Waymo’s latest safety numbers make a striking claim: its fully driverless vehicles are involved in fewer serious crashes than human drivers in the same markets. Independent research from the Insurance Institute for Highway Safety points in the same direction, but the evidence is more nuanced than a single percentage can show.
Through June 2026, Waymo reports 271.3 million rider-only miles across Phoenix, the San Francisco Bay Area, Los Angeles, Austin and Atlanta. The company says its vehicles were involved in 82% fewer injury-causing crashes and 95% fewer serious-injury-or-worse crashes than human-driver benchmarks. Those are company calculations, so they should not be read as independent proof by themselves.
What did the independent IIHS study find?
IIHS researchers looked at police-reportable crash involvement rather than every minor contact recorded by automated-vehicle reporting systems. Their July 2026 study found that Waymo vehicles in driverless operation had a police-reportable crash involvement rate 68% lower than human drivers operating in the same cities and years.
The researchers also found a lower injury-crash rate. That is important because raw automated-vehicle crash databases can exaggerate minor events compared with human-driver databases: autonomous fleets face stricter reporting requirements, so a scrape or low-speed contact may appear in one dataset even when a human driver would never have called police.
Why 68% and 82% are not contradictory
The figures answer different questions. IIHS focused on a cleaned historical set of police-reportable crashes and compared Waymo with human drivers over a defined study period. Waymo’s newer dashboard uses a larger and more recent mileage base and includes different outcome categories such as injury-causing crashes and serious-injury-or-worse crashes.
When two percentages come from different definitions, time periods and datasets, they should not be combined into one “true” number. The useful conclusion is narrower: both Waymo’s own analysis and independent IIHS research show substantially lower crash involvement in the categories they measured.
Does that mean robotaxis are proven safer everywhere?
No. The evidence applies to the specific Waymo system, operating areas, road types and time periods studied. IIHS explicitly noted that results from Waymo cannot automatically be applied to every autonomous-driving system. A company with different sensors, maps, software, operational limits or safety processes could perform very differently.
City-by-city variation also matters. IIHS reported strong reductions in several markets, while Austin’s estimate was less favorable and based on a smaller sample. As deployments expand, safety needs to be measured continuously rather than assumed from one national headline.
What about Waymo recalls in 2026?
Lower overall crash rates do not mean the technology has no serious failure modes. In 2026, Waymo reported safety recalls involving unusual edge cases. One recall involved vehicles entering or continuing into freeway construction zones after failing to respond correctly to closures. Another addressed behavior around standing water on higher-speed roads.
NHTSA documents show that these were software-related problems addressed through operational restrictions, mapping changes and software remedies. This is exactly why safety evaluation has to include both aggregate crash statistics and detailed failure analysis. A system can perform well on average while still needing fixes for rare but potentially dangerous situations.
What can human drivers do that robotaxis may struggle with?
Humans can improvise when road layouts become strange: a police officer gives hand signals, a construction worker waves traffic around a blocked lane, floodwater hides lane markings or a temporary barrier appears where the map expects an open road. Humans can also make terrible decisions, drive impaired, become distracted or fall asleep.
Autonomous systems exchange one set of weaknesses for another. They can maintain attention continuously and react consistently, but unusual environments, conflicting signals or sensor limitations can expose software assumptions that are difficult to test in advance.
Why the comparison with human drivers is difficult
Crash databases were built for human driving, not large fleets of autonomous vehicles. Human drivers underreport many low-severity incidents. Autonomous fleets may be required to report a wider set of contacts. Mileage exposure also differs by road, city, weather, time of day and traffic density.
IIHS researchers had to manually review crash narratives and decide which autonomous incidents would probably have been reported to police if a human had been driving. That improves comparability, but it also shows why simple totals can mislead.
What should a rider care about?
If you are considering a driverless ride in Phoenix, Los Angeles, San Francisco, Austin or Atlanta, three questions are more useful than asking whether the technology is “perfect.” First, does the service operate commercially in the specific area and conditions you need? Second, what happens when the vehicle stops, reroutes or encounters an unusual scene? Third, how can you contact remote support or emergency services if something goes wrong?
Riders should also follow normal seat-belt rules and avoid treating an autonomous cabin as a place where safety no longer matters. Automation changes who controls the car; it does not change the physics of a collision.
What should regulators measure next?
IIHS has called for better standardized data, including vehicle miles traveled and more consistent crash reporting. That matters because a safe system today still has to remain safe as it moves into new cities, weather patterns, road designs and larger fleets.
Regulators also need enough detail to separate fleet-wide performance from one recurring defect. If a system repeatedly mishandles construction zones, emergency vehicles or standing water, a strong overall crash rate should not hide that pattern.
So, are Waymo robotaxis safer?
The best current evidence supports a careful answer: in the studied markets and crash categories, Waymo’s driverless vehicles have shown lower crash involvement than human drivers. Independent IIHS research strengthens that conclusion. But the result is specific to Waymo, depends on definitions and operating conditions, and does not eliminate the need for recalls, oversight and better reporting.
The strongest case for autonomy is not that a robotaxi never makes a mistake. It is that, over enough comparable miles, it may make fewer harmful mistakes than humans. The 2026 data is encouraging, but continued measurement will determine whether that advantage survives much larger deployment.
Sources: IIHS, July 2026, Waymo Safety Impact, and NHTSA recall documentation. More transport analysis is available in Mobility.
