Since Tesla’s Autopilot was introduced back in October 2014, Tesla CEO Elon Musk has made extravagant claims about its superior safety. Last month, reacting to what Musk considers sensationalized media coverage of Tesla accidents, the company released the first of its promised quarterly safety updates.
The report claimed that, compared to the U.S. automobile crash rate of one every 492,000 miles, Teslas in the second quarter of 2018 had an accident “or crash-like event” every 1.92 million miles. With Autopilot engaged that dropped to one in every 3.34 million miles.
In other words, human-piloted Teslas had accidents at about one quarter the U.S. average, by miles traveled. The rate for Autopilot Teslas was about one-seventh of the U.S. average. (The company did not say how it arrived at the benchmark figure for the U.S. average number.)
According to the company numbers, cars on Autopilot were nearly twice as safe as human-pilot Teslas, buttressing Musk’s past claims.
Perhaps that’s because, as it turns out, Teslas on Autopilot could have a higher fatal accident rate than those driven entirely by humans.
Back in November, 2016 I analyzed fatal-crash data for the Model S, comparing cars driven by humans to those driven on Autopilot.
Our conclusion then: Autopilot had a fatal crash rate approximately double that of human-pilot Teslas—a conclusion radically different from the rosy Autopilot safety picture that Elon Musk and Tesla had so relentlessly promoted.
Tesla Model S Autopilot testing with IIHS [CREDIT: IIHS]
But exposure data and sample size for that analysis were small (as it was for Tesla’s upbeat claims), casting serious doubt on any statistical analysis of fatal-crash data for Autopilot-equipped Teslas.
Now, nearly two years later, a tweet by Musk suggests that the total number of miles driven by Tesla Autopilot has increased fivefold. As a result, comparative statistics are now somewhat more meaningful—though still far from definitive.
The good news: Autopilot’s fatal-crash rate has improved significantly in the last two years.
The bad news: It’s still may not be as good as human drivers.
To figure a fatal-crash rate, we need two numbers: fatal crashes and vehicle-miles driven.
Neither of these numbers is easy to come by.
Tesla, of course, does not make public its accident statistics. For my 2016 analysis, I got my crash number from bubslug, a contributor to the stock-analysis website Seeking Alpha who had compiled a list of fatal Tesla accidents in the U.S.
To figure accident rates, I also needed a measure of the cars’ accident exposure: total vehicle-miles driven.
Back in 2016, Tesla thoughtfully supplied those numbers itself. On October 7, Elon Musk tweeted “Cumulative Tesla Autopilot miles now at 222 million miles.” Earlier that year, Tesla had announced that total Tesla miles driven had passed the two-billion mark, and published a graph projecting future total mileage. The graph projected 3 billion miles by September 2016, so I used that figure.
Combining bubslug’s numbers and Tesla’s numbers, I calculated the following fatal U.S. Tesla accident rates as of October 2016:
- Autopilot: 0.45 fatal accidents per 100 million miles
- Human-driver: 0.21 per 100 million miles
(Put another way, Teslas on Autopilot had one fatality every 222 million miles; human-driven Teslas one every 470 million miles.)
Score one for the humans.
That was then. What about now?
We haven’t seen anything from bubslug for a while. And Tesla has recently held miles-driven data close to the vest. No tweets from Elon, and the total-miles driven graph seems to have disappeared from the website.
So, how do we find recent numbers to crunch?
Tesla Model 3 bought [Photo by reader AH]
Bubslug’s role as unofficial Tesla fatal accident compiler now seems to have been assumed by tweetster @ElonBachman, who has compiled a list of worldwide fatal accidents involving Teslas.
According to an @ElonBachman tweet, as of May 2018, there have been a total of three fatal Autopilot accidents and 13 human-pilot Tesla fatals, worldwide.
