An independent safety study just gave the robotaxi industry its strongest evidence yet. The Insurance Institute for Highway Safety, the same body behind the Top Safety Pick crash awards, analysed federal crash data from 2021 to 2024 and found that Waymo’s driverless vehicles were involved in 68% fewer police reportable crashes per mile than human drivers in the same cities. That’s a genuinely large gap, and it comes from an organisation with no commercial stake in the outcome.
It’s not a clean win, though. The study comes with a long list of caveats that IIHS itself flags, and reading past the headline number matters if you want an honest picture of where autonomous driving actually stands. Waymo runs Level 4 systems, meaning no driver is needed at all within its operating area. If the terminology around driving automation is unfamiliar, my guide to the SAE autonomous driving levels explains what separates Level 4 from the Level 2 assisted driving systems most European buyers actually have in their cars today.
| City | Waymo crash rate vs humans |
|---|---|
| Phoenix | 76% fewer crashes |
| Los Angeles | 71% fewer crashes |
| San Francisco | 35% fewer crashes |
| Austin | 4% higher crash rate (small sample) |
What the Study Actually Found
IIHS looked at roughly 50 million miles of fully driverless Waymo operation across Phoenix, San Francisco, Los Angeles and Austin, and compared it against human driving in the same areas over the same period. After stripping out duplicate reports and incidents that wouldn’t typically reach a human driver’s insurer, researchers landed on a rate of 4.06 crashes per million miles for human drivers against 1.28 for Waymo. That’s the source of the 68% figure, and it holds up as roughly three times fewer crashes overall.
The gap widens in specific categories. Waymo recorded 85% fewer single vehicle crashes and 81% fewer crashes resulting in injury. Its crashes also tended to be less severe on average, and Waymo was less often the party at fault. One genuinely surprising result cuts the other way: despite relying on lidar and radar, neither of which needs visible light to work, Waymo actually had a higher share of its crashes happen in the dark than human drivers did. IIHS didn’t fully explain why, though it’s a reminder that sensor hardware alone doesn’t guarantee good decision making in every condition.
The Caveats That Matter
This is the part worth reading slowly, because it changes how much weight the headline number deserves.
Crash reporting rules aren’t symmetrical. Autonomous vehicle operators are legally required to report crashes to regulators, but until a 2025 rule change, even minor contact like scraping a vehicle’s underbody counted as reportable. Human drivers routinely don’t report fender benders at all. Researchers had to manually clean the data to make a fair comparison, which introduces a degree of subjectivity no study can fully remove.
The concept of an “average human driver” is also doing a lot of work here. That average includes distracted, drowsy, drunk and impaired drivers, and a robot that only has to beat that pool isn’t the same as a robot beating how a careful, attentive driver actually performs. IIHS itself raises this point directly.
Road type matters too. Waymo doesn’t operate on highways, so highway miles were excluded from the human comparison as well, to keep things fair. But highways are statistically safer than city streets, so removing them makes human drivers look worse than they would in a full comparison. On top of that, half of Waymo’s crashes happened on streets with speed limits under 25 mph, against just 8% of human crashes, which likely flatters Waymo’s severity numbers since lower speed crashes are inherently less dangerous.
There’s also a structural effect from having no one in the car. Nearly half of Waymo’s studied miles were driven with no passenger aboard, and an empty car can’t produce an injured occupant. That’s a genuine safety benefit in one sense, but it also means the injury comparison isn’t measuring quite the same thing on both sides.
Finally, the sample size is still small next to human driving. Fifty million Waymo miles sounds like a lot until you compare it with roughly 222 billion human miles driven in those same four cities over the same years, or the more than 3 trillion miles Americans drive annually nationwide. Several crash categories in the Waymo data had just zero or one recorded incidents, which makes for a noisy dataset that a few more incidents either way could swing meaningfully.
Where Tesla Fits In, and Where It Doesn’t
Notably absent from the study is Tesla. Its Austin robotaxi service only began in 2025 with safety monitors aboard, and had racked up roughly 380,000 fully unsupervised miles at the time of writing, a tiny fraction of Waymo’s 50 million. Tesla has published its own internal crash numbers, but hasn’t submitted them for independent verification the way this IIHS study represents for Waymo. Until Tesla accumulates enough unsupervised miles and opens that data to outside researchers, there’s no independent way to know how its system compares.
