The big NHTSA database of all robocar crashes is interesting, but incomplete. By looking at the data and trying to estimate who is at fault in each crash, it reveals new information about the safety records of the biggest projects. While we mostly care if robocars are causing crashes, knowing about the crashes where they are probably not at fault reveals hints of important things, like how many miles the team is actually driving, and how good they are. The data on Waymos show lots of crashes where they were just sitting stopped and got rear ended by somebody. This is also happening with Teslas, but not enough . Which with some irony, is bad for Tesla.

Traditional crash safety research has always avoided trying to decide who is at fault in a crash. That’s because it’s quite difficult. Crashes can be complex, the law is complex and there’s very often just not enough information to figure this out. Fault will be assessed if there’s a lawsuit, or for insurance claims. In some states, insurance companies have pushed for a “no-fault” regime to avoid the complexity of doing that.

It’s much easier to get an accurate count of the number of crashes a vehicle is “involved” in, whether they did the hitting or were hit--we don’t care. That’s because for large-sample statistics, we get our result without the risk of biasing our data. Average humans will be at-fault in well over 50% of their crashes. For example, roughly 25% of all crashes involve a single vehicle, and that driver is almost always at fault. In two vehicle crashes (about 65-70%) we can presume fault is divided evenly, but a large fraction of crashes have more than one driver at fault, or will get assigned a “percentage blame” figure. With more than 2 vehicles it can be worse. You can see why the safety experts don’t want to go here.

In evaluating the safety of a robocar, though, it’s hard to avoid. Robocar teams are seeking to make vehicles that drive much safer than the average human. That’s new--while there are of course above average humans (and insurers try to find them) it doesn’t matter for status. With robotaxis, Waymo has a fleet of 3,000 vehicles all with the same computer driver. Some studies published by Waymo (but done by outsiders) have suggested the Waymo driver might be seven times safer than the average human driver in the same situations, though some have disputed this.

Let’s say in data on 1,000 crashes, we find the average human driver at fault in 700 of them, and not-at-fault in 300. The Waymo, with the claimed numbers, would have only 100 at-fault crashes (7x better) but it would still have been hit, not-at-fault, 300 times, for a total of 400. Now 400 is much better than 1,000, but if you are just counting “involved” crashes it’s only 2.5x better, not the 7 better which is the truth. That’s no small error.

With fatalities, it’s even stronger. Waymo has been “involved” in 2 fatalities. One it was clearly not at fault (a driver going 98mph crashed into a line of stopped cars, including the Waymo. One person and a dog in a different vehicle were killed.) In the other, the Waymo slowed down quickly in its lane to yield to pedestrians in the driveway it was turning into. A motorcyclist rear-ended the Waymo and was flung into the next lane, where a hit-and-run driver struck him fatally.) While it can be argued there are steps the superhuman computer in the Waymo could have taken to mitigate the crash, they are not things a human would do, and in the law a rear-ended vehicle in this situation is not-at-fault.

The two “involved" fatalities might raise eyebrows. Waymo has driven >220M miles, and humans are involved in about 2.7 fatalities on average in that distance, though fewer on city streets. But those are unreasonable eyebrows if the Waymo has in reality caused no fatalities.

I downloaded the NHTSA crash data. I asked an LLM to make a simple estimate of fault on all the crashes. Of course, this will have a number of errors and hallucinations. However, I and others did a human review of the results and found that a sampling the LLM results indicated they were good enough for some basic, high-level analysis. There is another significant bias here--the descriptions of the crashes are written by the companies, who are motivated to paint the best possible picture of themselves. I think a worthwhile step would be to require that independent parties write or review the crash details, based on the very detailed 3D recordings every robocar makes of any incident. In fact, I think that independent parties should actually come up with fault estimates.

A lot of the crashes are pretty easy to assess, though. They involve vehicles that were stopped at things like traffic light getting rear ended. Cars backing up into other stopped cars. Single car crashes where the vehicle hits a pole or static objects. (Road debris is more ambiguous, fault can be with the party who dropped it.) Quite a few crashes involve others hitting the doors of a robotaxi when passengers opened them up into traffic, often against the warning of the vehicle. I nonetheless assigned partial fault to the vehicle there.

