Table of Contents
Reading Crash Statistics: A Skill to Teach Your Kid
Reading crash statistics well means asking four questions. The July 2026 IIHS study of driverless cars shows why, including the city where the number flipped.
Reading crash statistics is a specific skill, and the July 2026 study from the Insurance Institute for Highway Safety is the best teaching text I have seen in years. Published on July 23, 2026, it found that Waymo’s driverless vehicles had a crash rate 68% lower than human drivers per mile travelled. That number went everywhere. The same study, same dataset, same researcher, also found that in Austin the Waymo crash rate was 4% higher than human drivers.
Both numbers are correct. If your kid can explain why, they have learned something that will serve them for the rest of their life.
Key Takeaways
- IIHS’s July 23, 2026 analysis compared about 50 million Waymo driverless miles against roughly 222 billion human-driven miles in the same four metro areas from 2021 to 2024. Overall: 68% lower crash rate, 85% lower single-vehicle crashes, 81% lower injury crashes.
- The city-level results ranged from 76% lower in Phoenix to 4% higher in Austin. The whole analysis rested on just 64 police-reportable Waymo crashes, which is why city-level splits should be read as rough, not precise.
- Human drivers underreport roughly half of their crashes. Driverless operators over-report, because federal reporting rules push them to. IIHS had to hand-clean the data and discard about 25% of the reports as redundant or not qualifying.
- Waymo’s own published figures are larger than IIHS’s: 95% fewer serious-injury crashes across more than 100 million miles. Both can be honest at once, because they use different comparison groups.
- RAND’s 2016 analysis found that proving a fatality-rate improvement by driving alone would take hundreds of millions to hundreds of billions of miles. That is the ceiling on what any mileage-based claim can prove right now.
What the July 2026 IIHS study actually found
A crash rate is a count of crashes divided by a measure of exposure, almost always miles travelled. IIHS’s study, led by Eric Teoh, the institute’s director of statistical services, covered San Francisco, Phoenix, Los Angeles and Austin from 2021 to 2024.
The exposure figures are worth saying out loud because the asymmetry is enormous. Roughly 50 million Waymo driverless miles. Roughly 222 billion human-driven miles in the same places over the same period. That is a ratio of about 4,400 to one. Within those 50 million miles, IIHS identified 64 police-reportable Waymo crashes.
Here are the headline reductions, from the IIHS release:
| Measure | Waymo vs human drivers |
|---|---|
| All crashes per mile | 68% lower |
| Single-vehicle crashes per mile | 85% lower |
| Injury crashes per mile | 81% lower |
| Phoenix | 76% lower |
| Los Angeles | 71% lower |
| San Francisco | 35% lower |
| Austin | 4% higher |
Two caveats belong next to that table permanently. First, the vehicles did not operate on interstate highways during the study period, so the comparison excludes the road type where high-speed fatal crashes concentrate. Second, in police-reportable crashes, Wikipedia’s summary of the underlying data records Waymo as at fault in 8% of cases with fault unclear in another 8%, meaning most of these crashes involved a human doing something to a driverless car rather than the reverse.
IIHS president David Harkey did not soften the data problem: “The present data collection system isn’t good enough to allow continuous monitoring of a large-scale expansion.” An institute that just published a favorable headline number told you, in the same release, that the measuring instrument needs work. That is what good research looks like.
The four questions that decide whether a crash statistic means anything
One: what is the numerator? “Crashes” is not one thing. Police-reportable crashes, insurance claims, injury crashes and fatalities are four different counts that move independently. A technology can cut fender-benders and do nothing for fatalities, or the reverse. When a number has no noun attached, it has no meaning.
Two: what is the denominator? Per mile, per hour, per trip and per vehicle give different answers. Per-mile favors anything that drives long distances on easy roads. Per-trip favors anything that makes long trips. Driverless fleets operate in dense cities at low speeds, which produces many short, slow, complicated miles. Choosing the denominator is choosing the conclusion, and honest researchers say which one they picked and why. If the levels of automation themselves are fuzzy to you, start with the automation levels explained for parents, because a statistic about “self-driving” cars is meaningless until you know which level is being measured.
Three: who is in the comparison group? This is where most published comparisons quietly fall apart. IIHS did the hard version: human crash rates for the same four metro areas, over the same years, rather than a national average. A national average would have included rural two-lane roads and interstates where crash severity is completely different. If you compare a robotaxi in downtown Phoenix to the average American driver, you are mostly comparing two cities.
