Youtube channel veritasium, on a video called "The real moral dilemma of self-driving cars" (ref) claims that if all cars in 2017 were self-driving, hundreds of lives could be saved.

So he is claming that the self-driving technology of 2017 would be less lethal (regarding number of deaths in car accidents versus miles driven) than human drivers. That a self-driving car has a higher mileage per death than human driven cars.

Is it so?

  • 1
    Well it would certainly lower driving accidents in my state since we couldn't use our self driving cars for all of winter. – JonK Mar 27 '17 at 15:05
  • It doesn't relate to completely self-driving cars but I can tell you that having a driving assist mechanism on my car (I-Sight) lowered my insurance rates... so if the insurance companies believe that autonomous features are beneficial, it must be the case. – Catija Mar 27 '17 at 21:38
  • Given the skill of Florida drivers, where making turns across multiple lanes of traffic is common, this is not hard to believe at all. – Pete B. Mar 28 '17 at 14:08
  • Technically lethality is the ability to cause harm, and the measure of it measures the amount or intensity of harm caused. A car driven by a computer delivers as much lethality as a car driven by a human. I don't believe 'lethal' is the appropriate word to use here. – TylerH Mar 28 '17 at 14:59
  • One of the significant differences between a world with all self-driving cars and a world with some self-driving cars is that in a world with all self-driving cars, the cars can assume full cooperation and operate in a fully coordinated manner, whereas today, self-driving cars move in an uncooperative environment. I think that this is such a fundamental shift that looking at statistics of self-driving cars performing in an uncooperative environment is not going to be good enough to answer the question. However, as it seems, even in an uncooperative environment, self-driving cars are at least … – Jörg W Mittag Apr 5 '17 at 23:21


A 2015 study pitting the ~50 autonomous cars in California against the nationwide statistics for conventional cars found the accident rate of autonomous cars to be about five times that of the national average, with four times the injury rate.

However, the study also notes the very small statistical base -- there were just 11 accidents on file for autonomous cars in 1.2 million miles. It also notes that the accidents were mostly at slow speeds with minor injuries.

Google claims that in none of the mentioned cases, the autonomously driving car was at fault.

A more recent article mentions other numbers. A federal inquiry to a deadly crash involving Tesla's autopilot found that Teslas with autopilot are crashing 40% less frequently than Teslas without autopilot.

All this is somewhat inconclusive as to whether today's self-driving cars are already safer than human-controlled vehicles. But that is not actually the claim here, is it?

Claim vs. Interpretation

You write that you interpret the claim to mean "that a self-driving car has a higher mileage per death than human driven cars."

But that is not the claim, is it? The claim is that hundreds of lives could be saved if all cars were self-driving. That is a hypothesis, really, and does not have to be backed by any current self-driving-cars-in-human-controlled-traffic numbers. It is about what would happen if the human factor were eliminated (as far as other cars are concerned, of course you still have pedestrians and cyclists etc.)

The wired.com article linked in the first section of this answer mentions that 40% of fatal car crashes in the US are the result of drunk driving, and 16% the result of distracted drivers. It also notes that as many as six million US drivers admitted to having hit another car on purpose. How many of these accidents would be avoided if all cars were self-driving?

  • 2
    The researchers, Brandon Schoettle and Michael Sivak, note the limitations, acknowledging their confidence levels could invalidate the overall finding of higher accident rates for self-driving cars. That's a different way of saying that the result of their study is in fact inconclusive. – gerrit Mar 27 '17 at 15:45
  • @gerrit: Which doesn' t change the second half of the answer, which I actually consider more important to the claim: data today is inconclusive, but the claim is plausible given that data. – DevSolar Mar 27 '17 at 16:49
  • The second half of the answer is speculative. I agree with your speculation, but it's not very well backed up by evidence. As for the first half of the answer, the question specifically mentions lethal accidents, for which we have even less data than for minor accidents (IIRC none whatsoever). – gerrit Mar 27 '17 at 17:05
  • 5
    @gerrit: The second half of my answer is pointing out that the claim is speculative. It is not about self-driven cars in today's traffic. It's about fully autonomous traffic, and AFAIK we have zero data on lethal auto-on-auto accidents. (You can do funny things when n=0...) – DevSolar Mar 27 '17 at 17:07
  • I also understand that most of those accidents were human-drivers rear-ending the autonomous cars, making them all legally not the autonomous fault. If I had to guess, I would say that people generally follow too closely, and auto-cars tend to stop or slow quickly when there is no need to, leading to "unpredictable driving" by human standards, and therefore more accidents. You are right to note that any perceived or real benefits from auto-drivers is with the assumption that nearly all vehicles are autonomous. – fredsbend Mar 27 '17 at 17:36

I agree with the assertion that we currently have insufficient evidence.

There has been one fatality associated with a self-driving car. Tesla said:

This is the first known fatality in just over 130 million miles where Autopilot was activated. Among all vehicles in the US, there is a fatality every 94 million miles. Worldwide, there is a fatality approximately every 60 million miles.

We're stuck relying on Tesla for the 130 million miles number. They're the only ones who aggregate that data. For the other US cars, we can consult Wikipedia, which refers us to NHTSA (National Highway Transportation Safety Administration). In 2014, there were 32,675 fatalities in 3,026 billion miles traveled. That's 1.08 fatalities per 100 million miles (from source). Calculating as million miles per fatality, we get about 93 million miles per fatality for 2014.

In the next software update after the accident, Tesla said that they made changes including:

Car will not allow reengagement of Autosteer until parked if user ignores repeated warnings

The way that the Tesla system is supposed to work is that the driver is supposed to be monitoring the situation and take control when necessary. The system can notify the driver that it is confused and ask the driver to take control. Prior to this update, it was apparently possible to refuse to take control. Maybe the system is even safer now. Or maybe not.

You can also read the NHTSA report (PDF) of the investigation into this accident for more information.

Not enough information doesn't mean that it is untrue. It means that we should try to get more information if we want to be able to answer that question. The little bit of information that we currently have is consistent with that claim.

  • The accidents per mile statistic is unreliable, because it compares highway traffic (Tesla) with assorted traffic (everyone else). – Peter Mar 27 '17 at 23:26
  • 1
    Going with the assumption that this answer is based on, namely that aggregate accidents per mile is a meaningful statistic here, this only shows that Teslas with Autopilot activated are approximately as safe as the average US driver in terms of accidents with fatalities. The difference between one accident per 130 million miles and one per 60 million miles is significant enough to be notable even if perhaps not conclusive, but IMO, the difference between 1/130Mmi and one per 94 million miles is not. I'm not sure you can extrapolate from that to that hundreds of lives could be saved. – a CVn Mar 31 '17 at 14:35

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .