Waymo Crash Rate Study: 68% Lower, With Real Caveats

The Waymo crash rate study that everyone quoted in July 2026 reported a single number: 68% fewer crashes per mile than human drivers. That number is real, published by the Insurance Institute for Highway Safety on July 23, 2026, and it is the best data anyone has.

It is also an average of four cities that disagree with each other. In Phoenix the crash rate was 76% lower. In San Francisco, 35% lower. In Austin, it was 4% higher than human drivers. Same company, same software, four different answers.

If you only remember one thing from this article, remember that the headline number is a mean, and the spread is the story.

Key Takeaways

• IIHS reported Waymo’s total crash rate 68% lower per vehicle-mile than human drivers, single-vehicle crashes 85% lower, and injury or fatal crashes 81% lower, for 2021–2024.
• The city-level results ranged from 76% lower in Phoenix to 4% higher in Austin, where IIHS noted the sample was “relatively small.”
• This is an observational comparison of two groups driving under different conditions, not a randomised trial. Waymo did not operate on interstate highways during the study period.
• The comparison rests on roughly 50 million Waymo driverless miles against about 222 billion human-driven miles in the same places and period.
• IIHS flagged that about half of all human crashes and a third of human injury crashes go unreported, and that AV companies over-report by comparison, a bias that cuts against Waymo in the raw counts and in favour of it once corrected.

What the study actually measured

An observational crash-rate comparison divides crashes by miles driven for two populations and compares the quotients. It does not randomly assign anyone to either group, which means every difference between the groups is baked into the result.

The IIHS release, published July 23, 2026 and reported in the paper “Rise of the machines: crash experiences of highly automated vehicles and human drivers,” states that “Waymo’s driverless vehicles deployed in San Francisco, Phoenix, Los Angeles and Austin were involved in 68% fewer crashes than human drivers,” and that “Waymo vehicles were involved in 85% fewer single-vehicle crashes and 81% fewer injury crashes per VMT than human drivers.” Eric Teoh, IIHS director of statistical services, was lead author.

The denominators: “about 50 million miles in driverless operation over the study period” for Waymo, against “about 222 billion miles by human drivers over the same period and in the same locations.” The study period ran 2021 to 2024.

The city spread nobody put in a headline

Here is the sentence that should have led the coverage. IIHS reported the reduction as “76% lower in Phoenix, 35% lower in San Francisco, 71% lower in Los Angeles, and 4% higher in Austin, where the sample size was relatively small.”

A 4% higher crash rate in Austin is not a scandal, and with a small sample it may well be noise. But it demolishes the idea that “68% safer” is a property of the technology. It is a property of the technology in a place, under that place’s weather, street geometry, pedestrian density and driving culture.

Phoenix is wide, flat, dry, grid-planned and has been Waymo’s home market the longest. San Francisco is hilly, foggy, dense, and full of double-parked delivery vehicles. The 41-point gap between those two cities is roughly the size of the whole headline effect.

What you were told	What the data says	Why it matters
”68% fewer crashes”	True as a four-city average, 2021–2024	It is a mean, not a constant
Safer everywhere	76% lower in Phoenix, 4% higher in Austin	Performance is city-specific
Compared like for like	Waymo drove no interstate highways in the period	Different road mix entirely
A study proving causation	Observational comparison, no randomisation	Confounders are not removed
Based on huge AV mileage	~50 million Waymo miles vs ~222 billion human miles	AV denominator is still small
Clean crash counts	~25% of reported crashes removed in cleaning	Definitions did heavy lifting

The reporting problem, in both directions

This is the part of the study that rewards slow reading, because the bias runs both ways and IIHS says so.

Human drivers under-report. IIHS states that “around half of all crashes and a third of injury crashes go unreported,” and that drivers are “generally only required to report crashes that cause an injury or result in more than $1,000 in property damage to the police, though the damage threshold varies by state.” A fender-bender settled in a parking lot never enters the data.