The three Autopilot fatal crashes are well known:
- May 7, 2016, Florida: Model S on Autopilot into semi-truck
- Jan 20, 2016, China: Model S on Autopilot into street sweeper
- March, 31 2018, Mountain View, California: Model X on Autopilot into highway divider
For the human-pilot Teslas’ crash data, we confirmed each accident on Bachman’s list through other independent sources. With a couple of corrections, we arrived at the following list of human-pilot fatal Tesla accidents:
- July 10, 2014, Hollywood, California: Stolen car
- Dec. 31, 2014, Sonoma, California: Drove off cliff on Highway 1
- Jan. 22, 2015, Los Gatos, California: Single-car crash
- June 22, 2015, Malibu, California: Drove off cliff
- Dec. 22, 2015, Canada: T-boned by dump truck
- Dec. 28, 2015, Houston, Texas: Unintended acceleration into pool
- Sept. 7, 2016, Netherlands: Drove into ravine
- Nov. 3, 2016, Indianapolis, Indiana: High-speed impact, fire
- Jan. 17, 2017, Germany: Autobahn pile-up
- Mar. 12, 2018, Norway: Unintended acceleration into parking garage wall
- May 1, 2018, Liberty, Ohio: Prom night speeding
- May 10, 2018, Ft. Lauderdale, Florida: Teen-age speeding, fire
- May 2018, Switzerland: High-speed impact, fire
That’s 13 accidents. (If any reader can confirm additional accidents, please let us know.)
Unfortunately, without Tesla’s total-miles data, we still had no way to figure an accident rate, which is the number that really matters.
MIT to the Rescue
Thanks to MIT researcher Lex Fridman, co-author of the MIT Autonomous Vehicle Technology Study, however, we now have highly educated guesses for both Autopilot and human-pilot Tesla miles driven.
Tesla Model X crash, Hwy 101 Mountainview, California
Fridman’s numbers, as of June 20: 1.21 billion Autopilot miles; 6.47 billion human-driven miles.
Combining our crash data with the MIT numbers, we arrive at the following fatal Tesla accident rates:
- Autopilot: 0.25 per 100 million miles
- Human Pilot: 0.20 per 100 million miles.
(Put another way, one Autopilot fatal crash every 400 million miles, one human pilot crash every 500 million miles.)
Either way you look at it, this is a big improvement for Autopilot. Contingent on our disparate data sets, the fatality rate is down by nearly half since 2016—although it still might not be quite as good as the human-pilot rate.
Is this big drop in the fatal crash rate due to improvements in Autopilot itself, better education of drivers about how to use it, or simply a statistical quirk?
No one can answer that question. But I’d put my money on the statistical quirk. When the sample size is small—only three accidents—random chance plays a huge role in any statistical result.
Consider the odd pattern of recent human-pilot Tesla fatalities. Of the 13 total since the car was introduced in 2012, four occurred in a ten-week period this past spring. Thus, on March 11, 2018, human-pilot Teslas had a fatality rate of just 0.14 per 100 million miles barely half the Autopilot rate. Ten weeks later, it rose to 0.20 per 100 million miles.
Did the Model S and X suddenly develop major safety defects? Did their human pilots abruptly begin a binge of reckless driving?
Of course not. It’s just the way stuff happens when the statistical sample size is small.
Tesla Autopilot sensor system
While these updated numbers for Autopilot are encouraging, it’s clear that Tesla’s claims of its vastly superior safety—at least in terms of fatal accidents—are still vapor. It’s way too soon to come to any firm conclusions about Autopilot safety.
Musk has compared Autopilot’s fatality rate to government figures for the overall U.S. traffic fatality rate. Surely someone as smart as Musk realizes that comparing Tesla’s Autopilot numbers to the NHTSA figure is not just apples-to-oranges. It’s apples-to-aardvarks.
NHTSA’s Fatality Rate per 100 Million Vehicle Miles Traveled, the figure repeatedly cited by Musk as the Autopilot benchmark for comparison, includes bicycles, pedestrians, motorcycles, and buses. Musk is essentially equating a Tesla Autopilot crash to a pedestrian getting run over by a bus.
Using this bogus comparison, Musk has claimed that “approximately half a million people would be saved (every year worldwide) if the Tesla Autopilot was universally available."
To be blunt, that statement can charitably be described as utter poppycock.
But compared to the more reasonable benchmark of human-pilot Teslas, after more than a billion miles of driving, it’s starting to look like the fatality rate of Teslas on Autopilot may at least be in the same ballpark as their human-piloted counterparts.
We certainly hope this positive trend continues, and we can’t wait to crunch the numbers in 2020.