The comparison problem
"Safer than a human" sounds like a simple claim and is not. Human crash rates vary enormously by road type, city, time of day and how crashes are counted; police records miss most minor collisions. A robotaxi drives only within its service area, in the conditions its operator allows, and reports every scrape. A fair comparison needs human data from the same streets, the same hours and the same reporting threshold, and it needs enough miles for rare, serious events to show up at all.
What the operators publish
Waymo has published the most: a running dataset of its driverless miles and crashes, and peer-reviewed comparisons, some with insurer Swiss Re, against human drivers in the same areas. Its figures show large reductions in injury crashes and in crashes where its vehicle was at fault, over tens of millions of miles. The studies are the company's own, and critics note that; independent researchers have broadly supported the direction while arguing about the size. Other operators publish far less.
Disengagements and crash reports
California requires companies testing with safety drivers to report disengagements, the moments the human took over. The figures were once treated as a league table and are widely misread: companies define a disengagement differently, and a low number can mean a mature system or a cautious test plan. Since 2021 NHTSA's Standing General Order has required crash reports from every operator of automated driving and Level 2 systems, producing the first national dataset, though one with uneven detail and heavy redaction.
What remains unknown
The rarest events, fatal crashes, are still too few in driverless miles to estimate a rate with confidence. New cities, new weather and new vehicle types reset the evidence. And the question that matters to most people is not the average but the failure mode: whether the car makes mistakes a human never would, as the Cruise dragging incident suggested it might. The honest summary is that the best operators have shown fewer injury crashes than humans in their areas, and that the record is still short.
Frequently asked questions
Are self-driving cars safer than humans?
In the areas where they operate, Waymo's published data shows substantially fewer injury crashes per mile than human drivers on the same roads. The data is the company's, the miles are still limited, and other operators publish much less, so the general question is not settled.
How many people have self-driving cars killed?
The 2018 Uber test-vehicle crash in Tempe, Arizona killed a pedestrian, and several fatal crashes have involved Tesla Autopilot or FSD with a human driver responsible. No fully driverless robotaxi in commercial service had been found responsible for a fatality as of this writing.