Thread by @bsilly3

Elon Musk wants us to believe that autopilot improves safety. I will show why the opposite is true.

To support its safety claim, Tesla reports accident rates comparing all miles driven on autopilot to all miles driven without autopilot. Here is the data for Q2 2019:

1/n

1 accident for every 3.27 million miles on autopilot, versus 2.19 for regular driving, suggesting that autopilot is 1.5x safer.

But is autopilot the only factor here?

Something twigged in my memory banks from a former life as a transportation engineer.

2/n

It is well known that freeway driving, despite higher speeds, is much safer than normal driving. The reason is simple: most accidents occur at intersections, and freeways don't have any. 

3/n

How much safer? Here is one study out of the U.K. that has some statistics.

4/n

bit.ly/33nHNm7

This shows that for all road types, you are 4.4X as likely to get in an accident versus driving on a freeway (motorway in U.K. parlance).

You can find similar data from the NHTSA. Numbers will vary state to state and over time, but are comparable. 

5/n

Tesla does not give us a breakdown of freeway miles. 

But we can get an approximation. Here is data from an MIT study, The Human Side of Tesla Autopilot, wherein vehicle data was collected was collected in and around Boston.

6/n

The study area:

Below is the money graph:

7/n

bit.ly/2ZFI5lN

What does the graph show? It categorizes road type by speed limit, and shows that over 90% of Tesla autopilot miles driven were on roads with speed limits over 55 mph.

These are mostly urban miles, and virtually all urban roads with speed limits over 55 mph are freeways.

8/n

Compare that to non-autopilot driving which tends to be a mix of mostly lower speeds.

We don't know exactly, but I think it is a reasonable approximation to assume at least 90% of autopilot miles are driven on freeways.

9/n

We have recent general driving data that shows the non-autopilot accident rate to be 2.19 accidents per million miles. If we adjust this by a factor of 4.4 we would expect the accident rate of non-autopilot driven freeway miles to be 8.76 acc/million miles.

10/n

We now can use these figures to estimate a 90% freeway / 10% general driving mix, comparable to the typical mix for autopilot. For this mix, we would expect non-autopilot accident rate of 8.1 acc/million miles, which compares to an autopilot accident rate of 3.27 acc/mm. 

11/n

In other words, using autopilot decreases safety by a factor of roughly 2.5.

Now this is not entirely scientific, and you can quibble with numbers, but the general conclusion will remain the same. For the time being, you are safer in a Tesla with Autopilot turned off.

12/n

Tesla has the data. So do it @elonmusk: prove me wrong.

n/n

Some folks seem to be having trouble viewing posts 4 and 9. It could be because I provided direct download links to the papers.

Here are the graphs.

The first is from "Accident Trends by Road Type", from the Royal Automobile Club Foundation.

The second is from the MIT autopilot study "Human Side of Tesla Autopilot"

Google them for more info.

n+3/n

Linking my corrections (in reply to Kawasaki) back into thread for completeness...

https://twitter.com/bsilly3/status/1159263041537466368?s=09

minor correction 4 & 7 (not 9) were the posts with graphs & links to supporting data...

I need an editor.