How Manned Vehicles Are Much Safer Than Elon Musk and Others Admit — And Why That Matters

The prime defense for self-driving cars has been,

A: We have one traffic death for every 86 million passenger vehicle miles driven. (or 1.16 deaths per 100m miles; source; Elon Musk)
B: Therefore, we can save lives if we create a self-driving car that does better

I will show why Part A is technically true but misleading, and consequently Part B is even more problematic.

An AV would have to be nine times safer than the current average manned mile, or have roughly one death every 770 million miles, just to be as safe as the manned cars it would replace, assuming the fatal crash rate doesn’t improve.

The Fallacy: All Cars Are Equally Relevant

The central problem with the 86m figure is that it refers to the average vehicle on the road. Some drivers are drunk or texting. The average car is 11 years old, when newer cars are linked to deaths at far lower rates for various reasons.

So we have to ask:

What car and driver is the AV replacing?
What is that car and driver’s death rate?

Autonomous Vehicles: At What Cost?

To find an appropriate benchmark car, we have to know how much AV rides will cost. Waymo is the clear leader in actually deploying AVs, but they’ve been super-mum about future passenger pricing.

General Motors (GM) is recognized as a leading contender for runner-up in the race to put out self-driving cars, and it projects a price of $1 per mile by 2025. Ford, another top contender, has also projected a cost of $1 per mile. Between charging (if electric) and low demand at night, these vehicles will still get nowhere near 100% utilization. And there’s strong potential for unknown unknowns to drive up beyond those prices.

Waymo Jaguar: $1 per mile would be a bargain

At $1 per mile, that price is 250% of the average of $0.40* per mile that Americans currently spend on driving. So if we’re looking for a manned car to beat, it should be something that costs far more than the average new car, like a mid-sized, new-ish Mercedez Benz sedan, which costs roughly $1 per mile for the first five years, according to Edmunds. Alternatively, the fatal crash rate for Uber & Lyft rides might be an even better benchmark, but that figure is not available.

(*Per mile cost methodology: The Bureau of Labor Statistics (BLS) reports an average of $8,427 per household spent on vehicle costs, and households average 1.968 vehicles, and cars average 11,244 miles driven per year, giving us a total of $0.38 a mile. The Internal Revenue Service uses a figure of $0.545, though it’s based off vehicles used for business. The American Automobile Association quotes even higher figures, but they are based on much newer than average cars.)

How Often Does a Mercedes Benz Kill Someone?

The short answer: the average, fairly new large luxury car is associated with one death per 800 million miles, or 0.13 deaths per 100 million miles, one ninth as often as the average across all cars. This figure is calculated using the weighted average driver death rates for “large luxury sedans,” from the International Institute for Highway Safety (IIHS), an independent non-profit.

The IIHS figures: 6.83 driver deaths per million car years, weighted average

IIHS figures are stated in driver deaths per million registered vehicle years. I converted this to miles, using the average number of vehicle miles per year for cars. I then converted driver deaths to vehicle occupant deaths, based on statistics that tell what percent of fatal crash victims in a vehicle were drivers (74%). I then convert vehicle occupant deaths to total deaths, figures that measure what percent of crash deaths were vehicle occupants (64%).

To be sure, the percentage and vehicle mileage figures are just averages across all cars, so this is a best guess — more granular figures are not available. If the average driver death rate for all cars is used, we get approximately 1 death per 175 million miles, or 0.57 deaths per 100 million miles, roughly half the rate of all cars.

Source
% of passenger vehicle occupant deaths that are the driver: 74%. Source
% of motor vehicle crash deaths that are passenger vehicle occupants: 64%. Source

Tesla

Additionally, such a metric would mean that as of March 2018, Teslas are 2.4 times as deadly as peer vehicles, given how Elon Musk has stated that Teslas have a death every 320 million miles. Since that statement, there has been a spate of Tesla occupant deaths —Florida teens Baratt Riley and Edgar Monserrat Martinez, and a German man in Switzerland.

Other electric vehicles, namely the Chevrolet Volt and the Nissan Leaf, do markedly better than the vehicles of the same class, small four door cars, according to the IIHS. With a rate of 7–8 driver deaths per million vehicle years, this is only ~10% higher than luxury sedans. The Volt and the Leaf cost more than other cars, and attract a different clientele — presumably older and higher income, versus non-electric small cars which are often bought by younger and lower income people.

Unlike the Volt and Leaf, Tesla is strongly marketed with a generous wink towards dangerous driving —i.e. ‘ludicrous’ speed and use of a premature Autopilot.

Many experts say full autonomy will require lidar, which costs $75,000, and that makes their widespread use in privately owned cars dead on arrival, for now. Leading players like Waymo and GM rely on them. Teslas do not have them — yet the company says their cars have all the hardware needed for full autonomy.

Tesla marketing. Source

Without knowing the total vehicle miles and deaths, I can’t say how Teslas compare now, but given 3 deaths in such a short period since the statement (less than 2 months), it’s likely that Tesla’s death rate is even higher now. Tesla’s rank as the #1 most costly vehicle to insure does not exactly improve the picture. Tesla could easily clarify the situation by sharing its data.

How Many Miles Would It Take AVs to Yield Fewer Deaths?

The respected think tank RAND set out to answer a related question: how many miles would AVs have to drive, before we could conclude with 95% confidence (the standard threshold used in statistical academic inquiry) that they are safer than manned vehicles?

RAND used the fatal crash rate for all cars, of 1.09 deaths per 100 million vehicle miles, or a death every 92 million miles (a slightly lower rate than the latest figure I found) for its analysis. Using conventional statistics, it concluded 275 million vehicle miles were needed, or three times the expected period between crashes (i.e. 275/91.7 = 3.00).

