Tesla Autopilot Safety Report Q4 2019

[diplomat33]

Here is a quick summary:

Q4 2019:
AP On: 1 accident for every 3.07 million miles driven.
AP off (with active safety features): 1 accident for every 2.10 million miles driven.
AP off (without active safety features): 1 accident for every 1.64 million miles driven.
Tesla Autopilot crash rate increases, but still lower than without Autopilot - Electrek

Note that safety has dropped quite a bit from the previous quarter where it was 1 accident for every 4.34 million miles driven. If my math is correct, it represents a 29% drop in safety.

Any thoughts on why? Is it a sign that owners are becoming overconfident with AP?

It will be interesting to see what these numbers look like when "City NOA" is released.

 

[AlanSubie4Life]

These numbers are pretty meaningless, as has been discussed ad nauseum, but here's a more easily digested plot, FWIW. Assuming no methodology changes, I suppose the numbers can be compared to prior data - just not to anything else - like NHTSA statistics (the yellow line - do not compare these statistics to that line):

HyperCharts

I would guess these plots, specifically the peaks in "safety," are correlated with relatively benign weather in California. I'm not sure why people would use Autopilot when it's raining (using cruise control in the rain is strongly recommended against by safety authorities!), but people will do what people will do... We have had significant rain events in California in 2019 Q4, and significant rain events persisted through 2019 Q2 last year.

 

[Mardak]

If my math is correct, it represents a 29% drop in safety. Any thoughts on why?

Probably need to adjust for seasonality similar to Q1 deliveries tend to drop relative to Q4 but year-over-year, there's an increase. At least from the last 1.5 years of data, Q4 has "always" dropped from Q3.

18Q3 -> 19Q3:
+29.9% Autopilot
+40.6% Active Safety
-9.9% Non Autopilot
+1.2% United States

18Q4 -> 19Q4:
+5.5% Autopilot
+32.9% Active Safety
+31.2% Non Autopilot
+9.9% United States

 

[diplomat33]

Probably need to adjust for seasonality similar to Q1 deliveries tend to drop relative to Q4 but year-over-year, there's an increase. At least from the last 1.5 years of data, Q4 has "always" dropped from Q3.

18Q3 -> 19Q3:
+29.9% Autopilot
+40.6% Active Safety
-9.9% Non Autopilot
+1.2% United States

18Q4 -> 19Q4:
+5.5% Autopilot
+32.9% Active Safety
+31.2% Non Autopilot
+9.9% United States

Thanks

 

[Yang_H]

Just curious, how do they define "accident"? Based on insurance claim?

 

[AlanSubie4Life]

Just curious, how do they define "accident"? Based on insurance claim?

No one knows. Presumably it's the airbags going off. Is it the same metric as the NHTSA uses (putting aside all the other differences...)? No one knows.

 

[S4WRXTTCS]

I hate how they don't separate out TACC accidents versus AP accidents.

On the freeways I imagine most people will either be using AP or TACC. That way we could see how much driver inattention really played into the accident rate while on AP. Like you wouldn't use TACC to send a txt, and instead you would kick AP on.

A quarter to quarter increase is likely either just randomness or weather.

 

[Daniel in SD]

I'm curious if Teslas have the ability to avoid getting hit by other cars? It seems like that would be required to achieve a 4x reduction in accident rate!

If Tesla is using data sent by the car then it seems like they would be missing accidents that are severe enough to take out the 12V system and accidents that occur in places without cellular coverage.

 

[AlanSubie4Life]

I'm curious if Teslas have the ability to avoid getting hit by other cars? It seems like that would be required to achieve a 4x reduction in accident rate!

If Tesla is using data sent by the car then it seems like they would be missing accidents that are severe enough to take out the 12V system and accidents that occur in places without cellular coverage.

I think we can all agree these statistics are useless except (perhaps) to compare against themselves (perhaps we can see whether the safety of Teslas specifically is improving...). It's pretty clear that the stats are meaningless as a means of gauging safety relative to other vehicles.