents, let alone the conditions or circumstances that would be needed to account for bias or confounding variables and determine the corresponding comparable metric. Of the reports they make to NHTSA [1] which are public by default, they declare the vast majority of contained information as confidential business information and their lawyers force NHTSA redact it.
A company kills a bunch of people and nobody is allowed to know how safe it is because that is confidential. That alone should worry you.
But, that is not all. Tesla does not know either. If you look at the reports you will see that of the over 1,000 crashes, Tesla declined to investigate to determine if anybody was injured in ~95% of cases. Of the reported confirmed fatalities, Tesla telemetry missed the majority of them. I will say that again, they can not even detect the majority of fatalities while their system is active. Their data collection procedures are so grossly inadequate that you can not even determine the upper bound on the fatality rate, let alone more nuanced safety measures.
To quote the safety investigation mentioned in the article [2]: "Tesla largely receives data for crashes only with pyrotechnic deployment, which are a minority of police reported crashes. A review of NHTSA’s 2021 FARS and Crash Report Sampling System (CRSS) finds that only 18 percent of police-reported crashes include airbag deployments." Tesla bases their safety statistics on data collection conditions that under-report actual crash rates by a factor of 500% on average.
The statistics Tesla releases do not even attempt to account for the glaring discrepancies and under-reporting that is clearly visible in their fatality reports and data collection conditions and instead intentionally state the under-reporting that benefits them as simple fact. They are intentionally overstating product safety based on data that any competent professional knows is faulty and inadequate to push product. That level of misconduct is criminal.
[1] https://static.nhtsa.gov/odi/ffdd/sgo-2021-01/SGO-2021-01_In...
[2] https://static.nhtsa.gov/odi/inv/2022/INCR-EA22002-14496.pdf
We also want to know whether there are correlations between the accidents that happen that are outside of factors we would consider normal among humans.
A contrived the example to prove the point: We wouldn't be satisfied with an automatic driver that got into only half the accidents humans did... but also possessed an inexplicable tendency to charge at people wearing certain color-combinations of clothes.