Are Waymos safe?
Yes — but whether they’re making roads safer is a much more complicated question.
(Note that this is an article in which I go way outside my depth and almost certainly will get myself in trouble. All mistakes are my own).
I like Waymos. I recently went to Los Angeles, and took them everywhere (in part because taking public transit where I needed to go would have taken forever). My media diet has fed me a steady stream of strident claims about the safety record of Waymos, which go as far as claiming they are saving lives. But I realized recently that I hadn’t seen anyone actually dig into these claims, and when I spent just a few minutes digging into it, it seemed like a bunch of the safety claims were overstated, and the data present a much more complicated picture.
So, I’m going to try to disentangle a bunch of concepts that Waymo’s own safety data tangle together, and try to develop answers to more specific questions on Waymo safety. These answers are based on the state of Waymos today, in May 2026. They might get much better in the coming years, so while I think Waymo’s safety stats overstate their case, I feel optimistic about the technology.
What does Waymo say about its safety track record? It has a notably transparent safety report card, which indicates that as of writing (May 25th, 2026), the following: “Compared to an average human driver over the same distance in our operating cities, the Waymo Driver had:”
A 92% reduction in serious injury or worse crashes
83% fewer airbag deployments in any vehicle crashes (interpreted in a positive way — e.g. that the crashes weren’t severe enough to deploy airbags, not that airbags failed to deploy)
82% fewer injury-causing crashes.
92% fewer pedestrian crashes with injuries
85% fewer cyclist crashes with injuries
81% fewer motorcycle crashes with injuries
These are stunning accomplishments in traffic safety, taken at face value! Waymos on the street are preventing injuries! And Waymo has done an admirable job making these data transparent and analyzable by anyone. But these statistics shouldn’t be taken at face value, because they aren’t actually comparing to an accurate counterfactual.
The most important question to ask is “what is a Waymo ride replacing?”
Waymo’s safety data aren’t answering the question “how many accidents are Waymos preventing by being on the road right now.” Instead, Waymo is answering the question “if Waymos were randomly substituted in for cars in the same counties and sorta the same roads, driving at any given time of day, how many fewer injury-causing accidents would this cause?” While these questions sound similar, they are actually very different. To not bury the lede, here’s what I think:
If Waymos were randomly substituted in right now for cars driving on city streets in a given moment, would roads be safer?
Yes! This is the hypothesis their safety data mostly test directly, and they’ve produced good evidence for it.
Is Waymo’s autonomous driving technology making their cars safer?
I lean yes, but the data aren’t clear yet, and a significant portion of the safety improvement from Waymos isn’t due to the autonomous vehicle (AV) tech, but other factors like routing, car type, and a ride’s time of day.
Am I safer if I take a Waymo right now vs...
Biking/walking?
Definitely, but I’d also be safer if I drove myself, assuming I’m walking on roads. And if I take a Waymo, I’m increasing the risk for other people.
Driving myself?
Yes for a truly random person, but if I’m sober, attentive, and driving a modern car, the evidence is much weaker. For a sober person driving a modern car, the safety benefit of riding a Waymo is probably slight.
An Uber/Lyft?
Maybe, but the benefit is potentially slight, especially if my driver has a modern car, pays attention while driving, and is sober (which is likely true of most Ubers/Lyfts).
Public transit?
Definitely not.
Staying home?
Definitely not.
Are Waymos making roads safer right now?
For Waymo riders
Probably, but it might depend on what kind of car they’d be in otherwise, and if their driver would be sober.
For other drivers/riders in other vehicles
Probably!
Vulnerable road users (pedestrians, cyclists, motorcyclists, etc.)
It’s too early to say, but potentially.
Will Waymos make cities safer in the longrun?
I think it is too early to say, and is heavily dependent on the counterfactual and future improvements in the technology. Will cities underinvest in public transit because Waymos and other AVs are successful? If so, then there is a fairly large possibility that Waymos make roads less safe. But if Waymos only ever replace Ubers and Lyfts, and generate no new demand for rides, then yes, they likely will.
