Waymo Robotaxis Look Safer Than Human Drivers, According to IIHS – But the Numbers Are Messy

Waymo’s robotaxis sit in a weird spot these days – some people hail them as a public‑health miracle, others warn they’re rolling death traps. The latest twist? The Insurance Institute for Highway Safety (IIHS) has announced that, on paper, Waymo’s driverless cars get into far fewer accidents than the average human‑operated vehicle.

That headline sounds dazzling, but the road to the conclusion is riddled with statistical shortcuts. The problem isn’t a lack of crash data for human drivers – we have millions of police‑reported incidents each year. It’s the opposite for autonomous cars. While California has demanded crash reports from self‑driving testers since 2014 and NHTSA followed suit in 2021, the reporting standards differ from those used for conventional drivers, making apples‑to‑oranges comparisons a headache.

One key metric for crash rates is vehicle‑miles‑travelled (VMT). The Department of Transportation tracks VMT for the whole nation dating back to 2000, but autonomous fleets aren’t required to publish the same figures. Waymo, being unusually transparent, does share its mileage; most of its rivals keep that info under wraps. Because of this, the IIHS study zeroed in on Waymo alone, leaving out Tesla, Cruise, Zoox and other robotaxi hopefuls.

Cleaning the data turned into a nightmare. The researchers had to sift through every autonomous‑vehicle incident reported between 2021 and 2024, discard duplicates, strip out events where the automation wasn’t actually engaged, and toss any non‑public‑road occurrences. They even applied a rough “police‑reportable” filter: a minor fender‑bender where the driver walked away got tagged as unlikely to hit the police logs, while any crash that popped an airbag was automatically counted, regardless of injury severity.

After this massive pruning, barely a quarter of the original 736 public‑road crashes involving any level‑4 system survived the cull. The remaining 162 incidents broke down to 89 Waymo crashes, 50 Cruise, and a handful from Zoox and other firms. Because mileage data existed only for Waymo, the institute used those 89 events – 64 of which occurred while the car was fully driverless – to calculate a crash‑per‑mile figure.

The resulting headline numbers are eye‑popping: Waymo’s crash involvement rate was about 68 % lower than that of human drivers overall, with the gap widening to 76 % in Phoenix, 71 % in Los Angeles, and 35 % in San Francisco. Austin was an outlier, showing a 4 % higher rate – though the sample there was tiny. Single‑vehicle crashes were down 85 %, and injury crashes fell 81 % compared with human‑driven traffic, when measured per vehicle‑miles‑travelled.

But do those percentages really tell the whole story? NHTSA recorded roughly 24.3 million police‑reported crashes involving human drivers from 2021 to 2024. Waymo, by contrast, logged just 89 police‑reportable incidents in the same timeframe. In mileage terms, Waymo’s fleet covered about 50 million driverless miles, while human drivers amassed roughly 222 billion miles in the same cities.

Statistically, the two groups are not comparable. Waymo’s operations are confined to dense urban zones where speeds stay low and highways are off‑limits, whereas human drivers zip across highways, rural roads, and every speed bracket in between. The sample size for Waymo is orders of magnitude smaller, and the data had to be scrubbed heavily before it could even be used.

IIHS itself acknowledges the limits. Its authors warn that “the method of coding narratives is unsustainable as deployments and the number of incidents continue to grow,” and they call for better, more timely crash and VMT reporting for Level‑4 autonomous vehicles. In other words, the conclusion that Waymo is safer than the average driver is provisional at best – it’s based on a handful of cleaned‑up incidents, not a robust, longitudinal data set.

So, while it’s tempting to celebrate a headline that paints driverless cars as a safety miracle, the reality is messier. We need more comprehensive reporting, broader mileage data, and a larger pool of incidents before we can truly say autonomous fleets are safer than us human drivers.