Since Waymo burst onto the public stage nearly 20 years ago as Google’s self-titled autonomous-driving project, the promise that self-driving cars would dramatically reduce accidents has remained just that. A handful of companies operate autonomous ride-hailing services in some cities (San Francisco is downright lousy at autonomous cars), but high-profile accidents show how elusive real progress has been. Still, according to the Insurance Institute for Highway Safety, Waymo is moving in the right direction.
The non-profit organization, backed by the insurance industry and known for very thorough crash tests, recently took a look at Waymo’s safety record. In a search Published this week, researchers found that Waymo vehicles are involved in fewer accidents per vehicle mile traveled (VMT) than human drivers. It is based on an analysis of accident rates in the cities where Waymo currently operates – San Francisco, Phoenix, Los Angeles and Austin – using available data. And the limitations of those data are the real conclusions.
Waymo vehicles traveled about 50 million miles autonomously during the study period, while human drivers in the same locations traveled about 222 billion miles over the same time period. The overall rate of accidents for Waymo vehicles was 68% lower per VMT than for human drivers, and this was mostly on a city-by-city basis. Accident rates for Waymo self-driving cars were 76% lower in Phoenix, 35% lower in San Francisco, and 71% lower in Los Angeles. The Waymo fleet had 4% more crashes in Austin than human drivers, but the sample size was relatively small. And Waymo cars were involved in 85% fewer single-vehicle crashes and 81% fewer crashes resulting in injuries per VMT compared to human drivers overall.
But as the IIHS points out in an attached press release, it wasn’t easy to make any meaningful comparisons between human and autonomous driving. While companies are required to report accidents involving self-driving cars to the National Highway Traffic Safety Administration (NHTSA) and the State of California, they are not required to report the mileage covered by these vehicles, or how much is actually done in autonomous mode. This makes it impossible to know how common self-driving car accidents really are.
Waymo publishes VMT information that other companies do not, which is why this was the only study. And even with that information, researchers had to take into account differences in reporting standards for self-driving car accidents and accidents involving human drivers. Most states require drivers to report only accidents that cause injury or result in property damage of at least $1,000, and yet many drivers still do not report to police. The IIHS estimates that about half of crashes and one-third of injuries go unreported.
Meanwhile, until a 2025 amendment to its reporting requirements, NHTSA required companies to report minor fender-benders involving self-driving cars, and their expensive and vulnerable sensors often exceed the $1,000 limit if damaged. Researchers also found redundant reports, and reports in which the car might not be operating in autonomous mode. Once they removed those irrelevant reports, the researchers also tried to find out whether the accident would have been reported to police if a person had been driving, so that a real apples-to-apples comparison could be made.
The researchers concluded that only 22% of the 736 autonomous-car crashes reported during the study period were potentially “police-reportable.” In total, 89 Waymo vehicles were involved (the highest for a single operator), but only 64 of them were in autonomous mode. This is not a very large dataset, and the complex process of obtaining it shows how difficult it is to draw any real conclusions about the safety of self-driving cars.
“They are encouraging signs for the future of driverless vehicles,” Eric Teoh, lead author of the study, said in a statement about the results. “Now we need to get the data collection system right so we can ensure that level of security continues as these technologies become more prevalent.”
