Autonomous vehicles will be safer than human drivers

The claim asks whether mature vehicles operating without human control will cause fewer crashes, injuries, and deaths per mile than comparable human-driven vehicles—not whether automation can eliminate every collision or whether every present-day automated system is already superior. The central evaluative problem is to distinguish demonstrated safety within restricted operating domains from a prediction about mature autonomous vehicles more generally, while comparing like exposures and giving appropriate weight to both common crashes and rare fatal events. The evidence supports a balanced judgment: favorable real-world results show that superior performance is achievable, but restricted deployment, difficult benchmarking, rare-event uncertainty, and novel failure modes prevent a categorical class-wide conclusion. The strongest support is direct evidence that at least one driverless service has recorded fewer injury-causing and police-reported crashes than geographically adjusted human benchmarks in actual passenger operations. Across 7.1 million rider-only miles, Waymo's analysis reported substantially lower rates than matched human drivers, including estimated reductions of 85% for injury-causing crashes and 57% for police-reported crashes (see Figure 1). IIHS treated the mileage comparison as credible evidence that Waymo's driverless vehicles crash less often than human drivers where they operate, which provides independent institutional support for the direction of the developer-affiliated estimates. The agreement between matched crash analyses and an independent observer is more probative than simulation or engineering aspiration alone, although the supplied evidence does not substantiate the separate 56.7-million-mile comparison depicted in Figure 2. A second supporting line is mechanistic: autonomous operation can remove perception failures and incapacitation that account for a meaningful portion of human crashes. IIHS analyses estimated that eliminating perception errors and driver incapacitation alone could prevent roughly one-third of reviewed crashes, establishing a substantial potential safety floor before any benefit from better prediction, planning, execution, or risk choice. The combination of an identified causal pathway and observed lower crash rates strengthens the case beyond either line alone: the mechanism explains why improvement is plausible, while deployment data show that improvement has occurred in at least one bounded setting. The strongest objection is that available road mileage cannot yet establish superior fatality performance with confidence, even when ordinary crash and injury results are favorable. RAND estimated that demonstrating human-level fatality rates through road testing alone could require hundreds of millions to billions of miles and, for conventional test fleets, potentially decades or centuries. Accordingly, reassuring performance over 7.1 million rider-only miles can materially support claims about more frequent crash outcomes while remaining statistically compatible with uncertainty about rare catastrophic failures. Comparative inference is also vulnerable to unequal exposure and reporting rather than merely insufficient mileage. NHTSA warns that public automated-driving incident counts differ in fleet size, mileage, system use, operating domain, reporting capability, completeness, and possible duplication, so raw totals do not establish comparative crash rates. Matched per-mile analyses are therefore more informative than database counts, but their validity still depends on the quality of geographic, road, traffic, weather, severity, and reporting adjustments. Automation also substitutes technical and systemic risks for some human errors rather than simply subtracting risk from the road system. A peer-reviewed collective case study of fatal automated-system crashes identified recurring problems involving system limitations, human oversight, organizational controls, and unsafe deployment assumptions. Because software, design, and deployment assumptions may be shared across a fleet, a systemic defect presents a different risk profile from an isolated human mistake, even if its observed frequency is initially low. Finally, superior sensing alone is insufficient to establish the claim because most reviewed human crashes were not attributable solely to perception failure or incapacitation. IIHS estimated that only about 34% of crashes in its sample fell into those categories and concluded that preventing most crashes would require autonomous systems to outperform people in prediction, planning, execution, and risk-taking decisions. Thus, immunity to intoxication, fatigue, or distraction does not by itself prove net superiority; the automated decision stack must also avoid its own planning and risk-selection errors. The evidence is most favorable when the claim is limited to a validated system operating inside the roads, cities, speeds, and other conditions for which it was designed and tested. The leading favorable results concern Waymo in selected operational domains, and IIHS expressly cautions that they do not establish the safety of all autonomous vehicles or performance under every condition. Those results demonstrate local superiority, not reliable generalization to snow, rural roads, construction zones, emergency scenes, or other unfamiliar environments. Safety superiority must be defined comparatively rather than as the elimination of every collision. Case analysis shows that some crashes are physically unavoidable once another road user creates a hazard too late for braking or evasive action, even for an automated vehicle. The appropriate endpoint is therefore fewer and less severe crashes per comparable mile, not zero crashes or faultlessness in every incident. Aggregate involvement rates may also conceal a redistribution of crash types and responsibility. An early California study found that many crashes involving autonomous test vehicles were rear-end impacts in which human-driven vehicles struck automated vehicles that were stopped or moving conservatively. This suggests that cautious robot behavior can reduce some risks while creating unusual interactions with human road users, making fault and injury severity important alongside simple crash involvement. The principal gaps concern generalization, rare fatal outcomes, benchmark comparability, and the independence of the most favorable quantitative evidence. The bundle does not provide broad, standardized per-mile results across multiple autonomous developers, operating domains, weather regimes, road classes, and mature vehicle designs. It also does not resolve how to validate rare fatality risk without infeasible amounts of public-road exposure, or how to detect low-frequency software failures that may be correlated across a fleet. Developer affiliation in the core performance analyses remains an unresolved conflict-of-interest classification, although independent IIHS assessment partially mitigates rather than eliminates that concern. Finally, the frozen evidence bundle contains no admissible evidence for the figure concerning Level 2 partial automation and no evidence for the specific 56.7-million-mile study represented in the supplied figures, so neither can carry weight in the conclusion (see Figure 3). On the present evidence, the claim is plausible and demonstrated for at least one mature driverless system within restricted operating domains, but it is not yet established for autonomous vehicles as a general class or for deaths as confidently as for more frequent crash and injury outcomes. Confidence in this balanced assessment is high because strong institutional and peer-reviewed evidence supports both the favorable bounded finding and the principal limitations. The dominant uncertainty is whether favorable results from a leading, developer-evaluated deployment will generalize across systems and conditions while preserving superiority for rare catastrophic outcomes; unresolved conflict-of-interest classification intensifies that uncertainty.