Human drivers stop inappropriately far more often than Waymo's do. This is not a unique behavior of Waymos, this piece is just detailing them because it's providing a factual account of the causes of the (small) number of accidents. If this list also included every crash caused by human drivers under similar circumstances it would be many times longer!
The fact that each Waymo vehicle is in far fewer crashes per mile driven suggests that this is a per capita advantage for the Waymo. But there may be specific conditions that they are worse in, if we cut things finely enough.
If a car doesn't cut you off, stops short, and you hit it from behind, then you are tailgating and are responsible. If you were not tailgating, then you should have driven around it.
I couldn’t be more averse to self driving cars in general. First of all, who asked the citizens if this is what we wanted in the first place? Secondly, why are we paying for these technologies, which are being controlled remotely and most-likely in another country entirely, with our tax dollars? I’m just not okay with it at all, and overly cautious drivers usually cause the most accidents anyway, just saying. Sudden braking is still a HUGE safety risk that should not be ignored.
Waymo is a private company, so no tax dollars are involved. Also, the cars are not driven remotely. The remote assistance can not steer or move the car, only give suggestions.
Alright then let’s talk about the legality of them being able to record people without their permission and gather data. Who stores that data and controls the privacies of people who didn’t even sign up for their services? A class action waiting to happen, and I can’t wait.
You don’t seem to mind. Good for you. I do mind, and it’s pretty messed up that random companies can just gather data on all civilians with no disclosure or permissions/cooperation from the public. I’m a firm believer in privacy protections.
You must not care about public privacy matters at large, and that’s your personal preference, but that doesn’t mean THIS country has established ANY communication regarding this matter with the public. We are supposed to be a democracy, in case you forgot.
Most cars today have cameras on the outside, in lieu of mirrors in many cases. Public transit has cameras. Stores have cameras capturing the sidewalk outside. Nest cameras on the outside of many houses. If you're outside, you may be observed by other people - this is a fact of life, always has been, and legally preventing it would require significant impositions on both personal freedom and public safety.
Given that the main "pro" for self-driving cars is that they seem *dramatically* safer than human-driven cars, and the main "anti" seems to be a vague "I don't like it" accompanied by inaccurate statements about how the tech works and how it is funded, my personal vote as a "citizen" is YES. It is already approaching anti-vax levels of irrationality and will only continue in that direction.
Surely you could supply examples of these taxes? It actually goes the opposite way — the service is heavily taxed by the state.
When you believe something for particular reasons, and then those reasons turn out to be completely false, you should change what you believe, and you should also doubt your own ability to reason more.
From Google: Yes, cities with self-driving vehicle services like Waymo are implementing, or considering, new taxes that affect residents and users of these services. While Waymo is a private, for-profit company, cities are creating tax frameworks to address infrastructure, congestion, and revenue losses from traditional fuel taxes.
They are electric vehicles. They do not pay gas tax, because they do not purchase gasoline. The quoted excerpt is specifically describing how they are taxed more than others.
Can you not read what you quoted? Cities are creating new taxes for Waymos. How do you get that they are being subsidized from that? It says the exact opposite.
Wow - so confidently incorrect and it absolutely doesn't support the argument you think you are making. The taxes are going to come from users of Waymo and because the EV's do not pay gas taxes they have to address the loss of revenue. The average citizen who does not use the service will not be getting taxed to pay for this. Maybe read the articles, and not just the Google summary. You can be against this, but you don't have to be unreasonable or just make up nonexistent reasons to be against it...
You're talking about taxes paid by the self-driving taxi companies. Sometimes those are per-ride fees that are passed down to the consumer of the rides in their fares - the same way that there are fees paid by taxi companies that are reflected in fares that users voluntarily pay. That doesn't amount to "people" being broadly taxed to pay for Waymo to operate in their city
Citizens want this. These companies pay taxes. They are primarily controlled locally, by the AI system in the vehicle, not remotely.
