Sorry Elon Musk, There's No Clear Evidence Autopilot Saves Lives (arstechnica.com)

Timothy B. Lee writes for Ars Technica: A few days after the Mountain View crash, Tesla published a blog post acknowledging that Autopilot was active at the time of the crash. But the company argued that the technology improved safety overall, pointing to a 2017 report by the National Highway Traffic Safety Administration (NHTSA). "Over a year ago, our first iteration of Autopilot was found by the U.S. government to reduce crash rates by as much as 40 percent," the company wrote. It was the second time Tesla had cited that study in the context of the Mountain View crash -- another blog post three days earlier had made the same point. Unfortunately, there are some big problems with that finding. Indeed, the flaws are so significant that NHTSA put out a remarkable statement this week distancing itself from its own finding.

"NHTSA's safety defect investigation of MY2014-2016 Tesla Model S and Model X did not assess the effectiveness of this technology," the agency said in an email to Ars on Wednesday afternoon. "NHTSA performed this cursory comparison of the rates before and after installation of the feature to determine whether models equipped with Autosteer were associated with higher crash rates, which could have indicated that further investigation was necessary." Tesla has also claimed that its cars have a crash rate 3.7 times lower than average, but as we'll see there's little reason to think that has anything to do with Autopilot. This week, we've talked to several automotive safety experts, and none has been able to point us to clear evidence that Autopilot's semi-autonomous features improve safety. And that's why news sites like ours haven't written stories "about how autonomous cars are really safe." Maybe that will prove true in the future, but right now the data just isn't there. Musk has promised to publish regular safety reports in the future -- perhaps those will give us the data needed to establish whether Autopilot actually improves safety.
UPDATE (2/16/19): The study's underlying data reveals serious flaws in the methodology that undermine its credibility, according to new analysis from a research and consulting firm.

Comments

Errors

[LetterRip]

While Timothy normally does excellent articles, his reasoning and logic were severely flawed this time.

First off the NHTSA report focused on autosteer, not autobraking. Hence his attributing the reduction in accidents to autobraking is bizarre. What the NHTSA was disavowing was that they had not examined the entirety of Autopilot (which includes autosteer, autobraking, lane keeping, etc.). Timothy mistakenly thinks they were stating that they hadn't verified the effectiveness of autosteer installation in accident reduction (they didn't verify the actual usage, but drivers with autosteer installed use it about 50% of their driving time).

Secondly the Tesla's prior to the FSD update already had autobraking, so the 40% reduction in accidents after enabling FSD can't be attributed to the autobraking.

Re:Errors

[Mr D from 63]

Anybody who read the NHTSA report should clearly understand that the Autopilot safety data comparison was not done to demonstrate the safety of Autopilot, but rather to decide if there was indication that AP caused an increase. Also, 2/3 of the cars in the study didn't have any pre-AP data at all. It was entirely useless for the purpose of making any kind of safety claim. The NHTSA should not have had to clarify, but too many idiots made stupid claims based on that information. The media in general can be really stupid with statistics.

How can it not be safer?

[Xenx]

Unless the autopilot feature is actively instigating accidents, it's impossible for it not to be safer. Anything above and beyond relying solely on driver's response is an improvement, even if only minimally.

Re:How can it not be safer?

[Balial]

A Tesla will happily drive you into a stopped fire truck, or a turning semi trailer, or a freeway divider while you're driving at full speed. So yes, it's actively instigating accidents that humans are pretty good at avoiding. See google.

I'm a huge fan of Tesla, but their autopilot scheme is a farce.

Re:How can it not be safer?

[olsmeister]

happily drive

What kind of messed up AI have they developed over there?

Re:How can it not be safer?

[Xenx]

Autopilot requires the driver to be attentive. The driver should be stopping the vehicle in those cases. The car didn't actively seek out hitting the fire truck, it just didn't detect it and stop. Absolute worst case, autopilot never detects anything ever and we're no worse off than just having a driver behind the wheel. As soon as the autopilot detects and prevents any potential accident, we're now safer with than without.

Re:How can it not be safer?

[cascadingstylesheet]

and we're no worse off than just having a driver behind the wheel.

