Tesla Releases Q3 Vehicle Safety Report: Autopilot Reduces Crash Risk


Autopilot lessens the likelihood of a crash.

At Tesla, safety is paramount. As such, the automaker routinely releases safety-related reports (typically on crash-test results). But moving forward, the automaker now says it will release a quarterly safety report.

The vast majority of this first such report focuses on Autopilot. Here are some highlights:

Here’s a look at the data we’re able to report for Q3:

  • Over the past quarter, we’ve registered one accident or crash-like event for every 3.34 million miles driven in which drivers had Autopilot engaged.
  • For those driving without Autopilot, we registered one accident or crash-like event for every 1.92 million miles driven. By comparison, the National Highway Traffic Safety Administration’s (NHTSA) most recent data shows that in the United States, there is an automobile crash every 492,000 miles. While NHTSA’s data includes accidents that have occurred, our records include accidents as well as near misses (what we are calling crash-like events).

Check out the full report in its entirety below:

Q3 2018 Vehicle Safety Report

At Tesla, the safety of our customers is our top priority, which is why it’s critical that we design and build the safest cars in the world. Not only do we conduct extensive in-house testing and simulation to ensure our vehicles achieve top safety performance before they ever reach the road, we are also uniquely positioned to leverage the hundreds of thousands of miles of real-world data our fleet collects every month to continuously improve our vehicles and develop a more complete picture of safety over time.

Because every Tesla is connected, in most instances we are able to learn immediately when a Tesla vehicle has been involved in a crash. Additionally, our non-traditional sales model allows us to have a direct relationship with our customers for the lifecycle of ownership, providing an avenue for us to supplement our records and gain even more insight as needed. In contrast, automakers whose cars aren’t connected and who utilize networks of third-party franchised dealers may never know when a vehicle is involved in an accident. Through traditional channels, it can take months or even years for lawsuits or claims to be filed that provide automakers with insight into an accident that allows them to draw meaningful conclusions and improve safety.

Earlier this year, when we made the decision to begin publishing our safety data on a regular basis, we designed and introduced a completely new telemetry stream for our vehicles to facilitate these reports. This new data stream allows us to gather the most critical fleet-wide statistics from the exact moment a crash-related event is detected by our system. While there are still some unique cases in which crash data may not be available to us through this channel, we believe this system currently provides the best framework for safety reporting on an ongoing basis.

Here’s a look at the data we’re able to report for Q3:

  • Over the past quarter, we’ve registered one accident or crash-like event for every 3.34 million miles driven in which drivers had Autopilot engaged.
  • For those driving without Autopilot, we registered one accident or crash-like event for every 1.92 million miles driven. By comparison, the National Highway Traffic Safety Administration’s (NHTSA) most recent data shows that in the United States, there is an automobile crash every 492,000 miles. While NHTSA’s data includes accidents that have occurred, our records include accidents as well as near misses (what we are calling crash-like events).

Moving forward, we will publicly release these accident figures on a quarterly basis.

Given the degree to which accidents can vary in severity and circumstance, we’ve started an additional initiative to create a more complete picture of safety by gathering serious injury data from our customers following an accident. While we have long maintained the practice of calling our customers whenever our system detects a crash in order to see whether they need emergency assistance, we now also use these calls to understand if they sustained an injury in the crash, and if they have feedback on our current safety system. This will help us continue to improve our system and understand the rate of serious injuries over time.

We also encourage our customers to proactively contact Tesla Support if they are ever seriously injured in a Tesla vehicle, or if they have suggestions about improving safety features.

As we are working hard to make our cars the safest and most capable cars on the road in terms of passive safety, active safety, and automated driving, we must continue to encourage driver vigilance on the road – that is, by and large, the best way to prevent traffic accidents. Safety is at the core of everything we do and every decision we make, so we cannot stress this enough. We want our cars to not only lead the way to sustainable energy, but also make driving as safe as possible for everyone, and we are working as quickly as we can to achieve that. We look forward to sharing continued updates with our customers and community, and working together to make our vehicles as safe as possible.