Four: did both sides report the same way? This is the one almost nobody asks, and it is the heart of the IIHS paper. Human drivers do not report roughly half of the crashes they are in; the fender-bender in a parking lot never reaches a police report. Driverless operators report nearly everything, because NHTSA’s crash-reporting obligations require it. Left uncorrected, that asymmetry makes the robot look worse than it is. IIHS researchers manually cleaned the federal reports, discarded about 25% as redundant or below the reporting threshold, and applied what they described as a common-sense standard for whether an incident would have been police-reportable had a human been driving.
So when your kid reads “68% safer,” the honest translation is: in four cities, on non-highway roads, over four years, after adjusting for the fact that the two groups report crashes differently, the per-mile police-reportable crash count was about two-thirds lower. Longer. Truer.
Why the company’s number is bigger than the independent one
Go to Waymo’s own safety page and you will find different figures: 95% fewer serious-injury-or-worse crashes, 82% fewer injury-causing crashes, and 93% fewer pedestrian crashes with injuries, across more than 100 million miles. A Swiss Re analysis cited there reports 92% fewer bodily injury claims and 88% fewer property damage claims over 25 million miles.
IIHS said 68%. Waymo says 95%. Neither is lying.
They are measuring different numerators (all crashes versus serious-injury crashes), over different mileage totals, with differently constructed human baselines. The serious-injury number will almost always look better than the all-crash number, because low-speed urban operation produces lots of minor contacts and very few severe ones. This is the most useful thing a thirteen-year-old can learn about statistics: two true numbers about the same subject can differ by a factor of two, and the difference lives in the definitions.
The follow-up question is the one to teach. Not “which number is right?” but “who built the comparison group, and did they have a reason to want a particular answer?” IIHS is funded by insurers and has no stake in Waymo’s valuation. Waymo has an enormous one. That does not make Waymo’s numbers false. It does mean an independent replication carries more weight, and it is why the 68% figure should anchor your mental model rather than the 95%.
There is also a hard statistical ceiling here. RAND’s Nidhi Kalra and Susan Paddock calculated in 2016 that demonstrating an improvement in fatality and injury rates through test driving alone would require “hundreds of millions of miles and sometimes hundreds of billions of miles,” and that existing fleets would need tens to hundreds of years to accumulate them. Their conclusion was blunt: “Developers of this technology and third-party testers cannot simply drive their way to safety.” Fifty million miles is a lot of driving. It is not enough to settle the fatality question, and anyone claiming otherwise is going beyond the data.
How to Teach Your Kid About Reading Crash Statistics
Ages 5–8: the denominator game
Count something twice with different denominators. Who eats more cereal, you or your child? Per bowl, maybe they do. Per day, you do. Per kilogram of body weight, they win easily. Same cereal, three different true answers. Say the word “denominator” out loud and let it be a funny word they know. You are planting the idea that a comparison needs a per what.
Ages 9–12: find the hidden city
Give them the table above and ask one question: why is Austin different from Phoenix? Let them generate hypotheses. Different streets. Different weather. Fewer miles driven there. Different drivers. Then tell them the fact that resolves it: the entire study found 64 crashes total, so Austin’s slice might rest on a handful of events. Flip a coin ten times and count heads, then do it again. The number moves. Small samples wobble. That single exercise inoculates a kid against half the bad statistics they will ever read.
Ages 13+: rebuild the comparison group
Hand your teenager the IIHS release and the Waymo safety page side by side and ask them to write three sentences: what each one counted, what each one divided by, and who each one compared against. Then ask them to decide which number they would use if they were a city council member voting on a permit. There is no answer key. The output is a defensible paragraph, and that is the whole skill.
The question to ask: “If a company reports every scratch and ordinary drivers report about half of theirs, which group will look worse in the data, and how would you fix that before comparing them?”
The checklist, applied to three real numbers
| Claim | Numerator | Denominator | Comparison group | What it can support |
|---|---|---|---|---|
| IIHS: 68% lower crash rate | All police-reportable crashes, hand-cleaned | Per vehicle-mile, ~50M driverless miles | Human drivers in the same 4 metros, 2021–2024 | A reasonable claim about non-highway urban driving |
| Waymo: 95% fewer serious-injury crashes | Serious-injury-or-worse crashes only | Per mile, 100M+ miles | Waymo’s own human benchmark for its cities | A narrower claim about severe outcomes |
| Austin: 4% higher | Police-reportable crashes in one city | Per mile, a fraction of 50M | Austin human drivers | Very little; the sample is too small |
| RAND 2016 | Fatalities and injuries | Miles needed for statistical proof | Human fatality rate | A limit on what any mileage claim can prove |
The fourth row is the one to sit with. RAND’s finding is not about Waymo. It is about arithmetic. Fatal crashes are rare per mile, so detecting a change in them requires an enormous number of miles. That is why regulators and researchers increasingly argue for simulation, scenario testing and public incident data rather than waiting for mileage to settle the question.
What to do at home
Keep a family rule: no percentage without a “per”
When anyone in the house quotes a statistic, the standing follow-up is “per what?” Make it a reflex rather than a correction. It works on crash data, on sports arguments, on test scores and on advertising claims about everything from toothpaste to tutoring.
Teach the source hierarchy out loud
A peer-reviewed paper beats an institute release, which beats a company press page, which beats a news summary of a company press page. Say the hierarchy, then show it working on this exact story. IIHS’s underlying paper is titled “Rise of the machines: crash experiences of highly automated vehicles and human drivers,” and the fact that a paper exists at all is a meaningful signal about the release.
Practice the reporting-asymmetry question on things that are not cars
School discipline data, injury rates in two different sports, crime statistics across two cities with different reporting cultures. The same trap appears everywhere: the group that records more carefully looks worse. Our piece on what teen distracted driving statistics actually show runs the same four questions on a dataset where the reporting problem is even larger. A kid who has seen this three times in three contexts will see it the fourth time on their own.
Use the robotaxi decision as a live case
If driverless rides are available where you live, you have a real decision to make rather than a hypothetical. Our walkthrough of what parents should decide first about teen robotaxi riders lays out the questions. Reading the safety data is step one of that decision, not a separate hobby.
What not to do: don’t turn skepticism into paralysis
The failure mode of teaching statistical caution is a teenager who dismisses every number. The 68% figure is real, independently produced, and more favorable to driverless cars than most people expected. Caution about precision is not the same as doubt about direction. Say that explicitly, or you will raise a cynic instead of an analyst.
What to Watch For Over the Next 3 Months
- Week 4: Watch for the peer-reviewed version of the IIHS paper to appear in a journal, and for any published critique of its reporting adjustment. Methodological disputes are where the real information is.
- Month 2 red flags: Any new “X% safer” claim that does not name its denominator or its comparison group. Also watch for claims that extend an urban, non-highway dataset to highway driving, which the study explicitly does not cover.
- Month 3 self-check: Ask your kid to explain the Austin result in their own words. If they say “Austin is more dangerous,” the lesson did not land. If they say “there weren’t enough crashes there to tell,” it did.
Frequently Asked Questions
Does the 68% figure mean driverless cars are safer than me?
It means that in four metro areas, on non-highway roads, between 2021 and 2024, the per-mile police-reportable crash rate was about two-thirds lower than for human drivers in those same places. It is not a statement about you specifically, and it says nothing about highway driving, which the study excluded.
Why does Waymo quote 95% when IIHS says 68%?
Different numerators. Waymo’s 95% refers to serious-injury-or-worse crashes; IIHS’s 68% covers all police-reportable crashes. Severe crashes are rarer and show bigger reductions in low-speed urban operation. The mileage totals and human baselines also differ.
Is a company’s own safety data worthless?
No, and treating it that way is lazy. Waymo publishes methodology and has had analyses reviewed externally, including by Swiss Re. The correct posture is to weight independent replication higher, not to discard the primary data.
Why does underreporting matter so much?
Because it changes the direction of the error. If one group reports every event and the other reports about half, the careful group’s crash rate looks inflated. Comparing them raw makes the automated system look worse than it is, which is the opposite of the bias most people assume.
At what age can a kid actually understand this?
Nine or ten for the denominator idea, and twelve or thirteen for comparison groups and reporting asymmetry. The coin-flip exercise works much younger than that and does most of the heavy lifting.
What single number should I remember?
Sixty-four. That is how many police-reportable Waymo crashes the whole study rests on. Holding that in mind makes you appropriately confident about the overall direction and appropriately humble about any city-level split.
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
- 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
- Waymo. “Safety.” Waymo Safety Hub. https://waymo.com/safety/
- Kalra, N., & Paddock, S. M. (2016). “Driving to Safety: How Many Miles of Driving Would It Take to Demonstrate Autonomous Vehicle Reliability?” RAND Corporation, RR-1478. https://www.rand.org/pubs/research_reports/RR1478.html
- National Highway Traffic Safety Administration. “Automated Vehicles for Safety.” U.S. Department of Transportation. https://www.nhtsa.gov/vehicle-safety/automated-vehicles-safety
- Insurance Institute for Highway Safety. (2024 data). “Fatality Facts: Teenagers.” https://www.iihs.org/topics/fatality-statistics/detail/teenagers
- Wikipedia contributors. “Self-driving car.” Wikipedia. https://en.wikipedia.org/wiki/Self-driving_car