That's demonstrably untrue. Humans are terrible at partial attention. Partial automation (like Tesla's misnamed Autopilot) lulls humans into a state of inattentiveness.

Re:How can it not be safer?

[Xenx]

It's designed to operate in conjunction with the driver, at this point in time. The old adage, the whole greater than the sum of its parts, applies. Between you and autopilot, less mistakes are bound to happen.

Re:How can it not be safer?

[Waccoon]

#1 reason why I own a manual transmission and never use cruise control. I don't trust myself. It may be decidedly un-American to have a compact car with few gadgets and gizmos, but I always know how fast I'm going in my car with a manual transmission. When I drive automatics, I speed, and when I use cruise control, I zone out.

Emergency response features sound useful, but I don't want a car with half-assed automation. If they car is supposed to drive itself, either do it completely and properly, or not at all.

Re:How can it not be safer?

[grumbel]

The problem with a lot of safety measures is they make people behave less safe. Thus the benefit of the safety measure gets eaten up by people getting into more accidents. In the case of autopilot you don't just augment an attentive driver with additional features, you turn him into an inattentive driver when you give him autopilot abilities. See the recent Uber self-driving death, driver wasn't paying attention to the road and fumbling around with the phone, we can blame the driver, but that's ignoring the fact that that's just how humans behave when they get bored. So the technology doesn't just need to catch a few dangerous situations, it has to be able to fully replace an attentive driver. Furthermore people like to overestimate what the capabilities of the technology, a Tesla car is not a fully self driving vehicle, but many people treat it like one and go hands-free, despite the technology not being able to handle that.

All that said, the biggest issue in judging self-driving cars is the lack of data points. Not just in actual crashes, but also in what the self driving technology is actually capable off. We have a few videos of the technology behaving correctly, but hardly anything that explores the limits of their abilities. I'd much prefer it if those car companies would just stuff their self-driving software into a simulation and have the public have some fun with trying to exploit it, than slowly collecting data points by the means of dead people on the road. So far there also doesn't even seem to be an official driving test for self-driving cars, we just trust the manufacturer's to do the testing themselves.

Re:How can it not be safer?

[goose-incarnated]

Unless the autopilot feature is actively instigating accidents, it's impossible for it not to be safer.

The autopilot is actively encouraging inattentiveness, so yes, it is possible for the addition of autopilot to increase the accident rate.

Re:How can it not be safer?

[tomtomtom]

In some ways, driving with Autopilot (and it will get worse as these systems become more advanced) is like supervising a learner driver. Most of the time (in the limited scenarios it can be fully engaged) it will act like a normal fully-qualified driver but every so often it will fail to react properly or just do something really erratic. If you're the human in charge of the vehicle in those situations, then the role you are fulfilling is more akin to that of a driving instructor than to that of a regular driver. Except actually its worse because you can't communicate with the "actual driver" like a human would and you may well be driving down roads you've never seen in your life before so don't know where the "danger spots" might be which people tend to get wrong.

There are very good reasons why most countries require full time driving instructors to pass additional tests and re-certify relatively frequently. I think it's worth considering whether we ought to apply the same standards to some types of semi-autonomous driving tecnnology.

Flaw

[Anonymous Coward]

The system is flawed because it has to rely on a lazy, distracted driver who really doesnâ(TM)t want to drive his own car to begin with.

3.7 times lower

[gumpish]

3.7 times lower

What the hell are they trying to convey here?

If value A is "2 times less" than value B, does that mean A is 1/2 B? So "3.7 times lower" means a factor of 1/3.7 ?

While I'm on the subject, there is no unit for "coldness" or "slowness", so please stop saying nonsense like "twice as cold" or "ten times slower" and stick to "half the temperature" and "one-tenth the speed". FFS

Sorry?

[lucm]

"Sorry Elon Musk"? That's the most asswipy headline I've read in a long time. Was this mistakenly submitted to Slashdot instead of Jezebel or Salon?

Sorry Elon Musk?

[Peter P Peters]

Sorry for our really fucking lame attempt at writing clickbait headlines..
I refuse to participate in this discussion because of this shit. What's wrong with "There's No Clear Evidence That Autopilot Saves Lives"? Seriously, I want an answer.