Categories: Tesla

Tags: , , , ,

Leave a Reply

50 Comments on "Tesla Releases Q3 Vehicle Safety Report: Autopilot Reduces Crash Risk"

newest oldest most voted

*if used properly* I think AP can reduce the risk of a collision. For me when on freeways / divided highways it frees me up to look to the sides and far down the road for possible hazards. In other words, it frees up brain / vision bandwidth that would otherwise be spent watching the car directly in front of me and keeping in my lane.

That may be.

But I find the data distinctly unimpressive. I think most drivers use AP like you, most of the time. Hence, AP is mostly engaged when driving on highways, and mostly disengaged on other roads. But the “mean miles between accident-like events” is of course MUCH higher on the highway than on other roads, in any car and with human-only driving. Stating that mean miles between accident-like events is 74% higher when AP is engaged than when it isn’t therefore is extremely weak evidence that AP is safer than humans. In fact, my guess based on this data runs the other way.

Some day, however, I am pretty sure that computers will be much better drivers than we are. I’ve just never seen any good evidence that they currently are.

You didn’t provide a number. What is the average mileage between accidents on highways?

They don’t account for the safety biases in the different stretches of road. When autopilot is disengaged, most likely the road is a bit more tricky than the easy segments which auto pilot can handle.

Exactly! The data is real, but the implied comparison is stupid.

Unless we know the road condition of which the data are taken from, it is pointless to compare. But Tesla seems like to spin those data and facts in their favor without providing the proper comparison condition.

For example, autopilot is more likely to be used in open stretch of highways rather than where autopilot mode is not recommended. Those type of highways and conditions naturally have lower accident rate. But Tesla didn’t disclose that (even though their data logger combined with GPS data can potentially normalize that data). When autopilot is not used in crowded cities, the road conditions are more complex or more crowded which is more likely to have accidents. So, until we can normalize those two conditions, it is pointless to compare the two sets of data.

Lastly, comparing again the entire hwy data is stupid as we know that brand alone makes a difference in crash rate, much less the driving patterns. Compare a single brand and a single brand with a feature vs. the entire market is just marketing tactic.

The NHTSA has reported that, per Tesla’s data, Tesla cars with Autopilot + Autosteer merely installed have a nearly 40% reduction in the accident rate as compared to Tesla cars without AutoSteer. The accident rate was measured by airbag deployment, and ignores whether Autosteer is actually active at the time of the accident. So that eliminates any possibility of bias, and it’s comparing Tesla cars to Tesla cars — unquestionably an apples-to-apples comparison.

Serial Tesla bashers keep trying to tell us to ignore that data, because it “doesn’t actually prove that Autopilot + Autosteer is 40% safer”. And in a perverted way they’re right, because of course use of AutoSteer must reduce the accident rate by much more than 40%, in order to drag the overall accident rate down by 40% even including cars which didn’t have it activated at the time of the accident!

And now you’re trying to come up with another excuse for ignoring some pretty compelling evidence that Autopilot + Autosteer makes driving significantly safer.

I certainly hope you’re not in charge of making auto accident actuarial charts for insurance agencies!


Again, personal attack mixed your own lack of understanding. Did you actually read the actual study? Doesn’t seem so. The so called Tesla data isn’t direct apple to apple. It was taken over two different period of Model S with and without autosteer, not compared with the same vehicle whether with autopilot on and off. So, before you attack others again with your own ignorance, you should read up about facts before you expose yourself as an ignorant fanboi again.

Just to remind you, I have no problem exchange fire between you and I. Time after time, I have tried to make you learn about responding without personal attack. But it seems that you are INCAPABLE of doing so. If not, then you should maybe excuse yourself from this forum as you are unfit to participate.

Things like lane keeping and adaptive cruise control technology certainly do reduce accidents, which is why organisations like IIHS give cars with them extra points in their safety ratings and manufacturers are installing the technology in more and more vehicles.

That has nothing to do with his point though – using the data the way it’s been provided by Tesla is basically meaningless marketing.