And it just isn’t true that most collisions are caused by overly-cautious drivers. Do you believe that we could decrease traffic injuries if we forced everyone to take a reckless driving class?
I was in a Waymo that stopped inappropriately last summer, and there was no accident, but it was very scary. Like - we stopped on the middle of an intersection because the light turned red. Fortunately, the traffic was light and the other drivers adjusted, but it (along with a couple of other experiences on that trip) really made me feel like it was the mix of autonomous and manned vehicles that was extra dangerous. Kind of like accidents in warehouses with robots. Dynamic judgment and decision making is very hard to program.
I did have an odd experience in a Waymo in Los Angeles last time I rode a few weeks ago. I was riding with my nephews to the La Brea Tar Pits, and there were a couple school buses parked outside the parking lot. There was one more turning to arrive there, and our vehicle cut it off to make the final approach, and then stopped to let us out next to the school buses, since they were occupying the entire curb lane, rather than pulling up another 100 feet.
It still was overall fine, but those moments of interaction were scary, completely unlike any other moment I’ve been in Waymo’s before.
Hi Rick! I do not think that human drivers run into Waymos more often than they run into other human drivers. Quite the contrary. The overall rate of serious Waymo-related crashes (regardless of who caused them) is around 80% lower than for a typical human-driven vehicle. It wouldn't possible to reduce the crash rate that much if Waymo was only avoiding crashing into other cars as much as human drivers. Waymos also seem to be driving defensively so as to reduce the likelihood that other vehicles run into them.
I'm remembering a story (maybe a novel, maybe a short story) from the '50s by Arthur C. Clarke in which a visitor to a city is shocked to find, there, that operating one's own car within the city is still legal. It's sort of like our horror at thinking, of the past, that slavery was legal, or that we used lead acetate in paint, or (continuing with lead) that we added tetra-ethyl lead to gasoline (poisoning millions). It's clear that human-operated vehicles are, right now, much more dangerous than autonomous vehicles. But autonomous cars are only a kind of interim solution. Already they are safer, in general, than human drivers. But it's easy to imagine a system where all cars are part of a hierarchical system in which the city > neighborhoods > roads > intersections tell the cars what to do and what routes to take. Imagine no stoplights or stop signs. Everyone gets to read/study/game/whatever while being transported instead of having their time enslaved by the need to drive. In such a system, no one would actually need to own a car, and there would be no need for parking lots or personal garages, freeing up immense amounts of real-estate . It's the ultimate public transportation system, free from the inefficiencies of defined routes, stops, hubs, or anything other than "where are you now?", "where do you want to go?", and "when do you want to do that?".
I've been making this same point for years, that in the not too distant future, humans will be prohibited from driving, first on highways and then on city streets.
This will then result in no need to OWN cars. This will kill off cars as status objects, will kill off the auto insurance industry, will kill off the need to manufacture new car models every year, will kill off auto sales advertising and much more!
This is what transition friction looks like. The machine may already be better at the task in aggregate, but civilization has not yet rebuilt the environment around that fact. Until it does, a significant share of failure will occur at the boundary between synthetic discipline and human disorder.
Referring to an old comedy trope that the most dangerous thing about stoned drivers is how slow they move, and that they're prone to like "waiting for the stop sign to turn green."
Why don't they add more of a human backup? In the rare case where there is an accident, or something goes wrong, why not have the vehicle alert well trained humans, who will see through the car's sensors and be able to instruct the computer on what to do, or take over driving?
In-fact, there are some amounts of human backup. When a car gets confused, it will send a question to a remote operator (in the US or Philippines), and the person will respond. In the cases of an accident, I'm assuming that Waymo operational personnel are immediately notified (because the car can sense it's in an accident).
Taking over driving is a little bit more iffy because of latency challenges, so Waymo remote operators basically can't do that, except for one very small edge case that Waymo has said has never happened. If you're curious for more, Tim wrote a piece about this recently:
I have an adjacent question, and I'm not sure if it's addressed in existing documentation. I appreciate that the fairest comparison is to the average (or mean) driver, particularly in the hypothetical "what would happen if we replaced all drivers with self-driving cars?".
