The early numbers on robotaxi safety look genuinely impressive – but according to Henry Kowal (pictured), director of insurance product at Arity, the data behind those numbers is still too narrow for the industry to treat it as the full picture.
"I think the early data that we're seeing is encouraging. However, I think the data that we're currently seeing is limited," Kowal said in an interview with Insurance Business. That gap may prove especially relevant for Canada: snowy and rainy conditions are exactly what robotaxi safety data hasn't captured at scale so far, and that could soon change as fleets expand into harsher climates.
Waymo remains the dominant player in the robotaxi space, and Kowal pointed to the scale of its safety data as the strongest evidence available today.
"They have over 20 million paid rides so far. They're doing about half a million a week now. And they've logged 200 million fully autonomous miles... their peer-reviewed safety impact data reports that about 93% fewer pedestrian injury crashes and about 94% fewer serious or fatal injury crashes versus the same-city human benchmark," Kowal said, citing Waymo's own published safety data.
Read more: Waymo recalls nearly 3,800 robotaxis after driverless car drove through a flooded road
A separate, independent study released by the Insurance Institute for Highway Safety reached a similar, if more conservative, conclusion, finding Waymo vehicles were involved in 68% fewer police-reportable crashes per mile than human drivers across San Francisco, Phoenix, Los Angeles and Austin.
The trouble, Kowal said, is how narrow the conditions behind those numbers actually are. Waymo remains far ahead of every other entrant – Uber, Tesla and Zoox included – meaning almost all of the safety evidence collected to date comes from a single operator. Tesla, for instance, is currently running only a few dozen vehicles compared with Waymo's roughly 4,000. On top of that, robotaxis currently run almost exclusively in mapped, temperate cities.
"They typically exclude things such as heavy snow and flooding... it's data being collected in a very controlled environment," Kowal said.
That's a meaningful gap for a Canadian audience. Cities like Toronto see the kind of heavy snow, ice and seasonal flooding that robotaxi fleets have largely avoided so far, and Waymo's own service has only recently expanded into its first real winter-weather market, Denver – where the vehicles are currently designed to pull over rather than drive through snow. Insurance Business Canada previously reported on a Waymo recall triggered by a vehicle that entered a flooded highway in Texas, an example of exactly the kind of weather-related edge case still largely untested at scale.
Kowal pointed to an incident in San Francisco this past July 4th, when a number of Waymo vehicles became stranded amid holiday fireworks, heavy congestion and a cell service blackout, as a clear illustration of where the real risk sits.
"It wasn't like the driving software that failed, it was really more the human environment that overwhelmed the system," Kowal said.
That matters for how the industry should be reading the safety data, he said. The vehicles themselves performed as designed; it was the unpredictable conditions around them, generated by human activity, that created the disruption. It's a dynamic he expects to keep surfacing as robotaxi fleets scale into busier, messier, and more varied environments than the ones they've mostly operated in so far.
Read more: AI in clinical settings may become a malpractice defense asset
"How do these cars, how do these vehicles coexist with the human drivers that are on the road?" Kowal said.
Even the researchers behind the IIHS study aren't calling the picture complete. Kowal noted that IIHS itself flagged a limitation in how robotaxi safety is currently being tracked.
"They did say that, at least from their perspective, they didn't believe the monitoring system is good enough yet to track robotaxis at a full-scale rollout," Kowal said.
That caveat is part of why Kowal keeps coming back to the same underlying issue: the vehicles themselves may be performing well, but the environment they're driving through is still the wildcard.