What This Means for Europe
Waymo itself doesn’t operate in Europe, but Europe already has its own live commercial robotaxi service, and it launched in Zagreb rather than London or Berlin. Verne, a Croatian company spun out of electric hypercar maker Rimac, launched Europe’s first commercial robotaxi service there in April 2026, running Pony.ai’s seventh generation autonomous system in a small fleet of electric Arcfox Alpha T5 SUVs, in partnership with Uber. For now, a trained safety operator still sits behind the wheel of every Verne vehicle, so it isn’t yet the fully unsupervised Level 4 operation Waymo runs in parts of the US. Verne has said it wants to remove that operator entirely by the end of 2026, and it’s already in permitting talks for eleven more cities across the EU, the UK and the Middle East. If it gets there, the same questions the IIHS study raises about crash reporting, sample size and what “safer than a human” actually means will need answering on Croatia’s narrow, cobbled streets rather than Phoenix’s wide grid.
The assisted driving systems already arriving in European showrooms raise a smaller scale version of the same legal responsibility question. BYD’s God’s Eye system has already put that question in front of European regulators and courts, and I covered the liability implications in detail in a separate piece on God’s Eye and who’s actually responsible when it’s engaged. Xpeng is taking a similar bet with its VLA assisted driving system, which the company plans to bring to Europe from 2027 as a direct challenger to Tesla’s FSD, a launch I covered in my piece on Xpeng’s European ambitions. None of these systems are Level 4, and the driver remains legally responsible in every case currently sold in Europe. But the direction of travel is the same one Waymo and Verne both represent, and the same reporting and comparison problems IIHS ran into will show up again the moment a European regulator tries to judge whether any of these systems are actually safer than the drivers they’re assisting, or eventually replacing.
Conclusion
The honest read of this study is that Waymo’s safety case is real but not yet proven beyond reasonable doubt. Sixty eight percent fewer crashes is a meaningful number from a credible, independent source, and it’s a considerably stronger result than anything Tesla can currently back up with outside verification. At the same time, the caveats around reporting rules, the “average human” comparison, road type, and sample size are large enough that this shouldn’t be read as a final verdict. It’s a strong data point in an ongoing experiment, not the end of the conversation, and the same scrutiny will need to apply to every assisted and autonomous system heading for European roads over the next few years.
FAQ
How much safer is Waymo than a human driver, according to IIHS? IIHS found Waymo’s driverless vehicles were involved in 68% fewer police reportable crashes per mile than human drivers across the same cities, roughly three times fewer crashes overall.
Was Tesla included in the IIHS study? No. Tesla’s Austin robotaxi service only launched in 2025 and had accumulated around 380,000 unsupervised miles at the time of writing, far too little for a meaningful comparison, and its safety data hasn’t been independently verified the way this study covers Waymo.
What is a Level 4 autonomous vehicle? A Level 4 vehicle can handle all driving tasks without a human driver, but only within a specific, geofenced operating area and set of conditions. My guide to the SAE autonomous driving levels covers the full scale from Level 0 to Level 5.
Does this study mean Waymo is definitely safer than human drivers? It’s strong evidence in that direction, but IIHS itself flags several caveats, including how crash reporting rules differ between humans and autonomous vehicles, and the relatively small sample size compared with total human miles driven.
Is Waymo operating in Europe? No, Waymo currently operates only in the United States, though it’s targeting London later in 2026. Europe already has its own commercial robotaxi service through Croatian company Verne, which launched in Zagreb in April 2026, though Verne’s vehicles still carry a safety operator onboard for now.
What is Verne, and how does it relate to this study? Verne is a Croatian autonomous mobility company spun out of Rimac that launched Europe’s first commercial robotaxi service in Zagreb in April 2026, using Pony.ai’s autonomous driving technology. It currently runs with a safety operator onboard and aims to go fully driverless by the end of 2026, which would make it Europe’s first genuine test case for the same safety questions IIHS raised about Waymo.
Featured Image Credit: Waymo
Latest News