But I wonder what the comaprison is to the median (or typical) driver. I would expect the poorest safety outcomes are concentrated very disproportionately in the worst drivers. That would answer a related question which is something like "what would happen if we replaced the typical driver by a self-driving car?"
In fact, it would be fun to answer this at different percentiles e.g. a 90th percentile, 50th percentile (median) and 10th percentile driver.
This is a really good question, and one that I wonder about a lot. I will try to address this in my next article about the broad statistical safety record of Waymo and others (that should come out sometime this summer, but no promises haha), but I don't think we have very good data on this.
One of the challenges about quantifying a 90 percentile driver is that there's just so much random luck that determines car crashes. If someone never gets in a crash, is that because they're a good driver or because they just got lucky?
Another challenge is just the human data we have around driving isn't particularly high quality.
The formal category error about "inappropriate stopping" is a structural failure of the Observational Sufficiency Principle (OSP). In the 2018 Uber crash, the system failed because it couldn't map the "pedestrian" atom. Modern AVs are making a reciprocal error that they 'Stand-Still' because their internal Fully Specified Stochastic Process (FSSP) cannot assign a probability to human variance that falls outside their σ-algebra. They aren't being "safe" as opposed to that they are effectively hitting a logical wall. This Stand-Still Error is precisely where the machine "flatters" its own internal model until it forces a physical confrontation with the real world. I have referred this as a structural isomorphism error to the Monty Hall Problem, Patient Zero to the current AI Pandemic.
Waymo cars make mistakes due to excessive caution is probably true, yet those mistakes can be very costly to others, as shows the example of the truck that fell off the road. Passengers in a less sturdy vehicle could have been killed in this scenario.
The actual more cautious behavior of the Waymo car would have been to drive on at minimal or regular highway speed and exit at the next ramp, as probably most cautious humans would have done. (For example, a teenage driver whose parents forbade them to go on the freeway.)
Let's hope that Waymo will iron out those bugs (I've heard of similar geofencing problems elsewhere) before they roll out to more cities and increase their number of cars (and areas around each city...).
I’m not necessarily anti self driving vehicles. Although I see the obvious anxiety surrounding it, the data shows that they are safer than I would have thought.
The issue im having is that they are using the idea of the vehicle’s being overly cautious as a “pro argument” while I would argue that being overly cautious is still dangerous along with reckless driving.
A safe driver, is a confident and PREDICTABLE driver.
First, thank you for this great article! Nice to have a full picture of the Waymo accidents and how humans actually play a bigger part in them. I think a lot of the anxiety and uncertainty people feel come from a fear of accepting new things which is very understandable.
About the Chinese government doesn't release public data about crashes, I think this is a misunderstanding. While China doesn't provide a centralized database like the DMV, the relevant data is not withheld from the public. The operation of autonomous vehicles in China is strictly regulated. For instance, companies like Apollo Go have explicitly stated that traffic accidents, such as being rear-ended, are reported to relevant regulatory authorities for the record immediately. Management committees or transportation authorities in certain operating regions, such as Wuhan and Beijing, periodically respond to incidents involving registered vehicles. Official data also indicates that the accident rate for driverless vehicles is approximately 1/14 that of human drivers.
As for "censorship", it's more of a broad censorship to tone down the public anxiety toward autonomous vehicles. Even then, on Chinese social media platforms like Douyin, Xiaohongshu, and Weibo, there is a vast amount of user-generated video and photographic content regarding Apollo Go vehicles getting stuck, involved in collisions, or causing traffic congestion, which frequently triggers large-scale public discussion. Mainstream Chinese media outlets, such as The Paper, Cailianshe, and Sina Finance, also track and report on autonomous driving accidents and obtain formal responses from either the government or the involved companies. There's a level of censorship, but it's no where close as "everything is taken down."
