I basically agree with this article but I will make two complaints.
1. We certainly have lots of data about crashes from Waymo, but there's lots of other data we don't have that's safety relevant. To start with, crashes are reported in detail but miles driven is just something you get from their advertising blog posts. We don't have breakdowns of highway vs city street, deadhead vs passenger, etc.
More significant for safely is that we don't know anything about disengagements and remote assistance. This doesn't impact how we should assess waymos safety record so far but is significant for how to assess it going forward.
2. I think footnote 3 is wrong because Waymo is a service not a car. If we were considering the safety record of a similar level AV that people owned I would agree that the impact on public transit is not the same as the safety record and belongs in a different conversation. But instead Waymo effectively is a public transit system. So you have to consider whether it's getting people out of Ubers significant safety win), their own car (big safety win but way more variance), or off of the bus (safety negative).
I'm unhappy with your framing of #2. If you're disappointed that AVs are diverting people from safer buses/trains, then you should improve your bus/train system to compete with AVs on convenience. If you can't get your shit together enough to do that, banning AVs (but not human-driven personal vehicles) in order to force people (but only those who can't drive themselves) into buses/trains anyway is absurdly cruel, unfair, and anti-abundance.
(This topic triggered me because I have serious driving anxiety and used to take an incredibly inconvenient train to work before Tesla FSD arrived in my life, diverted me from public transit, and in so doing saved my career and my sanity. If someone tried to take my FSD away on the basis that trains are safer, without either 5xing the frequency of said trains or stopping my coworkers from driving to work manually, I would go to nuclear war with them.)
I didn't say that we shouldn't allow Waymo or that we shouldn't consider convenience or any of the other things you're suggesting. What I said is that when thinking about the safety of the transportation system, Waymo is actually more like a public transit service rather than a private car, and so safety comparisons should include shifts from other systems, rather than just "is this car safer than that car".
In contrast Tesla FSD is fundamentally making individual private cars work differently, and so comparing primarily to other private cars is generally the right frame. (It's always the case that changes in the car market will induce some mode shift but that shouldn't be the primary analysis.) I think that analysis is not that great for Tesla but we don't really have good enough data to tell.
Right, acknowledging that Waymo is safer than other AVs without also acknowledging there’s an army of people behind the scenes operating them is tantamount to lying.
First of all, that sounds like a gross exaggeration, but anyway—why? Every single automated Tesla has an entire human behind the scenes ready to operate it *from the steering wheel* at all times. No one is testing automated vehicles that have no possibility of human intervention under any conditions, because that would be an insane thing to try to do.
How is it a gross exaggeration? It is a fact that Waymo has as many employees as they have vehicles on the road. Your Tesla comparison is nonsensical in that respect. The promise of Tesla is that you just straight up don’t have to drive the car, it takes care of that. The crashes that happen with Teslas in FSD mode are almost certainly because the driver took that too literally and didn’t intervene when required. Waymos don’t work like that. Someone, or maybe several someones, we don’t know, is monitoring the vehicle at all times. Not acknowledging that is, like I said, tantamount to lying.
I’m not sure where you got this “fact,” but it might go back to an ancient report about Cruise, which counted all employees of any kind, not remote operators. We don’t actually seem to have these data publicly, but anecdotally, I’ve used Waymo about 40 times, mainly in 2022-2023, and had a remote intervention twice, once my fault and once the car’s. You’re writing as if the cars are all being remote operated in real time, but they literally cannot be remote operated at all, the latency doesn’t allow it.
It's not lying at all -- the safety record is the safety record, regardless of how it works. Also, other AVs also have lots of people behind the scenes -- opacity about this is universal in the industry (which is bad). Questions about disengagments and remote assistance are important for understanding the future, and for thinking about how this will relate to private cars, but it doesn't change the actual accident numbers.
I agree that it would be helpful to compare the safety of public transit to Waymo. For example, in 2023 88 people were killed when struck by an NYC subway train. You would also need to account for the number of assaults, sexual assaults, murders, and suicides (roughly 1/2 of the subway deaths) that occur occur on or by contact with a public transit vehicle vs Waymo. When you do that, I suspect public transit will be less safe on every one of those metrics.
Why in god's name would you consider suicides to be a relevant stat when considering passenger safety unless you were deliberately trying to juke the results?
Because it is very difficult to commit suicide by Waymo (it has never happened) and very easy to do so by train. As an individual ride you may not care, but it absolutely is relevant if you are trying to assess the harm of Waymo vs train on a population level (and consider the harm to drivers who are involuntarily used to facilitate a couple suicides per week on MTA. That must be devastating.). But even if you exclude suicide Waymo is still safer….and Public transit still subjects riders to assaults, sexual assaults, and murders that Waymo does not, in addition to fatalities and injuries from the mode of transport itself.
If we take the National Transit Database figure of 8 billion passenger miles in 2023 for the NYC subway, and 97 fatal train collisions + 5 murders in 2023 as our denominator, that gets us 78 million passenger miles per fatality. About as unsafe as driving.
If we estimate that half of the fatal train collisions are suicides and decide to discount them for that reason, that gets us 150 million passenger miles per fatality. You're right, and honestly I was surprised - not as good as Waymo!
Of course, this should all come with the caveat that drivers travel greater distances than subway riders, so you could have the denominator be passenger hours instead, but then you run into the problem that subway riders spend more time commuting... so it all gets a bit muddy. Maybe the best way to do it is comparing the median car commuter v. subway commuter in the NYC metro.
The good thing is fatal train collisions are mostly a solved problem for new metros, since you can just build platform screen doors. The annoying thing is the MTA has said only a minority of NYC subway stations would be capable of having them retrofit. In my opinion it's probably worth it to install it in those minority of stations, but that doesn't seem to be the plan currently; instead they're doing these iron railings, which clearly have no deterrent power for suicides.
Goods are tangible items, like cars and milk and books and iPads. Services are activities other people do for you, like driving you around and cooking food and delivering Internet.
