Autonomous vehicles will be safer than human drivers

What's this about?

People disagree about whether self-driving cars will cause fewer crashes than people do. Early results look good, but only in some places and conditions.

What supporters say

• Waymo’s driverless taxis have had fewer crashes than human-driven cars in the same service areas.
• These results come from real streets, not just test tracks or computer models.
• Self-driving systems do not get tired, drunk, distracted, or angry like human drivers can.
• Companies can use fleet data to spot problems and send tested software fixes to many cars quickly.

What critics say

• The best safety results so far come from only some driverless taxi services, especially Waymo.
• We do not yet know if every self-driving vehicle will stay safer on every road.
• These cars work within set areas and conditions that their makers designed them to handle.
• Safer driving depends on good data, careful tests, and companies releasing updates in a safe way.

The bottom line

Evidence suggests some Waymo driverless taxis now crash less often than similar human-driven cars. But we cannot yet say all self-driving cars will be safer than humans everywhere.

Autonomous vehicles may eventually reduce crashes, injuries and deaths compared with human drivers. But the evidence so far supports that conclusion only for some carefully limited driverless services, not for every automated vehicle on every road.

The case for

The strongest evidence comes from Waymo’s driverless taxi service, which has recorded lower collision rates than comparable human-driven vehicles in the areas where it operates. Studies comparing similar locations, times and driving conditions found lower risks across several kinds of crashes. An analysis reviewed by the Insurance Institute for Highway Safety also found lower rates of police-reported and injury-related crashes for Waymo robotaxis in comparable service areas. 2

That matters because these are results from real streets, rather than simulations or closed-course tests. A Waymo One case study likewise found lower observed crash rates in several severity categories. A peer-reviewed study reached similar findings after accounting for where and when vehicles were driving, suggesting the advantage was not simply the result of choosing easier routes.

There is a straightforward reason automation could help. Human drivers become tired, distracted, impaired or impatient; a properly functioning automated system does not. It can continuously monitor its surroundings and follow driving rules. Reconstructed simulations of fatal crashes have suggested that Waymo’s system could have avoided or reduced the severity of many incidents within the areas and conditions it was designed to handle. 1

Automated fleets may also improve faster than individual human drivers. Their operating data can identify safety-relevant events, and software updates can potentially distribute validated fixes across many vehicles at once. That creates a plausible path toward safer driving over time, though it depends heavily on reliable data, careful testing and responsible deployment. 3

The case against

The central problem is that the evidence is narrower than the headline claim. Positive results concern a small number of mature, geofenced services operating in specific cities and under defined conditions. They do not show that all autonomous vehicles—including privately owned cars, systems made by other companies, or vehicles facing snow, severe weather and unfamiliar roads—will be safer than people. 4

Even the strongest studies found that performance was not better in every type of crash. Some unusual lighting conditions and turning situations remained exceptions. Research on “edge cases”—rare, confusing or highly unusual road scenarios—warns that infrequent failures can still be serious, and cannot simply be dismissed because average crash rates look favorable. 6

Fatal crashes present a particular challenge. They are rare, which is good news for the public but makes it difficult to prove a small safety advantage with ordinary road testing alone. Researchers say credible comparisons need huge amounts of mileage and close matching of routes, weather, traffic, vehicle types and crash definitions. RAND has argued that testing on public roads should be combined with simulation, modeling and other forms of validation, rather than relying on early crash totals as final proof. 5

The available data also have limits. Government crash reports are valuable, but they do not amount to a randomized contest between people and machines. Manufacturers hold important operational data, but their methods and commercial interests mean that independent replication remains important. There is still no universally standardized, independently audited system for comparing safety outcomes across companies and conditions.

Driver-assistance systems offer a separate warning. When a human is expected to take over from a partly automated car, complacency, confusion about the system’s limits and overtrust can create new hazards (see Figure 3). This does not directly show that fully driverless services are unsafe, but it does show why “automation” is not one single safety category. 7

The bottom line

The evidence strongly supports a qualified conclusion: certain mature driverless taxi services appear safer per mile than comparable human driving within their approved operating areas. Real-world Waymo studies provide meaningful support for that narrower claim. 2

But the broader prediction—that autonomous vehicles in general will ultimately be safer than human drivers in all settings—remains unproven. The biggest unanswered question is whether early gains will hold across different systems, regions, weather, road types and rare emergencies. Confidence is high in this cautious assessment: automation has shown real promise, but the evidence does not yet justify a universal verdict.