What causes car accidents and who gets injured in them?
Before we can ask if Waymos are safer than drivers, we have to ask, why do car accidents happen at all? And who is injured in them? The National Highway Traffic Safety Administration provides data on injury rates by vehicle type. In 2023, these figures were:
Passenger vehicle occupants
92% of injuries
59% of fatalities
Pedestrians
2.8% of injuries
18% of fatalities
Motorcyclists
3.4% of injuries
15% of fatalities
Bicyclists
2% of inuries
3% of fatalities
Other (large trucks, buses, etc)
~2% of injuries
~5% of fatalities
So, most car accidents injure or kill car riders. Although pedestrians, motorcyclists, and cyclists make up a small share of total injuries, they are more likely to be killed if they are in a crash, because they lack the protection of a vehicle. By being in a car, you’re both more likely to cause injury, and you’re also safer than when you are walking or biking.
Of course, cars and walking aren’t the only ways to get around. While the exact relative safety of transit is somewhat unclear, it’s estimated that buses are something like 60x-70x safer per mile than driving, and trains are 20x-30x safer (though this is contested).
Comparing walking to driving is more complicated — on a per mile basis, driving is something like 4x-8x safer than walking, and 10x-20x safer than biking. But per trip or per hour, walking is often similarly safe.
These stats are contested and not straightforward to estimate, in part due to the unclear substitutability between modes of transportation. But it seems safe to say that roughly, per mile travelled, transit is significantly safer than driving and walking is more dangerous for any given distance, but you’ll probably walk many fewer miles than you’ll drive.
Alcohol
Inebriation plays a major role in car accidents. Of the 40,901 traffic fatalities in the US in 2023, 12,429 involved at least one drunk driver. This has been trending down (from 48% in 1982 to 30% now). Of these 12,429 deaths, around 60% were the driver themselves, and 40% were someone else.
Pedestrian deaths involving alcohol are much more likely to involve a drunk pedestrian than a drunk driver: in approximately 24% of pedestrian deaths, only the pedestrian was impaired, while in only 10% was only the driver impaired. In 6% both were impaired, and in 60%, neither were impaired. (NHTSA 2023 Pedestrians fact sheet; NSC Injury Facts pedestrian breakdown)
Time of day
The peak window for car accidents is between 4pm and 8pm on weekdays (during rush hour). Fatal crashes, however, primarily happen in the evening and at night — there is a large peak between 6pm and 9pm, and another smaller peak between midnight and 3am on weekends. The per-mile fatal rate is 3x higher at night than during the day. 75% of pedestrian fatalities occur at night — a fact that matters for the Waymo discussion, since LiDAR can see in darkness when human eyes can’t.
What do we know about Waymo safety track record?
Waymo has a truly excellent and transparent safety report card. This report says, as of writing in May 2026, that Waymo achieves:
A 92% reduction in serious injury or worse crashes
83% fewer airbag deployments in any vehicle crashes (interpreted in a positive way — e.g. that the crashes weren’t severe enough to deploy airbags, not that airbags failed to deploy)
82% fewer injury-causing crashes.
92% fewer pedestrian crashes with injuries
85% fewer cyclist crashes with injuries
81% fewer motorcycle crashes with injuries
Notably, reductions in fatalities aren’t included here. This is good, and will be discussed below. Waymo’s methodology for these studies is as follows:
They gathered neighborhood-level data on accidents for roughly (but not exactly) the areas that Waymos drive on in the cities they operate in (Waymos often don’t drive on highways or high speed roads, etc), with spatial reweighting to match Waymo’s actual operating areas
They looked at the accident and injury rates on those roads.
They compared those rates to Waymo rates.