Autonomous vehicle operators over-report. “Until an amendment to the SGO that took effect in 2025, self-driving vehicle companies were required to log minor incidents as crashes. For example, a vehicle scraping its undercarriage as it turned into a parking lot would count as a crash.” IIHS also notes that “companies are still more likely than regular people to report all qualifying crashes due to the business risks of failing to comply,” and that “driverless vehicles require expensive sensors that can be damaged by minor impacts, making their crashes more likely to exceed the damage threshold.”

So the researchers had to clean the data by hand, deciding which incidents “a reasonable person would have reported” to police. That process “eliminated about a quarter of the reported crashes.” Of 736 public-road crashes in the raw set, only 22% qualified as police-reportable. The surviving counts: Waymo 64, Cruise 50, Zoox 10, others 10.

A quarter of the data removed by human judgement is not a flaw; it is the only way to make the comparison at all. It is also a dependency you should know about before you treat 68% as a physical constant.

What the limits mean for a parent deciding something real

Four limits matter if you are deciding whether to put your family in one of these cars.

No highways. Waymo’s vehicles “did not operate on interstate highways during the study period.” Highway crashes are a different distribution, higher speeds, higher severity. The study says nothing about them.

Geofenced, mapped territory. Waymo operates inside a defined service area it has surveyed intensively. Human drivers in the comparison drove everywhere, including roads nobody mapped at centimetre resolution.

Unoccupied vehicles. Wikipedia’s summary of the IIHS data notes that “at times the AVs had no people on board to be injured,” which mechanically lowers an injury-crash rate. How much, nobody has quantified.

Fault is not the same as involvement. In crashes serious enough for a police report, Waymo was judged at fault in 8% of cases, with another 8% unclear. Most crashes involving a Waymo were caused by a human hitting it. That is a point in Waymo’s favour and it is also why “crashes per mile” and “crashes caused per mile” are different metrics.

And the caveat from IIHS itself, which deserves to be the last word on the study: “The present data collection system isn’t good enough to allow continuous monitoring of a large-scale expansion.” IIHS also pointed out that “Waymo provides driverless VMT information voluntarily, but other companies do not,” which makes the same analysis impossible for most of the industry.

Waymo’s own numbers, and why they are a different kind of claim

On September 24, 2026, Waymo published its own safety data: “over 270 million miles of fully autonomous operations through the end of June 2026,” with “841 fewer injury-causing crashes, an 82% reduction compared to human drivers,” “95% fewer (20x fewer) serious injury or worse crashes,” and reductions of 93% for pedestrians, 86% for cyclists and 82% for motorcyclists, across Atlanta, Austin, Los Angeles, Phoenix and San Francisco.

Those are bigger numbers over a bigger mileage base, and they point in the same direction as the IIHS result. They are also a company blog post about its own product, using its own benchmark construction, and the 841-injuries figure rests on an explicit assumption of “at least one injured person per crash.”

None of that makes it wrong. It makes it a different evidentiary class. When the independent analysis and the company’s analysis agree on direction, that is meaningful. When you need a precise number, use the one you can audit.

For the broader context on how much we actually know about driver automation, the IIHS research base is sobering: Cicchino (2025) “did not find any crash-reduction advantage for vehicles equipped with partial driving automation” over comparable models with ordinary crash-avoidance features. Full driverless operation and partial automation are different animals, and conflating them is the most common error in this whole subject. We unpack that in the driving automation levels explained for parents.

What to do with this at home

Quote the range, not the mean

If your teenager repeats “68% safer,” add the Austin number. Not to undercut the finding, but because a kid who learns that a single statistic can hide a 41-point spread has learned something they will use on every statistic for the rest of their life.

Ask what the two groups were doing differently

This is the whole skill of reading observational research. Waymo drove mapped city streets at moderate speed in good weather in four metros. The human comparison group drove everything. Any difference in outcome includes that difference in conditions.

Check whether the measurement changed mid-study

The SGO amendment in 2025 altered what counted as a crash. Whenever a trend line crosses a definition change, be suspicious of the slope. This happens constantly in education data, crime data and health data too.