Thanks for the clarification! If I (or a Chinese citizen) wanted to look through Chinese crash data, what would I need to do? Are there any databases I could look at, even of just certain deployments or cities? What's the official data that you are indicating here that leads to 1/4th the crashes?
I know that the reporting requirements are quite comprehensive for Chinese companies (probably more comprehensive than the US -- I've heard engineers complain to me about how hard it is to keep the relevant data on the car). The thing I wasn't able to find when I was looking into this was whether any of that reporting ends up being publicly accessible. Would love any pointers on where to look!
Two of baidu (China's wikipedia) pages recording self-driving car accidents.
Again, China doesn't have a live, centralized public database like NHTSA’s. This isn't a policy unique to autonomous vehicles. The most authoritative data available to the public typically comes from the annual or semi-annual statistical reports of the National Bureau of Statistics or the Ministry of Public Security. These reports generally only contain the total number of accidents, casualties, economic losses, etc. and do not disclose specific vehicle models or technical details of individual accidents. Although local government open data platforms have been launched in places like Shanghai, Zhejiang, and Shenzhen—opening up some 'transportation sector datasets'—the content is mostly focused on business licenses, basic vehicle information, or traffic condition broadcasts, rarely involving detailed individual accident databases. It's just a different regulatory and data-sharing infrastructure.
I know that searching for data can be tiresome in a situation like this. I hope these materials help and I wish you the best of luck.
Thank you for linking all of these! I'll definitely do some more digging the next time I write about robotaxi safety records. The big Chinese companies are the only ones with similar amounts of experience as Waymo, so are very relevant to the global story of robotaxis.
(In case it's not clear, I don't mean to criticize Chinese regulations here. Nor China specifically. As far as I can tell, there aren't any countries outside of the US that have a centralized database of crashes like NHTSA releases, even in Europe.)
The pattern here is fascinating because it flips the usual intuition. We judge autonomous vehicles harshly for caution, but tolerate absurd levels of human error as background noise. If the data keeps showing fewer injury crashes, the cultural adjustment may be harder than the technical one.
So that would still mean that they are vastly less involved in crashes than humans.
Human drivers stop inappropriately far more often than Waymo's do. This is not a unique behavior of Waymos, this piece is just detailing them because it's providing a factual account of the causes of the (small) number of accidents. If this list also included every crash caused by human drivers under similar circumstances it would be many times longer!
The fact that each Waymo vehicle is in far fewer crashes per mile driven suggests that this is a per capita advantage for the Waymo. But there may be specific conditions that they are worse in, if we cut things finely enough.
If a car doesn't cut you off, stops short, and you hit it from behind, then you are tailgating and are responsible. If you were not tailgating, then you should have driven around it.
Driving slowly doesn't cause accidents. Tailgating causes accidents. I would rather have a slow driver than an accident.
I couldn’t be more averse to self driving cars in general. First of all, who asked the citizens if this is what we wanted in the first place? Secondly, why are we paying for these technologies, which are being controlled remotely and most-likely in another country entirely, with our tax dollars? I’m just not okay with it at all, and overly cautious drivers usually cause the most accidents anyway, just saying. Sudden braking is still a HUGE safety risk that should not be ignored.
Waymo is a private company, so no tax dollars are involved. Also, the cars are not driven remotely. The remote assistance can not steer or move the car, only give suggestions.
Honestly… you think we aren’t paying for it? lol
Alright then let’s talk about the legality of them being able to record people without their permission and gather data. Who stores that data and controls the privacies of people who didn’t even sign up for their services? A class action waiting to happen, and I can’t wait.
On public roads there is no expectation of privacy. It's legal to record, although it might differ in your state.
You don’t seem to mind. Good for you. I do mind, and it’s pretty messed up that random companies can just gather data on all civilians with no disclosure or permissions/cooperation from the public. I’m a firm believer in privacy protections.