How is a Waymo a service? That's doing a lot of work here and it's wrong. When you take a Waymo you're taking a car from one place to another, the corporate structure or the technology (that's it's networked) or the app are irrelevant.
I agree that the analysis you propose in 2) would be ideal. Maybe different scenarios could be calculated, e.g. "scenario A: if Waymo diverts 20% of public transit users, 50% of taxis, and 30% of drivers, the reduction in serious injuries per mile would be X." I assume that Waymo would come out positive for all scenarios, and the AV critics would find the least positive scenario and write articles about how Waymos aren't that safe after all, but it's the most comprehensive analysis you could do considering our uncertainty about how much AVs will be used in the future.
Thanks so much for writing this piece! May I gently suggest that Zipper deserves a little more holding-to-account than your admirably generous response? His opening sentence:
“If a chorus of wide-eyed boosters and enthralled journalists are to be believed, self-driving cars from companies like Waymo, Tesla, and Zoox can bring about road safety nirvana — if only US regulators would get out of their way.”
“wide-eyed” and “enthralled” are ad-hominem attacks on the legitimacy of those Zipper disagrees with. He shouldn’t do that even if he has evidence to support his description — but to my reading he doesn’t supply any. Next, his close:
“Let’s give AV companies yet another benefit of doubt and assume that their technology proves so powerful that it produces a net reduction in crash deaths even with an increase in total car use. Still, that is not enough to justify government leaders prioritizing AVs as a road safety solution. The reason is quite simple. Good policymaking entails choosing the most cost-effective ways to address a public problem, in this case traffic deaths. Self-driving technology is only one of many tactics available to reduce crashes, and it is not at all clear that it offers the highest return on investment. To offer just a few alternative approaches: Cars could be outfitted with Intelligent Speed Assist — a far simpler technology that automatically limits the driver’s ability to exceed posted limits. Regulators could restrict the size of oversized SUVs and pickups that endanger everyone else on the road. Cities could build streets with features that are proven to reduce crashes, such as protected bike lanes, wider sidewalks, roundabout intersections and narrower travel lanes. States could legalize the installation of automatic traffic cameras that deter illegal driving. Bus and rail service could be expanded. Unlike autonomous vehicles, these strategies have been reliably shown to reduce crashes — and without AVs’ mitigating risk of expanding total car use.” [new lines removed for space]
Zipper considers “expanding total car use” a “risk. He falsely presents *allowing Waymo to operate* as *prioritizing AVs as a road safety solution*. He lists his preferred policy changes as options that “regulators” / “cities” / “states” all “could” do without acknowledging the obvious first obstacle that *voters don’t want them to implement those policies*. As Zipper is surely aware, normal people like buying giant SUVs that can go 100mph, and many jurisdictions have removed automatic traffic cameras at the behest of angry voters (BTW, this fight seems worth having). I mean, dude: putting in roundabouts just doesn’t conflict with letting Waymo serve its customers. The biggest conflict is in the last sentence: voters will overwhelmingly reject the proposition that because car trips can produce accidents, public policy should restrict voters’ access to car trips.
Zipper’s piece is rife with evidence of bad faith. I think the burden is on him to reply with reasons why we should treat his essay with respect rather than dismissal.
“the abundance movement may need to expand its ambitions: from making things affordable to rebuilding the communities and shared purposes that make abundance worth having”
This sounds good and righteous in tone, but when you actually think about it, isn’t it just everything bagel economic reform? I don’t really see how building more infrastructure and housing, and lowering the cost of electricity and healthcare, go hand-in-hand with building community and shared purpose, unless one makes an argument that those things come from not struggling to find housing and healthcare?
I don’t really want debates over zoning reform or Medicaid drug price negotiation to center around building shared purpose. I want them to center around lowering costs…
Yeah, I suspect you’re right that this is a dangerous idea for the movement that should be shut down. It’s not in the abundance movement’s field of competence to have an opinion about, say, whether people should be encouraged to go to church.
It's not everything-bagel reform to have a million justifications for the same thing. If something you're doing to reduce prices also builds community and shared purpose as a side effect, sell it as such.
Are there any analyses out there that compare Waymos to vehicles that drive in similar ways? (Eg, Ubers or taxis) Not sure if that data is available, but it might make people feel even better that the comparison is apples-to-apples. It seems plausible to me that Ubers/taxis have lower incident rates than people driving themselves.
however the difference is not even close to statistically significant; with only two fatal crashes observed, the 95% confidence interval for fatalities per 127M VMT is (0.24,7.22), so the CI for waymo's VMT per fatality is (17.5M,529M).
This whole piece is about making good apples to apples comparisons. Recognizing what your apple is, and clearly comparing it to other apples.
It suffers greatly for not defining its apple for the reader.
It focuses on Waymo. But it does not explain what Waymo is. It does not say that Wemo is not merely software, but rather a vehicle whose cost to deploy is doubled because of all the extra sensor technology that is installed. That would help most readers.
More importantlyLy, it does not explain that Waymo vehicles cannot drive just anywhere. They are limited to carefully mapped streets, rather than being simply deployable in some novel area. It does not address the weather conditions in which Waymos are taken off the road. It does not explain that sometimes a remote human might help the automated systems in a particular vehicle.
It fails to provide that one or two paragraphs for the reader that gives them understanding of the context and problem that issue.
Some of that would narrow the problem for the reader, making the particular claim more clear. Some would reassure the reader that the so-called Robo taxis are not entirely autonomous at all times, thus redefining the problem. The weather question cuts against the strength of the offered conclusion, so it is particularly important to include— if one is truly presenting the case in good faith
Perhaps the fundamental mistake is assuming that the reader knows as much about the problem as the author— in this case, what Waymo’s are and how they work. But very little of the country has access to a Waymos, so that’s a pretty poor assumption.
I agree with a lot in this article. But were my state and locality to follow its advice, I don’t think I would gain much access to waymos. I don’t think the company is itching to come to New Haven, let where I live three towns away. The article should make clear where Waymo is even interested in deploying.