The question that these data answer is “if you replaced random cars on these roads with Waymos, how much safer would these roads be?” They aren’t directly measuring the impact of the AV technology — they are just asking “Waymo vs a random car — which is safer?” This is important because Waymos are not just replacing random cars, but we’ll get to this later.
Waymo doesn’t report their impact on fatalities. This is good — it would be statistically inappropriate to do so. Waymos have driven around 200 million rider-only miles total. In the US, there are currently around 1.20-1.26 deaths per hundred million miles driven. By the rule of three, this means that Waymos would have to have driven around 300 million miles without a fatality to claim with confidence that they are causing fatalities at approximately the same rate, and more to claim a reduction. AVs likely need hundreds of millions to billions of miles to statistically demonstrate fatal-rate reduction depending on the magnitude claimed.
And have Waymos not had a fatality? That’s unclear — there have been two crashes Waymos were involved in where someone died (SF, January 2025; Tempe, October 2025). In both cases, the Waymo was reportedly not at fault. The comparison to the human baseline depends on what you’re measuring. The national 1.20-1.26 deaths per 100M miles figure counts every road death divided by total miles travelled by all cars. The closer comparison is the fatal-crash rate per vehicle-mile, which is roughly 1.0 per 100M miles nationally. By that measure, Waymo’s 2 fatal involvements in 200M miles puts it at roughly parity with humans — neither clearly better nor worse, with very wide error bars. By an at-fault comparison (Waymo zero, humans roughly 0.5-0.7 per 100M miles), Waymo looks substantially better, but the sample is too small for confidence on any of these statistics.
What causes Waymo’s apparent safety benefits?
From Waymo’s data, it’s clear that Waymos are safer than humans if they replaced random cars. Discourse on Waymos has conflated the question “are Waymos safer than the alternative?” with the question “is Waymo’s AV technology making roads safer?” If Waymo got all its safety benefits without AV technology (say by using remote drivers at all times or by using human drivers, like Ubers and Lyfts), the data would speak only to the first question, and not the second. But many futurists are using the first question as a proxy for the potentially far more interesting second question.
Disentangling these is actually really hard. Unfortunately, Waymo’s data do not actually isolate the AV technology in its analysis. Instead, their data include several confounding factors. These confounding factors, like car type (Waymos are modern cars with modern safety features), make Waymos genuinely safer — but an Uber driver using a modern car would also be safer than the average driver. Waymo might be reducing accidents, but the interesting question is how much that safety gain is caused by the AV technology.
Car type
Waymos are all Jaguar I-PACEs, a modern car with many safety features, like side airbags and pedestrian-protection features that reduce pedestrian fatalities and head injuries.
Waymos seem to have disabled the “Advanced Driver Assistance Systems” that I-PACEs ship with (sensors and cameras that modern cars use to keep cars in lanes, brake to prevent rear-end crashes, etc) and replaced them with Waymo technology. Most newer cars also have these. But many older cars on the road don’t have them. I can’t find great statistics, but around half of cars on US roads don’t seem to have them.
So, when a Waymo is replacing a random car, there is probably around a 50% chance that car wouldn’t have many modern safety features. Replacing those random cars with any modern car with ADAS would likely increase the safety of roads.
Waymo does not control for this in their comparisons — I think this could be a fairly substantial contributor to Waymo’s safety statistics, given how much safer modern cars are than older ones. For example, forward collision warning and automatic emergency braking together reduce front-to-rear injury-causing crashes by around 53%. So it is unclear how much of the effect is due to modern cars with ADAS being safer, versus due to the AV technology. If Waymo gave random drivers Jaguar I-PACEs without the AV technology, there would still be a significant safety benefit.
Driving time of day
Time of day significantly impacts the likelihood of injurious or fatal crashes, especially for pedestrians. Waymo doesn’t control for this in their data (and acknowledges it as a problem, to their credit), though we might expect that their peak hours somewhat align with human drivers.