What they should do is either analyse it themselves or give the data to a group to analyse (i’m sure a university would love to do it) properly, so proper conclusions can be reached (which as I mentioned in another post probably will evidence that AP is safer in the right circumstances – it may also evidence that it increases accidents/unsafe driving in specific circumstances as well, but we won’t know until it’s actually analysed properly and/or results are released).

I’m sure Tesla do analyse it themselves in that way as method of making AP safer – where are it’s weaknesses, where are it’s strengths, what should they be concentrating on from a safety point of view etc – they just don’t release it.

Does the report you are linking to also record the age of the drivers? Is they studied the age of the drivers? If it did you can probably prove that driving a car with autosteer makes teenagers ten years older…. I see many teenagers driving old Teslas that are worth under $40k but I don’t see them on $100k talked with autopilot.

It isn’t that either of these groups are intentionally trying to mislead with data. It is simply very difficult or impossible with current data to rope out confirming factors that unfortunately ruin the effectiveness.
I haven’t yet seen a convincing study.

Why don’t you go and read it, and make up your own mind about what it says, instead of asking others to read it and tell you what it says?

Far better to link to the actual report that Electrek based its article on. Especially since Electrek is not exactly an unbiased source when it comes to Tesla. The report is freely available directly from NHTSA: https://static.nhtsa.gov/odi/inv/2016/INCLA-PE16007-7876.PDF, and the relevant section is 5.4: “5.4 Crash rates. ODI analyzed mileage and airbag deployment data supplied by Tesla for all MY 2014 through 2016 Model S and 2016 Model X vehicles equipped with the Autopilot Technology Package, either installed in the vehicle when sold or through an OTA update, to calculate crash rates by miles travelled prior to21 and after Autopilot installation.22 Figure 11 shows the rates calculated by ODI for airbag deployment crashes in the subject Tesla vehicles before and after Autosteer installation. The data show that the Tesla vehicles crash rate dropped by almost 40 percent after Autosteer installation.” This is the best – well, frankly the ONLY – evidence I have seen so far that AP does decrease accident rates. I have trouble squaring this with the data Tesla reported here (and IEVs forwarded in this article). A 40% reduction in crash rate (by miles travelled, regardless if AP is engaged or not) is equivalent to a 67% increase… Read more »

Also interesting to note are the NHTSA’s crash rates, stated in airbag deployments per million miles, translated to the inverse, million miles between airbag deployments:

Before AutoSteer: 1.3 deployments/mm => 769,231 miles between airbag deployments.
After AutoSteer: 0.8 deployments/mm => 1,250,000 miles between airbag deployments.

This is rather a rather interesting discrepancy that seriously undermines EITHER the NHTSA’s report, or Teslas numbers in this quarterly report – you take your pick. Tesla claims 1,920,000 miles between “accident-like events” when AP is not engaged, and 3,340,000 miles when AP is engaged.

Since Tesla claims far more miles between each “accident-like event” than NHTSA does miles between airbag deployments (based on data that was delivered to NHTSA by Tesla) it seems certain that Tesla must have excluded a whole lot of events that include at least one airbag deployment from their “accident-like” events.

In my view this casts a very strange light on Tesla’s data in this quarterly safety report.

I think the point here is that we’ll get a quarterly report so we can track improvements over time regardless of any biases.

That is indeed the only interesting part. But it would have been so much better if it came from an academic third party with access to the full Tesla database – doing their own selection and classification and interpretation of the raw data. We have no guarantee that Tesla will use identical methodology next time. Even so, this is by far the most interesting part of the statement.

If Tesla did divide the data up by some arbitrary manner, would there be data for other cars to compare that to? Are there overall accident fatality numbers grouped by highway vs. urban vs. rural roads?

Who would be making the decision about what is a “highway” and what isn’t, or what is a “rural” road vs an “urban” road? It’s well established that different people apply the exact same rules differently — remember the problem of counting Florida ballots with “hanging chads”? Once you start introducing arbitrary categories, you introduce bias, like it or not.

The real world isn’t a pristine laboratory where it’s possible to eliminate every variable but one. The real world is messy. The best we can do is look at the available evidence and make judgements as best we can based on that evidence. Refusing to look at the evidence we have because it’s not an absolutely 100% perfect apples-to-apples comparison, is downright stupid.