This is kinda nonsense because the author is specifically responding to someone else (David Zipper at Bloomberg). The article's purpose is not to explain the background of robotaxis generally or Waymo specifically.
Why not leave it up to Waymo the company to explain where they are "even interested in deploying"?
Seems totally fair to evaluate the data that is available and point out, as this article's author does, that it is not especially ambiguous?
If this were a blog post 15-20 years ago, I would agree with you. But I think this is not a blog post.
I think it is an article or newsletter. Moreover, it is not just about refuting someone else. I think it aims to present a case, and the refutation mostly just provides an opportunity to do so.
I agree with a lot of what you say, but you’re arguing against a different point. Kelsey is saying Waymo as it is currently run provides safe automated taxi service.
There is another argument about when we’ll get true, unlimited self-driving tech that fully replaces human drivers. There I’m more sympathetic to you, in that it seems like we’re still years away (if not more).
Well, I agree that that is her intent. Only she is also calling for more approval for it, and when she starts to go there, she should be a bit more clear that such approval (or lack of bans) is not enough to expand the service they currently offer. They have to make big investments to serve in a geographic area before they can do so. No, it's not public funds—at least not yet—, but it is still a barrier.
Unfortunately, the article she is responding to doesn’t do any of that either!
I think there’s a meaningful question of whether the fact that Waymo’s are safer in the conditions they have operated in than human drivers tells us anything about whether they are safer in other conditions. But those other conditions aren’t ones they are considering operating in yet.
I don't buy the suggestion that any of us should make our arguments as poorly or misleadingly as our adversaries. Yeah, the other article is dumb. I accept that. I'm not even bothering to read it.
This article should be good. This article should contain a well-presented argument. And that should include a little bit of disclaimer/context. In fact, acknowledging the limitations of an argument improved credibility among the thoughtful set. And it prevents readers from dismissing one's argument for noticing that they've left something important out—which can cause readers to question what else is left out that they don't already know about.
I'm just saying make a good solid argument that stands on its own and give away credibility that it doesn't need to.
I agree with that. I think this article would have been better if it made clear the range of types of driving that we do have evidence about, and that this doesn't extend to all driving.
It sounds like you want to see a different analysis about cost/benefit of Waymo vs other transport options. That’s a worthwhile goal, but it doesn’t mean we shouldn’t understand the safety data on its own merits. On the safety data alone, Waymo appears to be significantly safer than human-driven cars. That is a useful thing to understand for people who primarily care about riding in a safe vehicle.
No, I explicitly said that my concerns could have been addressed with a just one or two paragraphs. "It fails to provide that one or two paragraphs for the reader that gives them understanding of the context and problem that issue."
Moreover, I am not suggesting in any way that any additional or alternative analysis is provided. Just the context and limitations of this one. You know, like any well presented analysis would include.
So, let me say it again: you, like Kelsey, leave out the fact that Waymo vehicles seem quite safe, provided they are only deployed in well mapped contexts for which they are specifically trained and in some subset of weather conditions. They do not even attempt other circumstances, so we cannot actually compare their safety there; it would like dividing by zero.
And because only a very small tiny part of the country is well mapped and people who primarily care about riding in safe vehicles particularly care about safety in the least safe conditions, those limitations make Waymos a good answer for but a fraction of the public who primarily care about riding in safe vehicles.
The "well mapped contexts" and "weather conditions" disclaimers you ask for are not needed, as they are irrelevant. The safety data that we have covers the conditions under which Waymo *actually drives*.
When discussing the safety of commercial air travel, nobody would consider necessary a disclaimer that "we don't know how safe the 747 is flying above active war zones or at 50,000ft of altitude. They simply don't travel there.
Waymo is being compared to human drivers. You analogy is not a comparison. It is as valid as pointing out that we don’t need to list calorie counts. No, I take that back. It’s less volcanic than that.
There is not a trained researcher in the world who would not say that you have to explain where the comparison is valid and where it is not. But more importantly, a trained professional communicator would take into account what they know and what the audience knows and try to write in a way that will be understood by the audience. No, you make smaller claims that will be misunderstood by a typical audience member, given their knowledge of the subject. That’s just bad communication.
Can we get an AMA on how Kelsey raises like 10 kids, runs a school, and writes two bangers a week? Because like I'm very impressed but also jealous.
Also David has a podcast with Wes Marshal. Wes is an engineer and definitely the less reactionary of the two. I know he is a busy professor, but maybe Wes could be talked into doing a mad libs?
I used to take Adderall but it's not recommended while breastfeeding.
Quite seriously, as a young adult I followed a lot of women who I admired for being employed full time in intellectually productive work, emotionally healthy with lives enviably full of exciting projects, while raising a large family. I was desperate to learn the secret of unlocking my true inner potential so I could do it all like them. And in every case, when I got close to them and learned more about how they did it, they had either a relative who lived close by and helped out a ton, or a nanny or several nannies. I want to be direct about this because I don't want young parents to think that they're just not trying hard enough: the families that 'have it all' are generally paying for a ton of help around the home. My family isn't an exception here though we're a bit unique because of the multiple couples coparenting thing (but we also have paid help on top of that). Raising a large family and running a household are hard work, and almost no one is effortlessly doing it on top of a full time job without a lot of help. If we want this to be an option for people who aren't rich, we need to subsidize families.
narrowly re the characterization of cruise — I think the report found that the execs were neglectful in their reporting duties immediately following the incident, not that they lied / intentionally deceived anyone
The full report has a lot of details, basically what everyone knew and did across the whole timeline of events.
This is a sample size of two over two point something years, so not very meaningful, but the other Bloomberg News story that sticks in my mind is a 2023 story stating that 94% of new jobs at S&P 100 companies the year after George Floyd went to persons of color.
Which was not true, and which understandably engendered a lot of white resentment, partly because the article seemed to celebrate this alleged event as companies fulfilling their BLM-related promises, rather than criticizing what would necessarily have been massive discrimination against white job applicants.
