For the past 10 months, the National Highway Traffic Safety Administration has been collecting data on crashes that involve the use of Advanced Driver Assist Systems (ADAS), also known as Level 2 semi-automated driving systems according to the SAE’s very confusing classification system. NHTSA actually put out two separate reports, one for Level 2 systems, (which require constant driver attention and are commercially available today) titled “Summary Report: Standing General Order on Crash Reporting for Level 2 Advanced Driver Assistance Systems” and one called “Summary Report: Standing General Order on Crash Reporting for Automated Driving Systems” for Level 3 and up systems, which are currently still in development and are intended to be true automated driving systems that do not require a driver’s attention. These reports prompted a number of news outlets to mention Tesla, mostly because the Level 2 report mentions that the company has been involved in more L2 ADAS crashes than any other automaker, by a significant margin. This sounds pretty bad for Tesla, but this is also one of those cases where a lot more context is necessary to really assess what’s going on.
Before I go into this report – I’m just looking at the Level 2 one now, because the full ADS one deals with situations too different to be compared together– I should make clear where I personally stand when it comes to these systems: I think L2 driver assist systems are inherently flawed, which I’ve explained many times before.
My opinion on L2 is that it is always going to be compromised, not for any specific technical reasons, but because these systems place humans into roles they’re just not good at, and introduce a paradox where the more capable a Level 2 system is, the worse people tend to be at remaining vigilant, because the more it does and the better it does it, the less the human feels the need to monitor carefully and be ready to take over. So the fundamental safety net of the system, the human that’s supposed to be alert, is in some ways, lost.
I’m not the only one who feels this way, by the way; many people far smarter than I am agree. Researcher Missy Cummings goes into this phenomenon in detail in her paper Evaluating the Reliability of Tesla Model 3 Driver Assist Functions from October of 2020:
“Drivers who think their cars are more capable that they are may be more susceptible to increased states of distractions, and thus at higher risk of crashes. This susceptibility is highlighted by several fatal crashes of distracted Tesla drivers with Autopilot engaged (12,13). Past research has characterized a variety of ways in which autonomous systems can influence an operator’s attentional levels. When an operator’s task load is low while supervising an autonomous system, he can experience increased boredom which results in reduced alertness and vigilance (ability to respond to external stimuli) (14).”
With that in mind, let’s look at some of the data NHTSA found. It’s worth starting by looking at how the reported crashes were categorized:
Observations from reported crashes of Level 2 ADAS-equipped vehicles are presented in this section using data reported through May 15, 2022. It is important to note that these crashes are categorized based on what driving automation system was reported as being equipped on the vehicle, not on what system was reported to be engaged at the time of the incident. In some cases, reporting entities may mistakenly classify the onboard automation system as ADS when it is actually Level 2 ADAS (and vice versa). NHTSA is currently working with reporting entities to correct this information and to improve data quality in future reporting.
So, this is a little confusing: crashes are categorized by the ADAS system the car was equipped with, and not necessarily what was reported as being active at the time of the crash? Also, the mistake made by “reporting entities” of calling Level 2 ADAS semi-automated systems as “ADS,” or full self-driving systems (Automated Driving Systems), is a good reminder about how much public education still remains to be done regarding these systems.
The report also seems to contradict this a bit, as the criteria for what crashes get reported must include a Level 2 ADAS system being active at any time within 30 seconds of the crash (emphasis mine):
The General Order requires that reporting entities file incident reports for crashes involving Level 2 ADAS-equipped vehicles that occur on publicly accessible roads in the United States and its territories. Crashes involving a Level 2 ADAS-equipped vehicle are reportable if the Level 2 ADAS was in use at any time within 30 seconds of the crash and the crash involved a vulnerable road user or resulted in a fatality, a vehicle tow-away, an air bag deployment, or any individual being transported to a hospital for medical treatment.
I think this means that the crashes only get reported if a Level 2 system was active within 30 seconds of the incident, but the categorization may just be organized based on the installed equipment. At least, that’s what I think it’s saying.
Let’s go right for the most headline-tempting chart, the total number of crashes by car brand using Level 2 Advanced Driver Assist Systems:
So, obviously Tesla looks pretty bad here, with 273 crashes, over three times the next closest carmaker, Honda, which is nine times more than the next one, Subaru, with 10 crashes. Does this mean that Tesla’s Autopilot and Full Self-Driving Systems (FSD) are significantly more dangerous than everyone else’s Level 2 systems?
Well, as you likely guessed, it’s not all that simple. This chart of number of crashes doesn’t tell us the total number of miles driven or how many cars are equipped with these systems, or what percentage of drivers actually use them, and without any data like that, we really can’t make an assessment if the Tesla system is really three times as dangerous as, say, Honda’s systems, which we in turn can’t say if those are nine times more crash-prone than Subaru’s and so on.