It's not about minding or not. I've lived in countries with stricter privacy laws than the US, and I prefer them.
But when you talk about "legality" and "class actions", what matter is the current law, not your or my personal preferences.
You must not care about public privacy matters at large, and that’s your personal preference, but that doesn’t mean THIS country has established ANY communication regarding this matter with the public. We are supposed to be a democracy, in case you forgot.
Do you believe it is illegal for humans to have dash cams on their cars?!
Most cars today have cameras on the outside, in lieu of mirrors in many cases. Public transit has cameras. Stores have cameras capturing the sidewalk outside. Nest cameras on the outside of many houses. If you're outside, you may be observed by other people - this is a fact of life, always has been, and legally preventing it would require significant impositions on both personal freedom and public safety.
Given that the main "pro" for self-driving cars is that they seem *dramatically* safer than human-driven cars, and the main "anti" seems to be a vague "I don't like it" accompanied by inaccurate statements about how the tech works and how it is funded, my personal vote as a "citizen" is YES. It is already approaching anti-vax levels of irrationality and will only continue in that direction.
Ever heard of data being the new gold? Where is it being stored by these private companies and who gives them rights to sell it? Your point is null.
🤣🤣🤣 The more you post, the more foolish you look!
Point two is false, top to bottom. You should stop believing in false things.
In which city do you believe we aren’t being taxed for this service? Pretty sure we are paying for it somewhere somehow.
Surely you could supply examples of these taxes? It actually goes the opposite way — the service is heavily taxed by the state.
When you believe something for particular reasons, and then those reasons turn out to be completely false, you should change what you believe, and you should also doubt your own ability to reason more.
From Google: Yes, cities with self-driving vehicle services like Waymo are implementing, or considering, new taxes that affect residents and users of these services. While Waymo is a private, for-profit company, cities are creating tax frameworks to address infrastructure, congestion, and revenue losses from traditional fuel taxes.
They are electric vehicles. They do not pay gas tax, because they do not purchase gasoline. The quoted excerpt is specifically describing how they are taxed more than others.
No, it’s not. Do your own research.
Can you not read what you quoted? Cities are creating new taxes for Waymos. How do you get that they are being subsidized from that? It says the exact opposite.
Google it yourself. Search your city and tell me you won’t be taxed for it.
Wow - so confidently incorrect and it absolutely doesn't support the argument you think you are making. The taxes are going to come from users of Waymo and because the EV's do not pay gas taxes they have to address the loss of revenue. The average citizen who does not use the service will not be getting taxed to pay for this. Maybe read the articles, and not just the Google summary. You can be against this, but you don't have to be unreasonable or just make up nonexistent reasons to be against it...
You're talking about taxes paid by the self-driving taxi companies. Sometimes those are per-ride fees that are passed down to the consumer of the rides in their fares - the same way that there are fees paid by taxi companies that are reflected in fares that users voluntarily pay. That doesn't amount to "people" being broadly taxed to pay for Waymo to operate in their city
Citizens want this. These companies pay taxes. They are primarily controlled locally, by the AI system in the vehicle, not remotely.
And it just isn’t true that most collisions are caused by overly-cautious drivers. Do you believe that we could decrease traffic injuries if we forced everyone to take a reckless driving class?
Who is this "we" you refer to?
I want this so I gave Waymo some of my money to drive me around last time I was in SF
I was in a Waymo that stopped inappropriately last summer, and there was no accident, but it was very scary. Like - we stopped on the middle of an intersection because the light turned red. Fortunately, the traffic was light and the other drivers adjusted, but it (along with a couple of other experiences on that trip) really made me feel like it was the mix of autonomous and manned vehicles that was extra dangerous. Kind of like accidents in warehouses with robots. Dynamic judgment and decision making is very hard to program.
"Fortunately, the traffic was light and the other drivers adjusted ..."
The most important word here is "drivers". What would have happened if, say, a third of the them were various AV's, instead of human guided?