What's curious is that both that story and this Zipper story seem to feature very poor use of statistics. Which is not the sort of thing you would expect from a news organization founded by Michael Bloomberg.
As a meta point, what are we supposed to make of Zipper / Bloomberg? He’s a “senior fellow at the MIT Mobility Initiative” yet it seems like a high-schooler could reason about this better…
It might be a blind spot for a lot of the urban planning profession at the moment, just based on my conversations with grad school classmates. Many liberal urban planners (myself included) are mostly interested in design-based street safety, cycling/walking/public transit modal shift, and urban infill. Waymo does nothing for the first one, and potentially undermines the second and third if it makes taxis and exurban supercommuting more attractive.
In reality, the effects of Waymos on modal shift and housing demand are way too uncertain to matter for our decision-making now. And they can always be addressed with taxes and tolls. What we do know is that Waymos are very likely a good tool for addressing the public health crisis of traffic collisions in our country, so we should be embracing them wholeheartedly.
Might also be a blindspot for the Northeast especially since we have had no chance to ride Waymos yet (I recall there was a poll showing that people who have had contact with them are more likely to approve.)
“When measuring how dangerous human drivers are, Zipper used the first approach, citing data that says that for every 100,000 miles driven (by any vehicle), there are 1.26 fatalities.”
I think you meant to write “for every 100 million miles driven” here.
Kelsey, this is an outstanding piece of data journalism—and it perfectly illustrates something I've been working on in my forthcoming book.
You've identified a pattern I call "manufactured uncertainty"—where a mental structure filters data to create artificial ambiguity that justifies a predetermined conclusion.
The Bloomberg piece does exactly this:
Robust data ignored: 80-90% fewer serious injuries, 1/5 airbag deployment rate, 127+ million miles of statistics
Sparse data emphasized: 2 edge-case fatalities where Waymo wasn't clearly at fault
Manufactured conclusion: "We don't know if robotaxis are safer"
This isn't a math error. It's a mental structure problem. The author can't see the robust data because their framework filters it out.
I'm developing what I call the Tension Transformation Framework, which argues that tensions (like "new technology vs. safety concerns") are neutral—neither inherently good nor bad. What determines outcome is identity: whether you approach the tension as a victim seeking control, or an architect designing solutions.
The victim mental structure treats any uncertainty as grounds for paralysis or prohibition. Your reframe captures the architect response perfectly:
"We don't have perfect information, but we are not in a state of perfect ignorance either—and we're frankly much closer to the perfect information state than the perfect ignorance state."
This is exactly how we should approach AI, autonomous vehicles, and every other transformative technology. The question isn't "is there any uncertainty?" (there always is). The question is "what does the preponderance of evidence suggest, and how do we design for continuous learning?"
Utah's approach to both AI (their Learning Laboratory) and autonomous vehicles reflects this architect identity—sandbox experimentation with learning loops rather than prohibition based on manufactured uncertainty.
Thank you for modeling what rigorous, data-forward analysis looks like. We need more of this.
Chris Wasden is author of the forthcoming "How Chaos Makes America Work: Why Our Biggest Problems Become Our Greatest Advantages"
Zipper is doing the analytic equivalent of "Renee Good totally ran over that ICE agent." We can see the statistics! We know what they say! We know they say you're wrong!
Regarding the flipped numerator/denominator, which isn't the biggest error he makes but is pretty inexcusable: the lack of math knowledge or even basic competence among journalists (and what I sometimes see as kind of a perverse pride in it) is a problem.
I ride my bike every day in Los Angeles, and when I see a Waymo, I breathe a sigh of relief. I know it’s not going to do something crazy, dangerous, and weird like human drivers do all the time.
And there is a more subtle point where I think Zipper’s argument stands. As Zipper noted, a lot of Waymos’ driving has been at slower speeds and on surface streets (not highways). Furthermore, if Waymos behave fundamentally differently than human drivers, then we cannot extrapolate the human relationship between frequent injuries and rarer fatalities. Hypothetically Waymos could randomly bug out and drive into a brick wall for no reason every 90M miles of highway driving. Probably not but we don’t know yet. I would certainly be more confident if we had a few billion VMT on a more typical mix of roadways, but it will be years before we have that.
The article cited SF’s lack of progress toward its own vision zero goals without mentioning that SF — like most cities— has safer streets than the US average (high speed suburban arterials tend to be the most dangerous). As it points out, it’s a bit less than 100 million VMT per fatality US-wide. SF’s VMT in 2021 was about 7 billion while 26 fatalities occurred, equating to one fatality per more than 200 million miles driven!
I basically agree with this article but I will make two complaints.
1. We certainly have lots of data about crashes from Waymo, but there's lots of other data we don't have that's safety relevant. To start with, crashes are reported in detail but miles driven is just something you get from their advertising blog posts. We don't have breakdowns of highway vs city street, deadhead vs passenger, etc.
More significant for safely is that we don't know anything about disengagements and remote assistance. This doesn't impact how we should assess waymos safety record so far but is significant for how to assess it going forward.
2. I think footnote 3 is wrong because Waymo is a service not a car. If we were considering the safety record of a similar level AV that people owned I would agree that the impact on public transit is not the same as the safety record and belongs in a different conversation. But instead Waymo effectively is a public transit system. So you have to consider whether it's getting people out of Ubers significant safety win), their own car (big safety win but way more variance), or off of the bus (safety negative).
I'm unhappy with your framing of #2. If you're disappointed that AVs are diverting people from safer buses/trains, then you should improve your bus/train system to compete with AVs on convenience. If you can't get your shit together enough to do that, banning AVs (but not human-driven personal vehicles) in order to force people (but only those who can't drive themselves) into buses/trains anyway is absurdly cruel, unfair, and anti-abundance.
