A much larger real-world exposure set
Waymo published an updated safety analysis on June 24, 2026 covering more than 220 million miles of fully autonomous driving through March 2026. The data span five U.S. operating geographies, with Atlanta included for the first time. These are rider-only miles: the automated driving system performs the driving task without a safety driver behind the wheel.
Compared with human drivers in matched locations and periods, Waymo reported 94 percent fewer crashes involving serious or fatal injury, 82 percent fewer crashes with an airbag deployment and 82 percent fewer crashes involving any reported injury. It also reported fewer injury crashes involving pedestrians, cyclists and motorcyclists. The comparisons count crashes regardless of which road user was at fault.
These are large and consequential figures, but they are not simply the number of crashes divided by U.S. national mileage. The analysis constructs human benchmarks for the same operating areas and adjusts for differences between police-reported data and the records available for an automated fleet.
Why comparison design matters
Minor crashes are underreported in human-driving databases. An automated-vehicle operator may record events that would never reach police or insurance systems, making a naive comparison unfair to the robot. Geography matters too: urban streets, weather, vehicle mix and trip time change crash exposure. The chosen outcome also matters because a bumper scrape and a serious injury should not be treated as the same safety signal.
Waymo's data hub documents its benchmarking approach and links it to a 2025 peer-reviewed study in Traffic Injury Prevention. That paper examined 56.7 million rider-only miles and found statistically lower rates for injury-reported, airbag-deployment and suspected-serious-injury-or-worse crashes. The 2026 update applies the method to a much larger dataset, but the update itself is a company publication rather than a new peer-reviewed paper.
Statistical uncertainty does not disappear at 220 million miles. Fatal and serious crashes are rare, so the confidence range for the most severe outcomes is wider than for common events. Service also operates within defined areas, routes and conditions. Results there should not be assumed to apply unchanged to every road, snowfall, rural environment or vehicle platform.
A deployed autonomous system, not a driver-assistance feature
Waymo's system combines cameras, lidar, radar, maps and onboard computing. It plans and executes the complete driving task inside its operational design domain. This differs from consumer driver-assistance systems that require a human to supervise continuously.
Commercial operation supplies exposure to unusual interactions that closed-course testing cannot enumerate. Simulation and structured testing remain important for rare cases, while the road fleet provides outcome data. The combination creates a feedback loop: identify patterns, update the system, validate the release and continue monitoring after deployment.
What the evidence supports
The update provides evidence that this specific system, in the places and period measured, had substantially lower rates for the selected crash outcomes than the constructed human benchmarks. It does not prove that every autonomous vehicle is safer than every human, nor does it answer all questions about congestion, accessibility, emergency response or behavior outside Waymo's service areas.
Waymo develops the technology and publishes much of the underlying analysis, so independent review and public reporting remain valuable. The peer-reviewed 2025 paper strengthens the methodology, while regulators and outside researchers need sufficiently detailed, consistent data to test broader claims.
The practical milestone is scale with measurable outcomes. Autonomous driving has moved beyond a handful of demonstrations in these service areas. Safety assessment can increasingly use real injury, airbag and vulnerable-road-user events, while still stating the geographic and statistical limits of the comparison.