RAND formula for calculating number of miles needed to judge if AVs have a lower fatal crash rate

Using the same method but applied to our benchmark, we obtain a figure of 2.3 billion miles needed without a single death, to be 95% certain that AVs have a lower fatal crash rate than the manned vehicles they are most likely to replace.

This is simply a volume of miles that is too big to be covered by vehicle testing alone. To paraphrase RAND, with a fleet of 100 autonomous vehicles being test-driven 24 hours a day, 365 days a year at an average speed of 25 miles per hour, this would take just over 100 years, or 10 years with a fleet of 1,000 vehicles.

RAND goes on to ask about how many miles are needed to achieve greater certainty that AVs have a better safety record (using the 1.09 deaths per 100m miles, and not my figure of 0.13 deaths per 100m miles):

However, the RAND report continues, that to get much precision about the “true” rate of AV deaths, hundreds of billions of miles are required:

For low failure rates — 1 per 100 million miles, which is akin to the human-driven fatality rate — demonstrating performance to any degree of precision is impossible — requiring billions to hundreds of billions of miles. These results show that as autonomous vehicles perform better, it becomes harder — if not impossible — to assess their performance with accuracy because of the extreme rarity of failure events.

Waymo

Google’s AV effort Waymo, has racked up the most miles of any contender, at 6 million. If it were just as safe as our benchmark car and driver above, we’d expect… 0.008 people to have died by now. Meaning, its performance to date is entirely irrelevant for making any conclusions about whether it will result in any fewer crash deaths.

Waymo has announced that it will deploy “up to” (note the ambiguity) 62,000 vehicles. If Waymo were to actually deploy this many, it would take about 6 months to reach the 95% confidence level of 2.3 billion miles without a vehicle death, assuming a driving rate of 70,000 miles per year per vehicle, based on NYC taxicabs.

Such an aggressive rollout is a huge gamble. 62,000 cabs amounts, at 70,000 miles a year, is equivalent to nearly 400,000 privately owned cars driving 11,000 miles per year (from earlier figure). Whether that gamble is worth making right now is another matter. It’s also unclear when these will actually fully deploy.

The Status Quo: A Poverty of Will, not Robots

Self-driving cars must rely on complicated technologies that will be difficult to perfect. But they also require some that a) are relatively easy to implement and b) are easily imposed on human drivers, technically speaking anyway. (I tweeted as much, and it garnered a viral, overwhelmingly positive, response.)

For instance, limiting a car’s speed limit as a group within the European Union is hoping to do, on new cars by 2021:

Intelligent speed assistance communicates the current speed limit using satellite and speed sign recognition and helps keep the vehicle within the posted limit.
Speeding is a contributory factor in a third of fatal crashes in this country so such a measure could have a significant positive impact.

I’m just some guy with a keyboard and a spreadsheet, and I can come up with a number of solutions that will probably save lives and be much easier than full autonomy:

  • Localized speed limiters, as above
  • Forced braking to ~5 mph or less at stop signs and red lights (unless there are green arrows)
  • Classification algorithm that determines with 99%+ accuracy that a cell phone is being operated by a driver (based on kinematic data and other variables), and responds by locking phone except for emergency calls or limited voice functions
  • Attention detection system that alerts inattentive or sleeping drivers
  • Always on side view video display (if this actually helps; I witnessed one on an Uber ride recently, it seemed perfect for detecting cyclists without distracting the driver)
  • Mandatory breathalyzer installation in vehicles of all drivers convicted of a DUI (a policy that has had success in New Mexico in reducing DUIs)
  • Policy solutions like fixing broken, dangerous street designs, reducing street speed limits and requiring side guards on trucks

Or, as Sully, the pilot who landed a plane on the Hudson river, would call these: ‘guardrail’ technologies. And I’m sure experts could come up with a much more comprehensive ‘safety wishlist.’

Saving Human Drivers From Themselves Will Make the AV’s Job Easier

There are any number of technical and political interventions that can be done to save lives, yet many of the boosters for autonomous vehicles seem to have a studied ambivalence about such changes — though it would make autonomous mobility easier to achieve, by making drivers more predictable and compliant with the law, and reducing the risk when they do.

That said, the more such solutions are implemented, the fewer fatal crashes there are from manned driving, and the higher the threshold is for AVs to do a better job of preventing lives lost, than manned vehicles.

AVs will work, eventually… in certain places and times — it won’t be a switch flipped around the world. And there’s plenty that can be done now, when AVs aren’t available. Since AVs are not anticipated to eliminate cars everywhere upon arrival, preceding AV rollouts with such solutions would maximize safety improvements.

Towards a Smarter Rollout of AVs

The rate of crash deaths is far from uniform. AVs could be deployed selectively, to the places where they are most likely to have a positive impact. That means places where vehicle deaths are highest, provided they have environments that the AV can manage (i.e. probably not Alaskan winters). In such places, it would be easier to achieve reductions in death rates; they may also be easier, insofar as they’re less urban locations. Think of states like South Carolina and Kentucky — not DC, New York City and Boston (home to districts/states that are 48, 49 and 51 in the traffic death rate per mile). Sacramento, not San Francisco (14.4 vs 4.0 deaths per 100,000 people annually).

Additionally, if AVs are primarily replacing high risk trips, such as for nightlife when drunk driving rates are far higher, then the threshold for AVs doing better than the status quo is lower. If fewer traffic deaths is truly the goal, AVs must be recognized as a possible future solution when others are already proven, standing by, ready for our use. When we will it.