Tesla’s ‘Self-Driving’ System Is Likely Involved in Fatal Crashes More Often Than Human Drivers ⇥ prospect.org

Faiz Siddiqui and Jeremy B. Merrill, Washington Post:

The crash in North Carolina’s Halifax County, where a futuristic technology came barreling down a rural highway with devastating consequences, was one of 736 U.S. crashes since 2019 involving Teslas in Autopilot mode — far more than previously reported, according to a Washington Post analysis of National Highway Traffic Safety Administration data. The number of such crashes has surged over the past four years, the data shows, reflecting the hazards associated with increasingly widespread use of Tesla’s futuristic driver-assistance technology as well as the growing presence of the cars on the nation’s roadways.

The number of deaths and serious injuries associated with Autopilot also has grown significantly, the data shows. When authorities first released a partial accounting of accidents involving Autopilot in June 2022, they counted only three deaths definitively linked to the technology. The most recent data includes at least 17 fatal incidents, 11 of them since last May, and five serious injuries.

Ryan Cooper, the American Prospect:

[…] Back in April, he [Elon Musk] claimed that there have been 150 million miles driven with FSD on an investor call, a reasonable figure given that would be just 375 miles for each of the 400,000 cars with the technology. Assuming that all these crashes involved FSD — a plausible guess given that FSD has been dramatically expanded over the last year, and two-thirds of the crashes in the data have happened during that time — that implies a fatal accident rate of 11.3 deaths per 100 million miles traveled. The overall fatal accident rate for auto travel, according to NHTSA, was 1.35 deaths per 100 million miles traveled in 2022.

In other words, Tesla’s FSD system is likely on the order of ten times more dangerous at driving than humans.

I have been trying to reconcile the safety bonafides claimed by Tesla and autonomous car enthusiasts alike with the figures revealed here. Cooper takes into account only fatal collisions; perhaps Teslas are involved in fewer collisions of any time when “Full Self-Driving” or “Autopilot” are engaged. But that also appears to be untrue: the NHTSA says there were 181 collisions of any type per 100 million miles travelled in 2020 — the most recent data available in Table 24 here — and 207 in 2019, while the rate of this Tesla data set is 490 per 100 million miles. So it is about twice as likely to have a collision of any type and, as Cooper writes, perhaps ten times as likely for it to be fatal.

I do not see how these numbers square with claims about greater safety from autonomous vehicles — or, at least, Tesla’s vehicles. It seems plain to me that these systems are significantly more dangerous today than human drivers. And, I feel compelled to mention, this is effectively the danger rate of a single driver. Tesla’s software may have been improved, like how a human driver learns new things, but all of its cars run variations of the same software with the same instructions and similar hardware. This one driver seems to kill people at up to ten times the rate of the national average.

A caveat: the total “Full Self-Driving” figure reported by Musk in that investor call represents miles driven in Canada, but the NHTSA collision stats are U.S.-only. But I do not expect that to create a significant difference in the math. Less than 11,000 Tesla Model 3s were sold in Canada in 2022, compared to 156,000 in the U.S., and the FSD beta was not made available in Canada until April 2022. Also, as Cooper points out, this is an early data set; some of these crashes may not be found to be the fault of autonomous software, while other crashes involving its use may not have been reported.