Waymo Expands Driverless Safety Analysis to 271 Million Miles
The company reports lower injury-crash rates than adjusted human benchmarks across five metro areas, while its own methodology warns that no comparison is perfectly equivalent.
Edited by Tyronne Panaino
Waymo published an updated safety analysis on September 24 covering 271.3 million miles driven without a human driver through the end of June 2026. The company's release spans Atlanta, Austin, Los Angeles, Phoenix and the San Francisco Bay Area, and reports lower injury-crash rates than adjusted human benchmarks across those operations.
The update matters because autonomous-driving safety claims often rely on small samples or mismatched comparisons. Waymo's Safety Impact hub now exposes city-level mileage, outcome definitions, rates and methodological caveats. It is a larger operational record, but it remains a company-run analysis rather than an independent finding that applies to every road, city or driving condition.
A larger operating record changes the comparison
The dashboard lists 271.3 million rider-only miles: 92.121 million in Phoenix, 82.421 million in the San Francisco Bay Area, 67.098 million in Los Angeles, 21.064 million in Austin and 8.624 million around Atlanta. Waymo compares its crash rate with an estimated human-driver rate over equivalent distances in those operating areas.
Across all five locations, Waymo reports 0.67 injury-reported crashes per million miles for its vehicles and 3.77 for the adjusted human benchmark. It describes that difference as 82% fewer injury-causing crashes, or 841 fewer incidents than the benchmark would predict. For crashes involving serious injury or worse, the dashboard reports 0.01 incidents per million miles for Waymo and 0.21 for the benchmark, which it summarizes as a 95% reduction and 55 fewer incidents.
Those are modeled comparisons, not a count of crashes that can be observed in an alternate world. Waymo's blog turns the estimate into an avoided-injury claim using an assumption of at least one injured person per crash. The more cautious reading is that the company's selected data and benchmark produce an estimated difference of 841 injury-causing crashes.
The benchmark adjusts for reporting and road mix
Waymo says its own crash record is derived from incidents reported under the US National Highway Traffic Safety Administration's Standing General Order. That reporting threshold can include minor contact that would not necessarily appear in conventional police data, so a raw comparison would not be like-for-like.
For the any-injury outcome, the methodology applies a 32% underreporting adjustment to the human benchmark. It does not apply the same adjustment to serious-injury or airbag-deployment outcomes. The analysis also reweights city-level human data to better reflect where Waymo operates, because the difficulty and crash risk of streets vary within a metro area. Rates are expressed per million miles, and the dashboard publishes 95% confidence intervals around the estimates.
What the figures do not establish
Waymo explicitly says there is no perfect equivalent comparison between autonomous-vehicle and human data. The human benchmark is derived from police-reported records and estimated vehicle miles, while Waymo vehicles operate inside defined service areas and mostly on surface streets. National underreporting estimates may not describe each city equally well, and city-level adjustments cannot recreate every street, traffic pattern or weather condition encountered by the fleet.
The analysis is retrospective and ends in June, so it does not measure later route expansions or performance in places outside the five listed metro areas. It also does not by itself establish how another autonomous-driving system would perform. Independent replication using the downloadable incident identifiers and public reporting data would strengthen confidence in the result.
What operators and policymakers can use now
The useful contribution is not a blanket declaration that autonomous driving is solved. It is a more inspectable denominator, a documented outcome taxonomy and a benchmark that can be challenged or reproduced. Regulators, researchers and city officials can compare future releases against the same definitions while watching whether confidence intervals and city-level results remain stable as mileage grows.
The next verifiable checkpoint is another reporting-period update that preserves the underlying methodology while adding later operations. A separate independent analysis would also help distinguish a robust safety signal from choices made by the system developer about benchmarks, adjustments and operating scope.
Status
Confirmed. Waymo's official blog and methodology hub document the September 24 update and the stated results. Internal confidence is medium because both sources are controlled by the company, the comparison depends on modeling choices, and this reporting did not independently reproduce the calculations.
Sources
Update note: Last reviewed 2026-09-26. We will revise this post when Waymo publishes a later reporting period or an independent replication materially changes the interpretation.
Sources
- Waymo — September 24 safety-data update — official
- Waymo — Safety Impact data and methodology — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.