I think it’s appalling that anyone would advocate that data grouped by some arbitrary designation into one category or another, should be preferred to the unbiased total data.

The so called grouping of data doesn’t impact the integrity of the data itself. For the case of autopilot on/off, it is directly comparable if they cover the same segment or type of roads. That kind of data is directly comparable. If Tesla is capturing all those data, then it is possible it can dissect the data to compare the operation directly over the same stretch of road with autopilot on/off.

The data is just data. It is meaningless on its own without proper details, conditions or interpretation. Grouping or category doesn’t change the data. It only changes the interpretation of the data itself. The fact that the data is compared without information on how the data itself is “collectecd” ( in the sense of where the condition is applied) shows a pre-existing bias. That is the issue here.

As far as how another thinks one way dissecting is better than other, especially here, well that is an endless waste of time anyway.

As posted earlier. https://www.dot.ny.gov/divisions/operating/osss/highway-repository/Average%20Accidents%20Rates%20Table_2016.pdf NY State have data for each year from 2011 here: https://www.dot.ny.gov/divisions/operating/osss/highway/accident-rates There are literally hundreds of studies on the subject of safety in different environments – governments, cities, planners, insurers and a myriad of other groups will be generating and using them constantly. If you mean other car manufacturers then it may be harder to find, BUT this discussion is not actually about a car manufacturer (Tesla in this case), it’s about the safety of a system they have developed and are advocating for it’s effectiveness. If you’re advocating the benefits of a system then you need to provide evidence for that – the numbers above are not reliable evidence in and of themselves – they’re basically meaningless and don’t really move anything forward for the autonomous vehicle advocate (and I am one). And just to clarify, I’d argue this point if any other manufacturer of a safety system released the same type of numbers, without any real analysis of them. EDIT: and to further refute the points you’re making, especially that of “I think it’s appalling that anyone would advocate that data grouped by some arbitrary designation into one category or another, should be preferred… Read more »

Exactly. And the biases are worse than you give the impression of – at least if AP is mostly engaged on highways, mostly disengaged otherwise. EVERYONE drives *far* more miles between each incident on highways than on any other roads.

BS. You claim irrelevant statistics, but fail to provide any to prove your point.

There isn’t any relevant statistic, that I am aware of. It is routine when considering any data set to be on the lookout for potential biases and attempt to control for them, and take it into account if no control is possible. Tesla has done no such thing.

My “assumptions” here are so basic that I consider them to be known fact.

1) AP is more often engaged on highways/motorways than other roads.
2) People drive more miles between incidents on highway/motorways than other roads.

Which assumption do you dispute?

If you don’t dispute either, it follows that a 74% increase in miles driven between incidents when AP is engaged does not tell us how much, if any, of that increase is due to AP being engaged, and how much is due to the driving taking place on safer roads.

TSLA stock down to $280 Oct-4 3 PM.
May be a last under 300 buying opportunity, as the “financial” media continues to low-ball Tesla prospects.

Simple math is beyond the reach of some.
80,000 cars produced is a 320,000 Yearly Run Rate now.
If Tesla is even a bit in the red this quarter, it’s Green next quarter and beyond.

Remember, it was just a few days ago the Mr. Lutz “speculated” that TESLA production was only 200,000.
How can you be that far off reality?

The “Low-ball Tesla prospects”, could unfortunately continue for a while longer, as the “ShortSeller Enrichment Commission”, might have something more to say, now that Elon is back at it again, being his usual Twitter self.

Their market cap is not based on produced vehicles. If it was it would be significantly lower than what it is now.

There’s no reason to think it’ll magically leap over and stay over $300 tomorrow. Something has to actually happen. A list of things that would/could cause it:
1 – Analyst releases an estimate of what Tesla’s results will be.
2 – Analyst gives a new price target.
3 – Tesla releases Version 9
4 – Press talks about Version 9 – overwhelmingly positive reactions.
5 – Tesla announces financial details
6 – Tesla reveals updated products (IE, Model Y, Semi, or pickup)

So… yeah, you can buy for $280 right now. Probably you’ll be able to buy at a similar price tomorrow, and throughout next week, until one of those things happens. Obviously other things could happen, but those would be even harder to predict.