Route choice
Waymos primarily drive on city roads with lower speed limits. While exact routing information isn’t published, it’s also believed that Waymos avoid particularly dangerous or complicated roads when possible. A human driver might not make either of these choices, but if they did, they’d probably be a bit safer. Most serious injuries occur in higher speed accidents.
Waymo controls for this a bit in their comparisons with the “spatial reweighting” of crash data, but it is inexact, and is using county-level instead of street-level data, which would be necessary for a perfect comparison. A human driver making the same route choices as a Waymo might get some safety upside.
AV technology
I don’t think these confounding factors are enough to offset all of Waymo’s safety improvements. They likely significantly diminish them, but the AV technology probably is making Waymos safer! Of course, there are other confounding factors, like driver sobriety, which aren’t captured here — if a Waymo replaces a drunk driver, it is much more likely to reduce the likelihood of a crash than if it replaces a sober one. But the AV technology can be credited for some safety gains.
What rides are Waymos replacing or creating?
Even if Waymos are significantly safer than at least many of the drivers or cars they replace, they still risk making roads more dangerous by increasing the miles driven. For example, if a Waymo causes 1/10th the serious injuries per mile of traditional cars, but Waymos create 11x additional miles of driving, then you’ve increased serious injuries in expectation! If this new driving time is created exclusively by replacing transit, which is significantly safer than even the safest driving, then you’re not even getting the gain of helping more people get more places!
There is a decent probability that Waymos will create additional driving miles that otherwise would not occur. For example, Waymos might cause some people who would not otherwise have gone out to take a trip. Or, they’ll replace walking, biking, or taking public transit. This might increase road safety for passengers (e.g. when replacing walking or biking) or decrease it (when replacing public transit, or causing a new trip to occur that otherwise wouldn’t have).
So what evidence do we have about Waymos increasing driving miles? We can turn to the closest analogy we have — the introduction of ridesharing.
Given that ridesharing wasn’t exactly a novel invention (taxis already existed when they were introduced — Ubers are just more convenient), and Waymos similarly aren’t exactly novel (Ubers already exist, Waymos just are more pleasant in many ways), we might reason that if rideshares didn’t create new trips, then Waymos might not create new trips as well. But the evidence actually points the other way: the introduction of ridesharing significantly increased the number of miles driven on roads.
As a high-level example, for trips served by ridesharing, total miles driven more than doubled compared to what those trips would have required in other modes. The figures by metro area:
Boston: +157%
Chicago: +97%
New York: +114%
San Francisco: +134%
Suburban California: +118%
Because Ubers and Lyfts replace other modes of transportation, create new trips, and drive empty between rides, it takes roughly twice as many vehicle miles as the counterfactual would have required to deliver people via rideshare.
Replacing other modes of transportation
When ridesharing was introduced, it seems like a significant portion of rides replaced were other modes of transit. Over 50% of ridesharing rides replaced non-car transit (e.g. walking or public transit), thus making roads more dangerous. While this effect is difficult to study, it does seem like there is direct evidence that the introduction of ridesharing, despite the existence of taxis, switched many people from other modes of transit to driving.
Whether or not these switches are good or bad for road safety depends on the switch and whose perspective you view it from. A walker/biker taking an Uber/Lyft instead of walking is going to be safer in many cities. But because the additional drive is created, other pedestrians and cars become less safe. So on net, these switches likely make roads significantly less safe. Because public transit is the most substituted mode of transit, the degree of danger is even greater.
Additional trips induced
Around 5.8% of rideshare trips were entirely induced. Basically — by introducing rideshare, and increasing the convenience of leaving the home, more people took additional trips. This is a pure increase in risk — previously, a car wouldn’t be on the road, and now one will be.
Waymos might just replace the induced trips caused by Ubers/Lyfts, but presumably, if their costs come down as expected by the industry, they’ll induce more additional trips (indeed, intentionally low prices might be why Ubers/Lyfts caused more trips than taxis served).