(This topic triggered me because I have serious driving anxiety and used to take an incredibly inconvenient train to work before Tesla FSD arrived in my life, diverted me from public transit, and in so doing saved my career and my sanity. If someone tried to take my FSD away on the basis that trains are safer, without either 5xing the frequency of said trains or stopping my coworkers from driving to work manually, I would go to nuclear war with them.)
I didn't say that we shouldn't allow Waymo or that we shouldn't consider convenience or any of the other things you're suggesting. What I said is that when thinking about the safety of the transportation system, Waymo is actually more like a public transit service rather than a private car, and so safety comparisons should include shifts from other systems, rather than just "is this car safer than that car".
In contrast Tesla FSD is fundamentally making individual private cars work differently, and so comparing primarily to other private cars is generally the right frame. (It's always the case that changes in the car market will induce some mode shift but that shouldn't be the primary analysis.) I think that analysis is not that great for Tesla but we don't really have good enough data to tell.
I don’t care if there is AI driving or lots of exceptionally safe remote drivers behind the scenes. Either way, I’m safer riding in a Waymo.
I’m confused why you’re arguing to compare Waymo to public transport? Surely the best comp is taxis or Lyft/Uber?
The most direct analogy is pretty clearly Uber/Lyft, so that's probably the best thing to do safety comparisons with.
Investigating them encroaching on (safer) public transportation options also seems worthwhile, but feels pretty apples-to-oranges to me.
Right, acknowledging that Waymo is safer than other AVs without also acknowledging there’s an army of people behind the scenes operating them is tantamount to lying.
First of all, that sounds like a gross exaggeration, but anyway—why? Every single automated Tesla has an entire human behind the scenes ready to operate it *from the steering wheel* at all times. No one is testing automated vehicles that have no possibility of human intervention under any conditions, because that would be an insane thing to try to do.
How is it a gross exaggeration? It is a fact that Waymo has as many employees as they have vehicles on the road. Your Tesla comparison is nonsensical in that respect. The promise of Tesla is that you just straight up don’t have to drive the car, it takes care of that. The crashes that happen with Teslas in FSD mode are almost certainly because the driver took that too literally and didn’t intervene when required. Waymos don’t work like that. Someone, or maybe several someones, we don’t know, is monitoring the vehicle at all times. Not acknowledging that is, like I said, tantamount to lying.
I’m not sure where you got this “fact,” but it might go back to an ancient report about Cruise, which counted all employees of any kind, not remote operators. We don’t actually seem to have these data publicly, but anecdotally, I’ve used Waymo about 40 times, mainly in 2022-2023, and had a remote intervention twice, once my fault and once the car’s. You’re writing as if the cars are all being remote operated in real time, but they literally cannot be remote operated at all, the latency doesn’t allow it.
It's not lying at all -- the safety record is the safety record, regardless of how it works. Also, other AVs also have lots of people behind the scenes -- opacity about this is universal in the industry (which is bad). Questions about disengagments and remote assistance are important for understanding the future, and for thinking about how this will relate to private cars, but it doesn't change the actual accident numbers.
The point is Kelsey is complaining about misleading comparisons and she’s making one.
I agree that it would be helpful to compare the safety of public transit to Waymo. For example, in 2023 88 people were killed when struck by an NYC subway train. You would also need to account for the number of assaults, sexual assaults, murders, and suicides (roughly 1/2 of the subway deaths) that occur occur on or by contact with a public transit vehicle vs Waymo. When you do that, I suspect public transit will be less safe on every one of those metrics.
Why in god's name would you consider suicides to be a relevant stat when considering passenger safety unless you were deliberately trying to juke the results?
Because it is very difficult to commit suicide by Waymo (it has never happened) and very easy to do so by train. As an individual ride you may not care, but it absolutely is relevant if you are trying to assess the harm of Waymo vs train on a population level (and consider the harm to drivers who are involuntarily used to facilitate a couple suicides per week on MTA. That must be devastating.). But even if you exclude suicide Waymo is still safer….and Public transit still subjects riders to assaults, sexual assaults, and murders that Waymo does not, in addition to fatalities and injuries from the mode of transport itself.
(edited bc I misread the NTD)
If we take the National Transit Database figure of 8 billion passenger miles in 2023 for the NYC subway, and 97 fatal train collisions + 5 murders in 2023 as our denominator, that gets us 78 million passenger miles per fatality. About as unsafe as driving.
If we estimate that half of the fatal train collisions are suicides and decide to discount them for that reason, that gets us 150 million passenger miles per fatality. You're right, and honestly I was surprised - not as good as Waymo!
Of course, this should all come with the caveat that drivers travel greater distances than subway riders, so you could have the denominator be passenger hours instead, but then you run into the problem that subway riders spend more time commuting... so it all gets a bit muddy. Maybe the best way to do it is comparing the median car commuter v. subway commuter in the NYC metro.
The good thing is fatal train collisions are mostly a solved problem for new metros, since you can just build platform screen doors. The annoying thing is the MTA has said only a minority of NYC subway stations would be capable of having them retrofit. In my opinion it's probably worth it to install it in those minority of stations, but that doesn't seem to be the plan currently; instead they're doing these iron railings, which clearly have no deterrent power for suicides.
Then private car ownership is also a service...
No! Buying groceries is not a service. Going to a restaurant is purchasing services. There's an important difference.
I appreciate that you think that but I'm looking for some definition of what makes one a service for you and the other not.
Goods are tangible items, like cars and milk and books and iPads. Services are activities other people do for you, like driving you around and cooking food and delivering Internet.
How is a Waymo a service? That's doing a lot of work here and it's wrong. When you take a Waymo you're taking a car from one place to another, the corporate structure or the technology (that's it's networked) or the app are irrelevant.
What? Waymo is a service like Uber or like a taxi.
I agree that the analysis you propose in 2) would be ideal. Maybe different scenarios could be calculated, e.g. "scenario A: if Waymo diverts 20% of public transit users, 50% of taxis, and 30% of drivers, the reduction in serious injuries per mile would be X." I assume that Waymo would come out positive for all scenarios, and the AV critics would find the least positive scenario and write articles about how Waymos aren't that safe after all, but it's the most comprehensive analysis you could do considering our uncertainty about how much AVs will be used in the future.