Now, I do have some theories about what’s going on here, and why Tesla’s numbers are so high. Compared to all the other automakers on this list, Teslas are more likely to have some form of L2 driver assist systems, since all Teslas come with the basic Autopilot system, and similar systems are still more likely to be optional on other cars.
So, I think there are simply more L2 ADAS-capable Teslas on the roads, percentage-wise, than other cars, and, equally importantly, I think Tesla drivers have two cultural traits that factor in here: There’s much more focus in the Tesla community – hell, there’s more of a community, for better or worse, period, for Tesla owners than, say, for Hyundai or Honda owners – and that community pays a lot of attention to these sorts of driving automation systems, which makes them more likely to use them, and more likely to use them more often, and in more situations and driving conditions where a Toyota or Volvo or whatever owner might just drive normally.
This same culture is also one that encourages a lot of faith in an almost cult-like belief in an ongoing “mission” to achieve self-driving, and from there, a still-mythical generalized Artificial Intelligence, and this sort of focus and faith can lead L2 users to attribute more capability to their cars’ ability to drive than the car is actually capable of, which in turn leads to a lack of proper oversight and attention, which leads to crashes.
There’s also the possibility that Tesla’s data-logging and reporting capabilities, which are pretty extensive, may be a factor in how many crashes are reported. As the report states that carmakers that build in more advanced telematics and reporting systems will by their nature provide more data and communicate better, and cars with less advanced telematics and data logging capability may simply not be recording or reporting all incidents:
Manufacturers of Level 2 ADAS-equipped vehicles with limited data recording and telemetry capabilities may only receive consumer reports of driving automation system involvement in a crash outcome, and there may be a time delay before the manufacturer is notified, if the manufacturer is notified at all. In general, timeliness of the General Order reporting is dependent on if and when the manufacturer becomes aware of the crash and not on when the crash occurs. Due to variation in data recording and telemetry capabilities, the Summary Incident Report Data should not be assumed to be statistically representative of all crashes.
For example, a Level 2 ADAS-equipped vehicle manufacturer with access to advanced data recording and telemetry may report a higher number of crashes than a manufacturer with limited access, simply due to the latter’s reliance on conventional crash reporting processes. In other words, it is feasible that some Level 2 ADAS-equipped vehicle crashes are not included in the Summary Incident Report Data because the reporting entity was not aware of them. Furthermore, some crashes of Level 2 ADAS-equipped vehicles with limited telematic capabilities may not be included in the General Order if the consumer did not state that the automation system was engaged within 30 seconds of the crash or if there is no other available information indicating Level 2 ADAS engagement due to limited data available from the crashed vehicle. By contrast, some manufacturers have access to a much greater amount of crash data almost immediately after a crash because of their advanced data recording and telemetry.
There’s also the issue of usable domain; GM’s SuperCruise, for example, is designed to only operate on roads that have been mapped and approved for use by GM, where other systems, like Tesla’s Autopilot, are not limited this way.
As a result, there are just fewer places for a GM L2 ADAS vehicle to operate, and the places that it can likely have a greater margin of safety because they have been pre-vetted. So, does that mean that SuperCruise is safer than Autopilot? You could read it that way, but it’s still not really and apples-to-apples comparison.
Also interesting in the report is what the cars actually crashed into:
Passenger cars are, of course, at the top of the list, but the largest category is “unknown,” which sounds pretty mysterious. First responder vehicles are only listed twice, which is notable because NHTSA already is investigating Tesla L2-related crashes with first-responder vehicles in another study, and that one notes 14 crashes.
Also very telling is this chart that logs where the L2 ADAS-equipped vehicles were damaged:
Crashes involving L2 ADAS vehicles tended to be most commonly damaged in the front, which suggests that its the L2 vehicles that are crashing into things, as opposed to being crashed into. Maybe it’s my own biases, but I feel like this data supports the basic issues I see with L2 systems, that of driver inattention.
If most of the wrecks are caused by running into things, causing damage to the front of the car, then that would seem to suggest that drivers, perhaps lulled by their L2 systems doing the bulk of the driving for long, uneventful periods, are not being attentive to what the cars are always driving toward. In cases where the ADAS system disengages without warning or, arguably worse, doesn’t realize that it should disengage because it is making poor driving decisions, and the driver’s attention and focus have been eroded by an otherwise effective system doing most of the work, then wrecks caused by the car running into other cars or immobile barriers or other things seems a likely result.
This report is interesting, but also very incomplete, and I think it’s premature to use it to make any really sweeping generalizations about what brands of cars have more or less dangerous Level 2 driver assist systems. I think the one thing we can take from this report is that these systems are by no means flawless yet, and are no substitute for actually paying attention to the road, no matter what the hype suggests.
It’s wise for NHTSA to keep formal tabs on all of this, so I think this is a good start. I’ll be curious to see future reports that may be able to provide more data for a clearer picture of what is actually going on.