Note that this type of risk isn’t included in Waymo’s safety stats at all (and it’s not clear how it could be given that Waymo has no access to the counterfactual). If Waymo achieves 82% fewer injury-causing crashes per mile, but Waymos are responsible for 50% more total miles than the rides they replace (due to deadheading and induced demand), the net injury reduction shrinks from 82% to 73%. At 5x the miles, the reduction shrinks to just 10%.
Deadheading
Studies have suggested that somewhere between 40% and 50% of Uber/Lyft miles driven are without a passenger. For example, the Uber driver has to drive from their home to the center of the city, or between dropping off one passenger and picking up another. These are, depending on the passenger’s counterfactual transit mode, additional miles cars will drive (and additional risk). For example, if a passenger would otherwise drive themselves, this means that an additional 80%-100% of that distance will be driven by the Uber driver.
Waymos appear to have similar deadheading rates to Ubers/Lyfts. We might expect some improvements in the future due to optimization and increases in Waymo density, but some deadheading will likely always exist. This means that any driver who gives up their own car to ride in a Waymo is increasing the miles driven on the road. So, my expectation is that if Waymos just replace Uber/Lyft rides, then we wouldn’t see an increase in miles driven due to deadheading, but if Waymos replace trips driven with personal cars (or public transit, biking, etc), then we will see an increase.
Some Waymo deadheading time is idling — the cars will pull into safe spots on the side of the road and wait. But it’s unclear how much time is spent idling.
Waymo’s safety statistics are per mile — this means that for any ride where I replace my own car with a Waymo, given that we should expect a Waymo to drive empty for an additional 30%-100%, I should basically discount the per-mile safety advantage proportionally. A 50% deadheading rate roughly cuts the headline 82% reduction to a 73% net reduction in expected injuries from my trip.
Will more change in transport patterns happen with Waymo?
Waymo might entirely replace Uber/Lyft rides — if this is the case, then total miles driven wouldn’t increase (assuming Waymos don’t deadhead more than Uber/Lyft, which doesn’t seem evident), and any safety change would come purely from per-mile differences between Waymos and rideshare drivers. However, this isn’t what these companies seem to expect. Uber recently said in an investor presentation “AVs (as a new form of supply) will expand — not shrink — our total addressable market.”
What modes of transit will Waymos replace in the future?
Right now, Waymos are probably mostly replacing Ubers and Lyfts, and sometimes replacing public transit. But in the future, if cities choose to not invest in public transit due to cheap and abundant Waymos, the introduction of AVs might mean less public transit. If this occurs, cities will almost certainly be less safe unless Waymos get many times more safe than they currently are. Replacing public transit with AVs will probably make roads less safe by default.
Conclusion
Right now, I think the right way to think about getting in a Waymo is something like “I’m ordering an Uber, but I can click a button that basically guarantees that my driver is sober and attentive, and the car will be one with modern safety features.” That’s a button I would click every time, and a feature I would be willing to pay for in an Uber! It’s a real benefit! But claims about Waymos actually making roads safer right now, and certainly about Waymos saving lives are harder to defend from current data. Sometimes I call an Uber, and the driver is not attentive or a good driver, and my safety is probably at a bit more risk. But usually, the driver is sober and attentive. So, if my counterfactual to the Waymo is calling an Uber, I’m probably not getting much road safety benefit. But if my counterfactual would have been taking public transit, I’m probably increasing the risk to myself and others significantly.
For Waymo’s safety case to be fully convincing, we’d want a few things: independent data access (right now, every credible analysis uses Waymo-provided data), hour-by-hour mileage breakdowns to control for time of day, mode-substitution data for Waymo trips specifically, and another few hundred million miles of fatal-crash data. Until then, the best we can say is that Waymos might be modestly safer than realistic alternatives per mile, but the degree of net road-safety effects from Waymos AV technology are as of yet unknown.