Thanks so much for writing this piece! May I gently suggest that Zipper deserves a little more holding-to-account than your admirably generous response? His opening sentence:
“If a chorus of wide-eyed boosters and enthralled journalists are to be believed, self-driving cars from companies like Waymo, Tesla, and Zoox can bring about road safety nirvana — if only US regulators would get out of their way.”
“wide-eyed” and “enthralled” are ad-hominem attacks on the legitimacy of those Zipper disagrees with. He shouldn’t do that even if he has evidence to support his description — but to my reading he doesn’t supply any. Next, his close:
“Let’s give AV companies yet another benefit of doubt and assume that their technology proves so powerful that it produces a net reduction in crash deaths even with an increase in total car use. Still, that is not enough to justify government leaders prioritizing AVs as a road safety solution. The reason is quite simple. Good policymaking entails choosing the most cost-effective ways to address a public problem, in this case traffic deaths. Self-driving technology is only one of many tactics available to reduce crashes, and it is not at all clear that it offers the highest return on investment. To offer just a few alternative approaches: Cars could be outfitted with Intelligent Speed Assist — a far simpler technology that automatically limits the driver’s ability to exceed posted limits. Regulators could restrict the size of oversized SUVs and pickups that endanger everyone else on the road. Cities could build streets with features that are proven to reduce crashes, such as protected bike lanes, wider sidewalks, roundabout intersections and narrower travel lanes. States could legalize the installation of automatic traffic cameras that deter illegal driving. Bus and rail service could be expanded. Unlike autonomous vehicles, these strategies have been reliably shown to reduce crashes — and without AVs’ mitigating risk of expanding total car use.” [new lines removed for space]
Zipper considers “expanding total car use” a “risk. He falsely presents *allowing Waymo to operate* as *prioritizing AVs as a road safety solution*. He lists his preferred policy changes as options that “regulators” / “cities” / “states” all “could” do without acknowledging the obvious first obstacle that *voters don’t want them to implement those policies*. As Zipper is surely aware, normal people like buying giant SUVs that can go 100mph, and many jurisdictions have removed automatic traffic cameras at the behest of angry voters (BTW, this fight seems worth having). I mean, dude: putting in roundabouts just doesn’t conflict with letting Waymo serve its customers. The biggest conflict is in the last sentence: voters will overwhelmingly reject the proposition that because car trips can produce accidents, public policy should restrict voters’ access to car trips.
Zipper’s piece is rife with evidence of bad faith. I think the burden is on him to reply with reasons why we should treat his essay with respect rather than dismissal.
“the abundance movement may need to expand its ambitions: from making things affordable to rebuilding the communities and shared purposes that make abundance worth having”
This sounds good and righteous in tone, but when you actually think about it, isn’t it just everything bagel economic reform? I don’t really see how building more infrastructure and housing, and lowering the cost of electricity and healthcare, go hand-in-hand with building community and shared purpose, unless one makes an argument that those things come from not struggling to find housing and healthcare?
I don’t really want debates over zoning reform or Medicaid drug price negotiation to center around building shared purpose. I want them to center around lowering costs…
Yeah, I suspect you’re right that this is a dangerous idea for the movement that should be shut down. It’s not in the abundance movement’s field of competence to have an opinion about, say, whether people should be encouraged to go to church.
The more reasons, the better.
It's not everything-bagel reform to have a million justifications for the same thing. If something you're doing to reduce prices also builds community and shared purpose as a side effect, sell it as such.
Are there any analyses out there that compare Waymos to vehicles that drive in similar ways? (Eg, Ubers or taxis) Not sure if that data is available, but it might make people feel even better that the comparison is apples-to-apples. It seems plausible to me that Ubers/taxis have lower incident rates than people driving themselves.
in 2021-2022 uber was involved in 0.87 fatalities per 100M vehicle miles traveled, or 114M miles per involvement with a fatal crash, vs waymo's 63.5M.
https://uber.app.box.com/s/lea3xzb70bp2wxe3k3dgk2ghcyr687x3?uclick_id=0d6be752-6614-44b4-ac9e-ad1566cb8f3a
however the difference is not even close to statistically significant; with only two fatal crashes observed, the 95% confidence interval for fatalities per 127M VMT is (0.24,7.22), so the CI for waymo's VMT per fatality is (17.5M,529M).
Here are a couple slightly older analyses:
https://web.archive.org/web/20231003025000/https://getcruise.com/news/blog/2023/human-ridehail-crash-rate-benchmark/
https://arxiv.org/pdf/2403.14648
This whole piece is about making good apples to apples comparisons. Recognizing what your apple is, and clearly comparing it to other apples.
It suffers greatly for not defining its apple for the reader.
It focuses on Waymo. But it does not explain what Waymo is. It does not say that Wemo is not merely software, but rather a vehicle whose cost to deploy is doubled because of all the extra sensor technology that is installed. That would help most readers.
More importantlyLy, it does not explain that Waymo vehicles cannot drive just anywhere. They are limited to carefully mapped streets, rather than being simply deployable in some novel area. It does not address the weather conditions in which Waymos are taken off the road. It does not explain that sometimes a remote human might help the automated systems in a particular vehicle.
It fails to provide that one or two paragraphs for the reader that gives them understanding of the context and problem that issue.
Some of that would narrow the problem for the reader, making the particular claim more clear. Some would reassure the reader that the so-called Robo taxis are not entirely autonomous at all times, thus redefining the problem. The weather question cuts against the strength of the offered conclusion, so it is particularly important to include— if one is truly presenting the case in good faith
Perhaps the fundamental mistake is assuming that the reader knows as much about the problem as the author— in this case, what Waymo’s are and how they work. But very little of the country has access to a Waymos, so that’s a pretty poor assumption.