Look at who collected the numbers

IIHS is funded by auto insurers, which gives it a real interest in accurate crash data and no particular interest in flattering Waymo. Waymo’s blog is Waymo. Both can be useful; they are not interchangeable.

What not to do

Do not treat the study as a verdict on whether your family should ride in one. It measures crash rates in four cities over four years on non-highway roads. Your actual question, is this safe for my 15-year-old on a Tuesday night in my city, is narrower than anything this research answers, and depends on facts we cover in whether a teen can ride alone in a driverless car.

What to Watch For Over the Next 3 Months

• Week 4: Watch whether any operator besides Waymo starts publishing vehicle-miles-travelled data voluntarily. IIHS named that gap specifically. Without a denominator, no crash rate can be computed for Zoox, Tesla or anyone else.
• Month 2 red flags: Watch for the first large published dataset that includes highway miles. When that appears, treat the existing 68% figure as superseded for highway driving rather than extended to it. Also watch for city-level numbers from new markets: Atlanta and Nashville are the obvious tests of whether Phoenix or Austin is the better predictor.
• Month 3 self-check: Can you, from memory, name one reason the comparison favours Waymo and one reason it works against them? If yes, you are reading the study the way its authors intended. If you can only produce the headline, re-read the reporting section above.

Frequently Asked Questions

Is the 68% number trustworthy?

As a four-city average for 2021–2024 on non-highway roads, yes: it comes from an independent institute with a published methodology. As a general statement that driverless cars are 68% safer than people everywhere, no. The Austin result alone rules that out.

Why was Austin worse?

IIHS does not claim it was worse in any meaningful sense; it reports a 4% higher rate and notes the sample was “relatively small.” With few miles and few crashes, a handful of events swings the percentage hard. The honest reading is that Austin did not show a benefit in this dataset, not that Waymo is unsafe there.

Does this mean driverless cars are safer than my teenager?

Probably, on the roads Waymo drives, but the study does not compare against teenage drivers specifically. The human benchmark pools all drivers in those cities. Teen crash rates are well above the adult average in separate IIHS research, so the gap for a 16-year-old is likely larger than 68%, but that is inference, not a published finding.

What about highway driving?

Unmeasured. Waymo did not operate on interstates during the study period. Any claim about driverless highway safety from this study is an extrapolation.

Why do Waymo’s own numbers look better than IIHS’s?

Different mileage base (270 million vs 50 million), different period (through June 2026), different benchmark construction, and a company analysing its own product. Directionally consistent, methodologically not comparable.

Should I let my family ride in one?

That is a risk-tolerance question, not a data question, and the data supports the view that these vehicles crash less per mile than human drivers in the cities studied. The remaining uncertainty sits in highways, new cities, bad weather, and what happens when a vehicle gets stuck, which is a separate problem we cover in how robotaxis handle edge cases.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Insurance Institute for Highway Safety. (2026, July 23). “How safe are driverless cars?” IIHS News. https://www.iihs.org/news/detail/waymos-driverless-cars-crash-less-often-than-people
2. Insurance Institute for Highway Safety. “Advanced driver assistance.” IIHS Topics (citing Cicchino, 2025; Mueller et al., 2021, 2025; Reagan et al., 2021). https://www.iihs.org/topics/advanced-driver-assistance
3. Waymo. (2026, September 24). “Across Waymo’s 270 million autonomous miles, the data is unequivocal.” Waymo Blog. https://waymo.com/blog/shorts/safetydata-september26/
4. National Highway Traffic Safety Administration. “Driver Assistance Technologies.” https://www.nhtsa.gov/vehicle-safety/driver-assistance-technologies
5. National Highway Traffic Safety Administration. “Automated Vehicles for Safety.” https://www.nhtsa.gov/vehicle-safety/automated-vehicles-safety
6. Wikipedia contributors. (2026). “Self-driving car.” Wikipedia. https://en.wikipedia.org/wiki/Self-driving_car