IIHS: Waymo robotaxis in four cities had 68% fewer comparable crashes per mile driven than human drivers
IIHS: Waymo robotaxis in four cities had 68% fewer comparable crashes per mile driven than human drivers
The Insurance Institute for Highway Safety (IIHS) compared Waymo's driverless trips with human driving in San Francisco, Phoenix, Los Angeles, and Austin. In this dataset, injury-involved crashes per mile driven were 81% less frequent for Waymo.
On 8 September 2026, IEEE Spectrum published an analysis of a July IIHS study. The institute compared driverless Waymo robotaxi trips with human driving in four cities; the results apply specifically to that operational scope.
To make the comparison, IIHS first aligned two different crash-reporting records to a common threshold. Human drivers often do not call the police after minor collisions, whereas federal rules before a 2025 amendment required operators of automated vehicles to report even small incidents. The researchers read through incident narratives and selected those that, in their judgment, a reasonable driver would have reported to police. Any incident involving airbag deployment was automatically included.
Of 736 incidents on public roads with automation engaged, the researchers judged 22% to be comparable to police-reported crash statistics. Over the study period, Waymo vehicles covered approximately 50 million miles, while human drivers covered 222 billion miles in the same locations and timeframe. Waymo's rate of comparable crashes per mile was 68% lower.
By city, the difference was 76% in Phoenix, 71% in Los Angeles, and 35% in San Francisco. In Austin, Waymo's rate was 4% higher; the small number of cases there makes the estimate particularly sensitive to each individual collision. The IIHS conclusion applies to the operating Waymo robotaxi fleet in these cities and to this category of crashes.
As robotaxi operations expand, IIHS recommends establishing standardized collection of mileage, driving-mode, and crash data. Such data would allow comparable monitoring to continue once manual review of incident narratives can no longer keep pace with the volume of reports.