I agree with a lot in this article. But were my state and locality to follow its advice, I don’t think I would gain much access to waymos. I don’t think the company is itching to come to New Haven, let where I live three towns away. The article should make clear where Waymo is even interested in deploying.
This is kinda nonsense because the author is specifically responding to someone else (David Zipper at Bloomberg). The article's purpose is not to explain the background of robotaxis generally or Waymo specifically.
Why not leave it up to Waymo the company to explain where they are "even interested in deploying"?
Seems totally fair to evaluate the data that is available and point out, as this article's author does, that it is not especially ambiguous?
If this were a blog post 15-20 years ago, I would agree with you. But I think this is not a blog post.
I think it is an article or newsletter. Moreover, it is not just about refuting someone else. I think it aims to present a case, and the refutation mostly just provides an opportunity to do so.
Yeah, it should stand on its own.
I agree with a lot of what you say, but you’re arguing against a different point. Kelsey is saying Waymo as it is currently run provides safe automated taxi service.
There is another argument about when we’ll get true, unlimited self-driving tech that fully replaces human drivers. There I’m more sympathetic to you, in that it seems like we’re still years away (if not more).
Well, I agree that that is her intent. Only she is also calling for more approval for it, and when she starts to go there, she should be a bit more clear that such approval (or lack of bans) is not enough to expand the service they currently offer. They have to make big investments to serve in a geographic area before they can do so. No, it's not public funds—at least not yet—, but it is still a barrier.
Unfortunately, the article she is responding to doesn’t do any of that either!
I think there’s a meaningful question of whether the fact that Waymo’s are safer in the conditions they have operated in than human drivers tells us anything about whether they are safer in other conditions. But those other conditions aren’t ones they are considering operating in yet.
I don't buy the suggestion that any of us should make our arguments as poorly or misleadingly as our adversaries. Yeah, the other article is dumb. I accept that. I'm not even bothering to read it.
This article should be good. This article should contain a well-presented argument. And that should include a little bit of disclaimer/context. In fact, acknowledging the limitations of an argument improved credibility among the thoughtful set. And it prevents readers from dismissing one's argument for noticing that they've left something important out—which can cause readers to question what else is left out that they don't already know about.
I'm just saying make a good solid argument that stands on its own and give away credibility that it doesn't need to.
I agree with that. I think this article would have been better if it made clear the range of types of driving that we do have evidence about, and that this doesn't extend to all driving.
It sounds like you want to see a different analysis about cost/benefit of Waymo vs other transport options. That’s a worthwhile goal, but it doesn’t mean we shouldn’t understand the safety data on its own merits. On the safety data alone, Waymo appears to be significantly safer than human-driven cars. That is a useful thing to understand for people who primarily care about riding in a safe vehicle.
No, I explicitly said that my concerns could have been addressed with a just one or two paragraphs. "It fails to provide that one or two paragraphs for the reader that gives them understanding of the context and problem that issue."
Moreover, I am not suggesting in any way that any additional or alternative analysis is provided. Just the context and limitations of this one. You know, like any well presented analysis would include.
So, let me say it again: you, like Kelsey, leave out the fact that Waymo vehicles seem quite safe, provided they are only deployed in well mapped contexts for which they are specifically trained and in some subset of weather conditions. They do not even attempt other circumstances, so we cannot actually compare their safety there; it would like dividing by zero.
And because only a very small tiny part of the country is well mapped and people who primarily care about riding in safe vehicles particularly care about safety in the least safe conditions, those limitations make Waymos a good answer for but a fraction of the public who primarily care about riding in safe vehicles.
The "well mapped contexts" and "weather conditions" disclaimers you ask for are not needed, as they are irrelevant. The safety data that we have covers the conditions under which Waymo *actually drives*.
When discussing the safety of commercial air travel, nobody would consider necessary a disclaimer that "we don't know how safe the 747 is flying above active war zones or at 50,000ft of altitude. They simply don't travel there.
That’s garbage.
Waymo is being compared to human drivers. You analogy is not a comparison. It is as valid as pointing out that we don’t need to list calorie counts. No, I take that back. It’s less volcanic than that.
There is not a trained researcher in the world who would not say that you have to explain where the comparison is valid and where it is not. But more importantly, a trained professional communicator would take into account what they know and what the audience knows and try to write in a way that will be understood by the audience. No, you make smaller claims that will be misunderstood by a typical audience member, given their knowledge of the subject. That’s just bad communication.
Can we get an AMA on how Kelsey raises like 10 kids, runs a school, and writes two bangers a week? Because like I'm very impressed but also jealous.
Also David has a podcast with Wes Marshal. Wes is an engineer and definitely the less reactionary of the two. I know he is a busy professor, but maybe Wes could be talked into doing a mad libs?
I assumed copious amounts of meth..
I used to take Adderall but it's not recommended while breastfeeding.
Quite seriously, as a young adult I followed a lot of women who I admired for being employed full time in intellectually productive work, emotionally healthy with lives enviably full of exciting projects, while raising a large family. I was desperate to learn the secret of unlocking my true inner potential so I could do it all like them. And in every case, when I got close to them and learned more about how they did it, they had either a relative who lived close by and helped out a ton, or a nanny or several nannies. I want to be direct about this because I don't want young parents to think that they're just not trying hard enough: the families that 'have it all' are generally paying for a ton of help around the home. My family isn't an exception here though we're a bit unique because of the multiple couples coparenting thing (but we also have paid help on top of that). Raising a large family and running a household are hard work, and almost no one is effortlessly doing it on top of a full time job without a lot of help. If we want this to be an option for people who aren't rich, we need to subsidize families.
Yes, my comment was a joke, but I appreciate you saying that
Time is limited. There's always trade offs that have to be made.
For example, i'm sure I could have a higher pain job.If I accepted more travel, but I want to be here with my family, so I don't
appreciate the article!
narrowly re the characterization of cruise — I think the report found that the execs were neglectful in their reporting duties immediately following the incident, not that they lied / intentionally deceived anyone
The full report has a lot of details, basically what everyone knew and did across the whole timeline of events.