One can argue about a number of the assessments, but it doesn’t matter, because I’m only looking at the grand pattern. Even if some of the numbers change by major amounts, it doesn’t alter the basic conclusions. Nobody should treat the numbers found as precise .

You can look at the data and AI/human analysis in this spreadsheet . However, this is not represented as rigorous,

Driving Means Getting Into Crashes

The first thing that’s obvious is that the mere act of going out on the roads for a significant number of miles means you are going to get into a bunch of crashes, even if you never make any mistakes yourself. From June 2025 to June 2026, Waymo reports around 1,000 crashes they were involved in. But the fault analysis suggests they caused 12% of them, and probably no more than 15%. That’s a very good score, and in line with other research on their safety record. An earlier study found only 4% were sole fault of the robocar.

Nobody else has nearly as many miles, and only two other companies have more than 50 crashes. Zoox has 53, with only 3 measured as at-fault. That’s extremely good--perhaps too good and a result of Zoox writing, consciously or unconsciously, the narratives from their perspective. (It’s also because a lot of Zoox’s operations to this point are in their regular test vehicles, which are Toyota Highlanders with a safety driver supervising.)

AVRide also has over 50 crashes with roughly 20% fault, however, AVRide only tests with a safety driver (sometimes in the right hand seat, but that is still a safety driver no matter what people try to convince you of.) A properly supervised vehicle should not be getting into many crashes at all. In fact, ideally it should do better than a random human driver, both because the safety driver is professional, and because there are “two sets of eyes”--machine and human--looking out for trouble. Tesla has reported that even untrained supervisors (their car owners) drive about 1.5x better with their FSD system on than they drive on their own.

Tesla’s crash count of 22 is small because they simply haven’t driven many miles. They show fault in roughly half the crashes, though almost all those miles are supervised which suggests rather poor performance.

An interesting side-result emerges. This database has been criticized because teams are not required to report how many miles they have driven their vehicles, so there is no “denominator” if one wishes to calculate crash rates per mile driven. It turns out that the number of “involved, not-at-fault” crashes is a proxy for miles driven. While more research is needed, it seems that if you drive around 170,000 miles, somebody is going to hit you. It’s true that some cars can be better at avoiding being hit--Waymo did a famous study where they recreated all fatal crashes in Chandler and found that if they put themselves in the role of the not-at-fault car, they could still prevent most of the crashes. But it’s going to happen.

That Tesla only shows 12 not-at-fault crashes has one clear meaning--t hey just aren’t driving very much . This is confirmed by sites like Robotaxi Tracker which watches cars through public traffic cameras, and occasional data releases from Tesla itself. While Tela’s CEO regularly claims their vehicles are ready for wide deployment and will see massive deployment this year, if you look at what they do rather than what they say, we know they feel they are not ready.

Others have also asked whether robocars are doing something unusual and “inhuman” that might make them get hit more often than ordinary drivers. That could include things like “phantom braking” or just driving strangely or extra conservatively. Waymo said back in 2015 that their own research into this had not found a problem (they were curious about this, too) but I’m not aware of additional or more formal research on it. If it’s true, one would want to account for it. Less mature robocars tend to be tuned to drive very conservatively, and we want them to be that way for obvious reasons. That means they block traffic a bit more, and perhaps surprise other drivers more.

Of course, it would be much better if the standing order required some reporting of miles operated, broken down by city or road type, and by whether there was an in-car supervisor, remote supervisor, or no supervision. Then we would not even care about the crashes in which the vehicle had no fault, other than we want to make sure that’s true. With complete 3-D recordings and data on the “state of mind” of the robot, determining fault should be much easier than for human crashes. The problem is that companies are loathe to ever report they were at fault, as that could open up legal liability. That’s why having a mostly independent panel determine fault is a better approach.