I do continue to expect we’ll see $400 and beyond before the end of the year (possibly even before the end of the month, depending on when those different things I listed occur.)

Why are you talking about stock price when the topic at hand is about Tesla’s spin on safety data?

Even though I agree with you, pumping up the stocks on a non-stock related article is just as annoying as people who short stocks.

TL;DR: It’s all about sales, and future law requirements, which are rightfully part of equation stock market use for evaluating the worth of given shares.

Relative safety as compared to human driver-operated cars does have immense influence over the future of Tesla.
It’s not unreasonable to assume that at some point in the future some form fo driver assist will be mandated by law, in USA or in EU. Companies that have the best hardware will have an easier time meeting those requirements. Indeed law may copy-cat those capabilities into requirements, forcing other companies to license such technology if their own can’t be adopted soon enough. Last but not least, relative safety is a significant differentiator for some potential buyers, and other brands then Tesla also put immense emphasis on that aspect of their offerings. A significant shift in perception would thus influence potential buyers of Tesla cars.

It is now 6.79 times safer to be on autopilot than your average car in america. (3.34 million miles / 429,000 miles = 6.79) Elon Musk said that at 10 times safer, the BETA badge would be removed. I wonder how long it will take to get to 10 times safer?

Plenty of brand/models can make similar claim. Compare any single model/brand vs. overall market is stupid claim.

That is like comparing watermelon with the entire fruit basket.

Exactly. As a (non AP) example. Tesla owners are going to be older, more experienced drivers (most are probably male and >40) so will inherently crash less than a 16 year old who has just passed their test in their first used 10 year old Honda Civic.

More of a guess, the average Tesla driver (assuming the same demographic above) are more likely to be driving more highway miles and less town miles compared to stay at home mum in her minivan, popping in and out of car parks and residential streets, with a higher accident risk.

There are lots of factors that the press release completely ignores, as evidenced by the accident rate of Tesla’s not being controlled by AP, which is 2-3x lower than the general population.

Also ignored is the tendency for reckless driving by some Tesla drivers, because Teslae are high performance cars, which encourages irresponsible drivers to drive recklessly. That is shown by some truly spectacular crashes in Tesla cars, with the car going over 100 MPH at the time of the crash.

But of course, you’re ignoring that because it doesn’t support what appears to be your anti-Tesla agenda. (◣_◢)

I don’t have an Anti Tesla agenda, just pointing out that the way those figures are presented means little without further analysis.

Interesting though, are you suggesting that Tesla drivers are more dangerous than their similar demographic companions in other vehicles? May explain the increased insurance premiums I guess? 😉

Just ignore PuPu, he has nothing to offer beside attacking anyone who criticized Tesla spin. The simple fact that he can’t separate talking about the topic or rebut an argument without personal attack just shows his simple lack of understanding.

The problem I have with these claims is that Autopilot is usually engaged on highways or highway similar roads, which have the lowest risk of accidents per mile anyway and compare these to numbers of ALL driving including city driving where accidents occur in higher numbers

Exactly! Spin of statistics are typical marketing scheme used on stupid people.

You are the one advocating putting spin on the statistics, by claiming the data should be broken down into arbitrary categories of highway vs. non-highway miles.

No, I am advocating for providing MORE INFORMATION On the statistics taken without making any suggestion or comparing them. Not just taking the spin of the statistic itself. The so called accident rate per miles is a spin because there are NO INFORMATION provided on how those two set of data are taken. It was ONLY by autopilot on or off. There was NO information given on whether they are taken over similar set of samples (number of trips, trip length, time of data, highway vs. inner city, average traveling speed during the data collection…etc).

Not necessarily, highway accidents are many of the deaths as they are often single vehicles leaving the roadway, which AP can be helpful to prevent. Also, as AP gets better people will be more confident leaving it on in more situations.