Most probably they would also have stopped, like the waymo.
I did have an odd experience in a Waymo in Los Angeles last time I rode a few weeks ago. I was riding with my nephews to the La Brea Tar Pits, and there were a couple school buses parked outside the parking lot. There was one more turning to arrive there, and our vehicle cut it off to make the final approach, and then stopped to let us out next to the school buses, since they were occupying the entire curb lane, rather than pulling up another 100 feet.
It still was overall fine, but those moments of interaction were scary, completely unlike any other moment I’ve been in Waymo’s before.
Yeah, you just want to be able to whisper to it “in this case, I need you to trust my judgment”
IF
Waymos are hit by human drivers statistically more often than
human drivers are hit by other human drivers
then something IS going on with the Waymos. They are behaving differently in an important way.
“Legal” is not a high enough standard. Expectations of normal driving behavior matter.
Hi Rick! I do not think that human drivers run into Waymos more often than they run into other human drivers. Quite the contrary. The overall rate of serious Waymo-related crashes (regardless of who caused them) is around 80% lower than for a typical human-driven vehicle. It wouldn't possible to reduce the crash rate that much if Waymo was only avoiding crashing into other cars as much as human drivers. Waymos also seem to be driving defensively so as to reduce the likelihood that other vehicles run into them.
By normal behavior, do you mean speeding, cutting other people off, and road-raging?
I'm remembering a story (maybe a novel, maybe a short story) from the '50s by Arthur C. Clarke in which a visitor to a city is shocked to find, there, that operating one's own car within the city is still legal. It's sort of like our horror at thinking, of the past, that slavery was legal, or that we used lead acetate in paint, or (continuing with lead) that we added tetra-ethyl lead to gasoline (poisoning millions). It's clear that human-operated vehicles are, right now, much more dangerous than autonomous vehicles. But autonomous cars are only a kind of interim solution. Already they are safer, in general, than human drivers. But it's easy to imagine a system where all cars are part of a hierarchical system in which the city > neighborhoods > roads > intersections tell the cars what to do and what routes to take. Imagine no stoplights or stop signs. Everyone gets to read/study/game/whatever while being transported instead of having their time enslaved by the need to drive. In such a system, no one would actually need to own a car, and there would be no need for parking lots or personal garages, freeing up immense amounts of real-estate . It's the ultimate public transportation system, free from the inefficiencies of defined routes, stops, hubs, or anything other than "where are you now?", "where do you want to go?", and "when do you want to do that?".
I've been making this same point for years, that in the not too distant future, humans will be prohibited from driving, first on highways and then on city streets.
This will then result in no need to OWN cars. This will kill off cars as status objects, will kill off the auto insurance industry, will kill off the need to manufacture new car models every year, will kill off auto sales advertising and much more!
Thanks for the great detailed article. I really enjoy these.
This is what transition friction looks like. The machine may already be better at the task in aggregate, but civilization has not yet rebuilt the environment around that fact. Until it does, a significant share of failure will occur at the boundary between synthetic discipline and human disorder.
Sounds like Waymo has a "stoned driver" problem, and y'all published this story two days late.
Hi Jason! I’m not sure what you mean here. Can you say more?
Referring to an old comedy trope that the most dangerous thing about stoned drivers is how slow they move, and that they're prone to like "waiting for the stop sign to turn green."
Here are a couple of many examples:
https://www.instagram.com/reel/DIrJni6RZ2e/
https://www.instagram.com/reel/DPIKp2Gic5o/
Why don't they add more of a human backup? In the rare case where there is an accident, or something goes wrong, why not have the vehicle alert well trained humans, who will see through the car's sensors and be able to instruct the computer on what to do, or take over driving?
In-fact, there are some amounts of human backup. When a car gets confused, it will send a question to a remote operator (in the US or Philippines), and the person will respond. In the cases of an accident, I'm assuming that Waymo operational personnel are immediately notified (because the car can sense it's in an accident).
