The US Insurance Institute for Highway Safety has published an analysis of driverless Waymo operation covering roughly 80 million kilometres across four American cities, and it found the vehicles involved in 68% fewer crashes than human drivers on the same roads over the same period. It is the largest independent look at a deployed driverless fleet so far, and it is worth reading carefully — including the parts that complicate the headline.
What the study measured
The IIHS examined 89 police-reportable crashes involving Waymo vehicles in San Francisco, Phoenix, Los Angeles and Austin between 2021 and 2024. It compared those against human-driven crash rates in the same cities, drawn from roughly 355 billion kilometres of human driving.
| Metric | Waymo vs human drivers |
|---|---|
| All police-reportable crashes | 68% lower |
| Single-vehicle crashes | 85% lower |
| Crashes resulting in injury | 81% lower |
The single-vehicle figure is the most interesting of the three. Single-vehicle crashes — leaving the road, hitting a fixed object — are the ones most directly attributable to the driver rather than to someone else's mistake. An 85% reduction there is a stronger claim about driving competence than the headline number is.
The city-by-city split matters
The aggregate hides a wide spread.
| City | Crash rate vs human drivers |
|---|---|
| Phoenix | 76% lower |
| Los Angeles | 71% lower |
| San Francisco | 35% lower |
| Austin | 4% higher |
Austin is the outlier, and the IIHS attributes it in part to a small sample — Waymo's Austin operation is the newest of the four and has accumulated the least mileage. But the San Francisco figure is harder to wave away. A 35% reduction is real and worth having; it is also half the aggregate, in the densest and most complex of the four environments.
The obvious reading is that the benefit scales inversely with difficulty. Phoenix is wide, gridded and dry. San Francisco is none of those things.
What it does not tell you
The study is about Waymo, not about autonomy. Waymo operates a geofenced, lidar-equipped, remotely supervised fleet on roads it has mapped in advance, and it removes itself from conditions it does not handle. That is a fundamentally different proposition from a consumer car running a vision-only system anywhere its owner points it.
It also does not settle the comparison people most want settled. Tesla has published its own European figures claiming FSD is 5.2 times safer than manual driving across 65 million kilometres. Those numbers are self-reported and use a different methodology; the IIHS numbers are independent and use a matched-road comparison. They are not directly comparable, and the gap between self-reported and independently-audited safety data is precisely the issue currently playing out in the Dutch regulator's refusal to publish its FSD assessment.
The IIHS authors flag the limits themselves: police-reportable crashes undercount minor collisions, the human baseline includes vehicles far older than Waymo's fleet, and four cities in one country is not a general result.
Why it matters in Europe
Europe has no deployed driverless fleet at Waymo's scale, and the regulatory route here runs through type approval rather than state-by-state permitting. But the evidentiary standard the IIHS has set is portable, and it is the standard European regulators are converging on: matched-road comparison, independent analysis, published method.
Germany has already cleared Momenta for Level 4 testing nationwide and London has licensed supervised Uber-Wayve robotaxi operation. Both will eventually need to answer the question this study asks. Having a number that was produced by someone with no commercial stake in the answer is the useful precedent here — more useful, arguably, than the number itself.