Speaking of crashes, I wouldn’t want any EV with a slightly damaged battery pack recharging in my garage overnight.

You wouldn’t want a car with a leaky gas tank either. BMW apologized and recalled over 100,000 diesels in Korea as they were catching fire. 20 or so in about a month. Did it make the news? Not really.

Plus at least a few houses burned down. BMW refused for a year to even agree that there was a problem, with cars that sat for hours spontaneously combusting. Of course if Tesla’s were catching on fire it would be all over the, bought and paid for media. Despite dozens of reports, with video, BMW continued to deny there was a problem for almost a year, after initial reports came in..
Finally they did a recall, but are still foot-dragging to compensate owners.

“Because every Tesla is connected, in most instances we are able to learn immediately when a Tesla vehicle has been involved in a crash. Additionally, our non-traditional sales model allows us to have a direct relationship with our customers for the lifecycle of ownership, providing an avenue for us to supplement our records and gain even more insight as needed.”

You Tesla owners are being spied on! And I thought Fakebook and Google were bad. Yes, I know all manufacturers are creeping into the spy game.

Tesla is hardly alone in this. All GM vehicles with OnStar track crashes as well. And any vehicle with GPS tracks your location.

The more relevant question is: what is Tesla likely to be doing with the info they collect? Facebook and Google provide free platforms for social media posting and searching, respectively. So selling users’ browsing information is their primary business model and advertising is the only way they make money.

Tesla’s primary business model is selling cars, solar panels, and battery storage systems. It would seem logical that they would use the information they gather to further their income based on their business model, i.e. sell more cars, etc. Improving customer service and vehicle capabilities seems like it would offer a higher investment return than selling user location data or vehicle settings.

So it seems more likely to Tesla’s business model that they are using the information they collect to improve their product and sell more cars, and less likely that they are seeking to sell user information to a third party.

I don’t disagree much, but just a small point: “Any vehicle with GPS tracks your location” is a bit misleading in this context. Obviously it does track your position then and there, but this is a far cry from *logging* your position for permanent keeping and, crucially, submitting it to the manufacturerer for their own use.

Unfortunately the way they are portraying those stats is extremely wonky.

AP is generally used on open, divided roads, where crashes are 3-5x less likely than in towns and cities, where AP is generally not used.

One of the reasons non AP controlled Tesla’s are are in more crashes than when AP is used may well be because of that.

That’s not to say AP is not more safe than a human driver, just that those stats are meaningless unless they’re sorted by road type.

Here’s an example of what I mean, showing the significant variation in crashes between a rural controlled, divided highway and urban roads with lots of intersections.


Well stated.

What is Tesla’s Safety record compared to the German’s. R they 6 times better than BMW audi and Merc ?

Although I would like more details along the lines of the NHTSA Fatality Analysis Reporting System (FARS) data, this summary is a good start. FARS data includes when, where, and how with enough details to do a useful statistical analysis. Hopefully Tesla will expand the safety program to obtain automated records from other sources such a injury reporting systems.

The reason is “Bell the Hybrid” was passed using extremely doggy statistical analysis. An analysis that also supports the hypothesis of the windshield “A” pillar blocking in turning car-pedestrian accidents. When making a left turn, the “A” pillar can block vision in one eye leaving only the other eye with a blind spot to detect a pedestrian crossing. Instead, the “sound” pedestrian alert was passed into law:

In 2016, I replaced a perfectly fine, 2010 Prius with a 2017 Prius Prime to get the standard Toyota Safety Sense-P (TSS-P) Pre-Collision. Even this basic system makes both highway and city driving much easier. Of course I have to monitor what is going on and steer the car. But even this system supports driving strategic.

The only “real” bit in this commercial for Tesla’s AP was this little tidbit:

> Moving forward, we will publicly release these accident figures on a quarterly basis.

As each quarter goes by this will at least enable us to see whether the “AP engaged” MMBALE relative to the “AP disengaged” MMBALE improves over time or not, so we’ll get an indication of whether AP is getting better (than it was, not than a human) or worse.

Has the insurance rate decreased?