Taking over driving is a little bit more iffy because of latency challenges, so Waymo remote operators basically can't do that, except for one very small edge case that Waymo has said has never happened. If you're curious for more, Tim wrote a piece about this recently:
https://www.understandingai.org/p/waymo-just-revealed-a-crucial-statistic?utm_source=publication-search
I have an adjacent question, and I'm not sure if it's addressed in existing documentation. I appreciate that the fairest comparison is to the average (or mean) driver, particularly in the hypothetical "what would happen if we replaced all drivers with self-driving cars?".
But I wonder what the comaprison is to the median (or typical) driver. I would expect the poorest safety outcomes are concentrated very disproportionately in the worst drivers. That would answer a related question which is something like "what would happen if we replaced the typical driver by a self-driving car?"
In fact, it would be fun to answer this at different percentiles e.g. a 90th percentile, 50th percentile (median) and 10th percentile driver.
Love the substack :)
This is a really good question, and one that I wonder about a lot. I will try to address this in my next article about the broad statistical safety record of Waymo and others (that should come out sometime this summer, but no promises haha), but I don't think we have very good data on this.
One of the challenges about quantifying a 90 percentile driver is that there's just so much random luck that determines car crashes. If someone never gets in a crash, is that because they're a good driver or because they just got lucky?
Another challenge is just the human data we have around driving isn't particularly high quality.
The formal category error about "inappropriate stopping" is a structural failure of the Observational Sufficiency Principle (OSP). In the 2018 Uber crash, the system failed because it couldn't map the "pedestrian" atom. Modern AVs are making a reciprocal error that they 'Stand-Still' because their internal Fully Specified Stochastic Process (FSSP) cannot assign a probability to human variance that falls outside their σ-algebra. They aren't being "safe" as opposed to that they are effectively hitting a logical wall. This Stand-Still Error is precisely where the machine "flatters" its own internal model until it forces a physical confrontation with the real world. I have referred this as a structural isomorphism error to the Monty Hall Problem, Patient Zero to the current AI Pandemic.
Waymo cars make mistakes due to excessive caution is probably true, yet those mistakes can be very costly to others, as shows the example of the truck that fell off the road. Passengers in a less sturdy vehicle could have been killed in this scenario.
The actual more cautious behavior of the Waymo car would have been to drive on at minimal or regular highway speed and exit at the next ramp, as probably most cautious humans would have done. (For example, a teenage driver whose parents forbade them to go on the freeway.)
Let's hope that Waymo will iron out those bugs (I've heard of similar geofencing problems elsewhere) before they roll out to more cities and increase their number of cars (and areas around each city...).
I’m not necessarily anti self driving vehicles. Although I see the obvious anxiety surrounding it, the data shows that they are safer than I would have thought.
The issue im having is that they are using the idea of the vehicle’s being overly cautious as a “pro argument” while I would argue that being overly cautious is still dangerous along with reckless driving.
A safe driver, is a confident and PREDICTABLE driver.
First, thank you for this great article! Nice to have a full picture of the Waymo accidents and how humans actually play a bigger part in them. I think a lot of the anxiety and uncertainty people feel come from a fear of accepting new things which is very understandable.
About the Chinese government doesn't release public data about crashes, I think this is a misunderstanding. While China doesn't provide a centralized database like the DMV, the relevant data is not withheld from the public. The operation of autonomous vehicles in China is strictly regulated. For instance, companies like Apollo Go have explicitly stated that traffic accidents, such as being rear-ended, are reported to relevant regulatory authorities for the record immediately. Management committees or transportation authorities in certain operating regions, such as Wuhan and Beijing, periodically respond to incidents involving registered vehicles. Official data also indicates that the accident rate for driverless vehicles is approximately 1/14 that of human drivers.
