Waymo's safety numbers improved. The way we check them didn't

Waymo released its largest safety dataset this week: 271 million fully driverless miles across Atlanta, Austin, Phoenix, Los Angeles and the San Francisco Bay Area, logged through the end of June, 50 million more than its previous report. The headline numbers are strong — 82 percent fewer injury-causing crashes than human drivers, 95 percent fewer crashes causing serious injury or worse, 93 percent fewer injury-causing pedestrian crashes.

They are also self-reported, unaudited, and deployed as lobbying ammunition in cities that have not yet let Waymo in. That is not an accusation of dishonesty; it is the structural problem. The company that builds the system, chooses the comparison benchmark and publishes the result is also the only party with the raw data.

What independent research actually found

The strongest corroboration is not Waymo's. The Insurance Institute for Highway Safety published its own analysis in July, comparing crash involvements per vehicle mile traveled across San Francisco, Phoenix, Los Angeles and Austin, and landed on 68 percent fewer crashes than human drivers — lower than Waymo's number, and for instructive reasons. IIHS had to clean federal data first: it culled redundant records, removed incidents where automation was not engaged, and judged whether a crash was severe enough that a reasonable person would have called police. Only about 22 percent of 736 automation-engaged crashes met that bar.

The variance by city is the part Waymo's summary flattens. IIHS found 76 percent fewer crashes in Phoenix, 71 percent fewer in Los Angeles, 35 percent fewer in San Francisco — and 4 percent more in Austin, where the sample was small. A single fleet-wide percentage hides the city where the record is worst, which is exactly where a regulator would want to look.

Johns Hopkins puts the honest ceiling on all of it. Johnathon Ehsani, who studies road safety at the Bloomberg School of Public Health, notes there is no consensus standard for how safe is safe enough, no equivalent of a driver's test for an automated system, and no uniform reporting rules across companies. The data is "suggestive of significant improvements," he says. That is a different claim from the one made by the medical letter urging cities to deploy autonomous vehicles now, which he says oversimplifies it.

The federal record tells a less flattering story

The same mileage generates a second dataset that Waymo does not curate. Under NHTSA's 2021 standing general order, automated driving systems must report any crash involving physical contact with a vehicle, person or property. By that accounting, Waymo vehicles were involved in 797 crashes across all cities between March and September 2026. The company reported one fatality: a pedestrian in Dallas struck by a human driver and thrown into a Waymo.

Neither dataset is wrong. They measure different things — reported contact versus benchmarked crashes per mile against a human baseline — and the gap between 797 incidents and "95 percent fewer serious crashes" is precisely why a reader cannot adjudicate the claim from outside. IIHS president David Harkey's verdict on the monitoring regime itself is the sentence to remember: it "isn't good enough to allow continuous monitoring of a large-scale expansion."

That matters more this month than usual, because the expansion is happening while the measurement is still being designed. Waymo runs nearly 4,000 vehicles in more than 10 US cities, about two million driverless miles a week and over 500,000 paid rides weekly. It added Las Vegas and restarted San Antonio. Zoox's 100-vehicle cap in Nevada expires, clearing a commercial fleet it was not previously allowed to build. May Mobility is going public in a $1.4 billion SPAC deal that will hand shareholders a pure-play robotaxi company to judge on safety disclosure. NHTSA's answer is a consortium with SAE, called ASCEND, aimed at writing "ADS competency" standards — the first attempt to define what a competent automated driver even is, and the subject of our report when NHTSA and SAE signed on to write the first AV driving test. Administrator Jonathan Morrison has offered no timeline.

Who wins, and who is left holding the inference

Waymo wins the news cycle, and it has earned a real argument: two independent analyses now point the same direction, and its fleet reports crashes that human drivers routinely never mention to police. IIHS's own figures — roughly 60 percent of property-damage crashes and about a third of injury crashes go unreported by humans — mean the robotaxi record is held to a stricter standard than the human baseline it is compared against.

The losers are the cities trying to make a decision. Boston, New York and Washington have cited traffic disruption and job losses in resisting service; they now have a company dashboard, an academic caution and a federal incident database, and no agreed method for combining them. Carnegie Mellon's Phil Koopman has argued for years that aggregate dashboards function as corporate relations — his point is not that the numbers are false but that they never capture the embarrassing non-crash failures, and that a fleet ten to a hundred times larger is what it would take to say anything credible about fatalities. The Cruise precedent sits behind that skepticism: a company with favourable safety data and a serious disclosure problem, which resolved in neither direction investors expected.

There is a specific asymmetry worth naming. The public scrutiny is real but distributed unevenly. A Comma device involved in a fatal crash triggers a federal defect investigation and a preliminary evaluation file — our analysis of NHTSA's probe into Comma and what it means for open-source driving traced how that process works. A Waymo with 797 reported crash involvements produces a company press release. The per-unit risk is far lower; the accountability mechanism is also far softer. Meanwhile the people closest to the machinery carry the injuries: contract test drivers for Waymo and Zoox have reported sprains, whiplash and worse, as we covered in Waymo and Zoox test drivers are getting hurt as robotaxis scale — physical harm that never appears in either safety dataset.

What to watch

Three things would convert this from duelling percentages into evidence. Whether ASCEND produces a testable competency standard with dates attached, rather than a consortium communique. Whether NHTSA's standing general order gets the reporting upgrades IIHS asked for — normalised severity, mileage denominators, enough granularity to compute city-level rates independently. And whether a city that has said no, like New York or Boston, ends up setting the disclosure terms Waymo accepts; that is where the leverage sits, because the company needs the market more than the regulator needs the number.

Until then, treat 82 percent and 797 as two facts about the same fleet, and note that only one of them requires the operator's cooperation to check.

Would you trust a robotaxi more if the safety numbers came from someone other than the company running the fleet? Tell us in the comments.

Sources: Waymo Safety Impact · IIHS · The Verge · Carscoops · Johns Hopkins Bloomberg School of Public Health · TechCrunch Mobility