Dan Luu does a good job highlighting bits for a still long but somewhat abbreviated read https://danluu.com/cruise-report/
This is a sample size of two over two point something years, so not very meaningful, but the other Bloomberg News story that sticks in my mind is a 2023 story stating that 94% of new jobs at S&P 100 companies the year after George Floyd went to persons of color.
https://www.bloomberg.com/graphics/2023-black-lives-matter-equal-opportunity-corporate-diversity/
Which was not true, and which understandably engendered a lot of white resentment, partly because the article seemed to celebrate this alleged event as companies fulfilling their BLM-related promises, rather than criticizing what would necessarily have been massive discrimination against white job applicants.
What actually happened was a lot of white people retired, so the net increase in employment at these companies was disproportionately nonwhite. (As the Daily Wire, of all places, pointed out. https://www.dailywire.com/news/bloomberg-flubs-data-for-bombshell-report-that-only-6-of-new-corporate-hires-are-white )
What's curious is that both that story and this Zipper story seem to feature very poor use of statistics. Which is not the sort of thing you would expect from a news organization founded by Michael Bloomberg.
As a meta point, what are we supposed to make of Zipper / Bloomberg? He’s a “senior fellow at the MIT Mobility Initiative” yet it seems like a high-schooler could reason about this better…
Does he just have huge anti-tech blinders on?
It might be a blind spot for a lot of the urban planning profession at the moment, just based on my conversations with grad school classmates. Many liberal urban planners (myself included) are mostly interested in design-based street safety, cycling/walking/public transit modal shift, and urban infill. Waymo does nothing for the first one, and potentially undermines the second and third if it makes taxis and exurban supercommuting more attractive.
In reality, the effects of Waymos on modal shift and housing demand are way too uncertain to matter for our decision-making now. And they can always be addressed with taxes and tolls. What we do know is that Waymos are very likely a good tool for addressing the public health crisis of traffic collisions in our country, so we should be embracing them wholeheartedly.
Might also be a blindspot for the Northeast especially since we have had no chance to ride Waymos yet (I recall there was a poll showing that people who have had contact with them are more likely to approve.)
“When measuring how dangerous human drivers are, Zipper used the first approach, citing data that says that for every 100,000 miles driven (by any vehicle), there are 1.26 fatalities.”
I think you meant to write “for every 100 million miles driven” here.
You're right and we've now published the correction. Thank you.
Came here to say this.
Kelsey, this is an outstanding piece of data journalism—and it perfectly illustrates something I've been working on in my forthcoming book.
You've identified a pattern I call "manufactured uncertainty"—where a mental structure filters data to create artificial ambiguity that justifies a predetermined conclusion.
The Bloomberg piece does exactly this:
Robust data ignored: 80-90% fewer serious injuries, 1/5 airbag deployment rate, 127+ million miles of statistics
Sparse data emphasized: 2 edge-case fatalities where Waymo wasn't clearly at fault
Manufactured conclusion: "We don't know if robotaxis are safer"
This isn't a math error. It's a mental structure problem. The author can't see the robust data because their framework filters it out.
I'm developing what I call the Tension Transformation Framework, which argues that tensions (like "new technology vs. safety concerns") are neutral—neither inherently good nor bad. What determines outcome is identity: whether you approach the tension as a victim seeking control, or an architect designing solutions.
The victim mental structure treats any uncertainty as grounds for paralysis or prohibition. Your reframe captures the architect response perfectly:
"We don't have perfect information, but we are not in a state of perfect ignorance either—and we're frankly much closer to the perfect information state than the perfect ignorance state."
This is exactly how we should approach AI, autonomous vehicles, and every other transformative technology. The question isn't "is there any uncertainty?" (there always is). The question is "what does the preponderance of evidence suggest, and how do we design for continuous learning?"
Utah's approach to both AI (their Learning Laboratory) and autonomous vehicles reflects this architect identity—sandbox experimentation with learning loops rather than prohibition based on manufactured uncertainty.
Thank you for modeling what rigorous, data-forward analysis looks like. We need more of this.
Chris Wasden is author of the forthcoming "How Chaos Makes America Work: Why Our Biggest Problems Become Our Greatest Advantages"
Humans definitely drive into flood waters and through active crimes scenes.
Zipper is doing the analytic equivalent of "Renee Good totally ran over that ICE agent." We can see the statistics! We know what they say! We know they say you're wrong!
Regarding the flipped numerator/denominator, which isn't the biggest error he makes but is pretty inexcusable: the lack of math knowledge or even basic competence among journalists (and what I sometimes see as kind of a perverse pride in it) is a problem.
I ride my bike every day in Los Angeles, and when I see a Waymo, I breathe a sigh of relief. I know it’s not going to do something crazy, dangerous, and weird like human drivers do all the time.
And there is a more subtle point where I think Zipper’s argument stands. As Zipper noted, a lot of Waymos’ driving has been at slower speeds and on surface streets (not highways). Furthermore, if Waymos behave fundamentally differently than human drivers, then we cannot extrapolate the human relationship between frequent injuries and rarer fatalities. Hypothetically Waymos could randomly bug out and drive into a brick wall for no reason every 90M miles of highway driving. Probably not but we don’t know yet. I would certainly be more confident if we had a few billion VMT on a more typical mix of roadways, but it will be years before we have that.
The article cited SF’s lack of progress toward its own vision zero goals without mentioning that SF — like most cities— has safer streets than the US average (high speed suburban arterials tend to be the most dangerous). As it points out, it’s a bit less than 100 million VMT per fatality US-wide. SF’s VMT in 2021 was about 7 billion while 26 fatalities occurred, equating to one fatality per more than 200 million miles driven!
An AI-assisted search seems to confirm SF’s per-mile fatality rate is about one third the national average: https://kagi.com/assistant/226d41bc-79db-4e60-bf4e-27230e5b28f0?branch=00000000-0000-4000-0000-000000000000.