As for "censorship", it's more of a broad censorship to tone down the public anxiety toward autonomous vehicles. Even then, on Chinese social media platforms like Douyin, Xiaohongshu, and Weibo, there is a vast amount of user-generated video and photographic content regarding Apollo Go vehicles getting stuck, involved in collisions, or causing traffic congestion, which frequently triggers large-scale public discussion. Mainstream Chinese media outlets, such as The Paper, Cailianshe, and Sina Finance, also track and report on autonomous driving accidents and obtain formal responses from either the government or the involved companies. There's a level of censorship, but it's no where close as "everything is taken down."
Thanks for the clarification! If I (or a Chinese citizen) wanted to look through Chinese crash data, what would I need to do? Are there any databases I could look at, even of just certain deployments or cities? What's the official data that you are indicating here that leads to 1/4th the crashes?
I know that the reporting requirements are quite comprehensive for Chinese companies (probably more comprehensive than the US -- I've heard engineers complain to me about how hard it is to keep the relevant data on the car). The thing I wasn't able to find when I was looking into this was whether any of that reporting ends up being publicly accessible. Would love any pointers on where to look!
https://hb.chinadaily.com.cn/a/202405/16/WS66456933a3109f7860ddd9fc.html
The '1/14' safety figure was released by Apollo Go and has been widely cited by major Chinese news outlets.
https://13115299.s21i.faiusr.com/61/1/ABUIABA9GAAghfuvoQYogtvkywI.pdf#page=15.12
http://www.mzone.site/Uploads/Download/2022-02-10/6204917913b55.pdf
Here're the government reports from the Beijing High-Level Automated Driving Demonstration Area including incident breakdowns.
https://www.most.gov.cn/kjbgz/202507/P020250723532256831666.pdf
Ethical Guidance on Automated Driving Technology R&D.
https://finance.sina.com.cn/wm/2026-04-01/doc-inhsysmc5022728.shtml
Here's one of the news reports on the March Apollo Go incident.
https://baike.baidu.com/item/2%C2%B722%E9%95%BF%E6%98%A5%E7%90%86%E6%83%B3%E6%B1%BD%E8%BD%A6%E6%92%9E%E8%BD%A6%E4%BA%8B%E6%95%85/65433095?fr=aladdin
https://baike.baidu.com/item/6%C2%B717%E9%97%AE%E7%95%8C%E8%BD%A6%E4%B8%BB%E5%BC%80%E6%99%BA%E9%A9%BE%E6%92%9E%E8%BD%A6%E4%BA%8B%E6%95%85/66190506
Two of baidu (China's wikipedia) pages recording self-driving car accidents.
Again, China doesn't have a live, centralized public database like NHTSA’s. This isn't a policy unique to autonomous vehicles. The most authoritative data available to the public typically comes from the annual or semi-annual statistical reports of the National Bureau of Statistics or the Ministry of Public Security. These reports generally only contain the total number of accidents, casualties, economic losses, etc. and do not disclose specific vehicle models or technical details of individual accidents. Although local government open data platforms have been launched in places like Shanghai, Zhejiang, and Shenzhen—opening up some 'transportation sector datasets'—the content is mostly focused on business licenses, basic vehicle information, or traffic condition broadcasts, rarely involving detailed individual accident databases. It's just a different regulatory and data-sharing infrastructure.
I know that searching for data can be tiresome in a situation like this. I hope these materials help and I wish you the best of luck.
Thank you for linking all of these! I'll definitely do some more digging the next time I write about robotaxi safety records. The big Chinese companies are the only ones with similar amounts of experience as Waymo, so are very relevant to the global story of robotaxis.
(In case it's not clear, I don't mean to criticize Chinese regulations here. Nor China specifically. As far as I can tell, there aren't any countries outside of the US that have a centralized database of crashes like NHTSA releases, even in Europe.)
no worries. I'm glad to be of assistance.
The pattern here is fascinating because it flips the usual intuition. We judge autonomous vehicles harshly for caution, but tolerate absurd levels of human error as background noise. If the data keeps showing fewer injury crashes, the cultural adjustment may be harder than the technical one.