Waymo's 82% Safety Claim and the Conditions That Make It True

Waymo’s autonomous vehicles crash more frequently than human drivers. They also cause far fewer injuries. Both statements are true, and the tension between them tells you more about AI safety claims than either number alone.

A peer-reviewed study published in Traffic Injury Prevention in May 2025, covering 56.7 million miles of rider-only driving, found that Waymo vehicles were involved in 82% fewer cyclist and motorcyclist injury crashes than the human baseline. Serious injury crashes dropped 92%. Intersection injury crashes dropped 96%.

Those are remarkable numbers. They are also incomplete.

The Crash Frequency Paradox

The same body of data reveals something Waymo does not emphasize in press releases. Per-mile crash frequency for autonomous vehicles runs around 9.1 incidents per million miles, compared to 4.1 for human drivers. AVs get into more than twice as many crashes. The crashes are just less severe.

Why? Waymo vehicles drive slowly. They stop when confused. They yield when uncertain. These behaviors prevent catastrophic outcomes and create minor ones. A Waymo car that brakes unexpectedly in an intersection avoids killing a pedestrian and gets rear-ended. That counts as a crash in both columns. The safety column improves. The frequency column worsens.

That may be exactly the right engineering tradeoff. But “82% fewer injury crashes” and “more than twice as many total crashes” describe the same system. Which headline you choose depends on what you want the audience to believe.

The Domain Problem

Every Waymo statistic comes from a specific operating envelope: urban streets, speed limits at or below 35 mph, pre-mapped cities, fair weather conditions. The human driving baseline includes highways, rural roads, night driving, rain, snow, construction zones, and every other condition where accidents cluster.

Comparing Waymo’s curated operating domain against all human driving is like comparing a hospital’s surgical mortality rate to the national death rate. The hospital controls who enters the operating room. The national rate includes heart attacks on hiking trails.

Swiss Re published a supporting study based on 25.3 million miles of Waymo data: 88% fewer property damage claims, 92% fewer bodily injury claims. Strong numbers. Worth noting: Swiss Re insures Waymo. Their incentives are not perfectly aligned with independent verification.

The Statistical Impossibility

In 2016, RAND Corporation calculated that autonomous vehicles would need to drive hundreds of billions of miles to demonstrate, with statistical confidence, that they are safer than humans in fatal crash scenarios. Not millions. Not tens of millions. Hundreds of billions.

Waymo has driven approximately 170.7 million miles in rider-only mode across the United States. That is 0.005% of the 3.279 trillion miles Americans drive annually, according to FHWA 2024 data. A Waymo executive acknowledged this directly: the company does not yet have sufficient mileage for statistically rigorous conclusions about fatal crash prevention.

This does not mean Waymo is unsafe. It means the sample size cannot prove what the marketing claims imply. The data is consistent with Waymo being dramatically safer. It is also consistent with Waymo not having encountered enough edge cases to reveal its failure modes. We do not know which interpretation is correct, and at 0.005% of annual driving volume, we cannot know.

The School Bus Problem

In January 2026, the NTSB opened an investigation into more than 30 school bus safety violations by Waymo vehicles across Austin and Atlanta. Separately, a Waymo vehicle struck a child in a Santa Monica school zone.

The school bus incidents are instructive. In at least one case, a remote operator incorrectly told the autonomous vehicle that no active school bus signals were present. The AV proceeded past a stopped school bus with its stop arm extended.

The algorithm worked as designed. The governance around it did not. The failure lived at the handoff between human judgment and machine execution. The same pattern appears across AI systems: the technology performs well within its training distribution, then encounters an edge case that requires human oversight, and the oversight infrastructure is not ready.

As we documented in The Verification Tax, the cost of checking AI output is often hidden until something fails visibly. Waymo’s remote operators are verification infrastructure. When that infrastructure fails, children walk in front of cars.

”Better Than Humans” Is Always Conditional

The deeper question is not about Waymo specifically. It is about a claim pattern that appears everywhere AI is deployed.

“AI is better than humans at X” is never a universal statement. It is always conditional: better at X, under conditions Y, measured by metric Z. Change any variable and the claim may reverse.

Waymo is better than humans at preventing serious injuries in low-speed urban environments with pre-mapped roads and fair weather. That is a real and valuable achievement. It is not the same as “robotaxis are safer than human drivers,” which is what headlines communicate and what most readers absorb.

This pattern repeats across the AI industry. Coding assistants that accelerate output on familiar tasks and introduce subtle bugs on unfamiliar ones. Medical AI that outperforms radiologists on specific imaging tasks and misses diagnoses outside its training data. Each claim is real within its domain. Each becomes misleading when the domain constraints are dropped from the sentence.

In Domain Expertise Still Wanted, we showed that 99% of developers still verify AI output against other sources. They have internalized what the headlines ignore: AI performance is domain-specific, and the boundaries of the domain are where failures concentrate.

The Verification Infrastructure That Does Not Exist

What would responsible evaluation of Waymo’s safety claims require?

Independent crash data, not collected or funded by the company or its insurer. Controlled comparison groups driving the same routes at the same times in the same conditions. Transparent reporting of near-miss incidents and system disengagements, not just completed crashes. A regulatory framework that defines what “safer” means before companies choose their own definition.

None of this exists at scale. The NHTSA collects some autonomous vehicle crash reports. But the reporting standards are inconsistent, the comparison baselines are poorly matched, and the sample sizes remain orders of magnitude below what statistical rigor requires.

The same verification deficit we described in The AI Verification Debt applies here, except the infrastructure is physical instead of digital. Organizations deploy AI systems, measure their own performance, publish their own studies, and let favorable headlines compound into public conviction. The conviction may turn out to be correct. It is not yet earned.

What This Means For AI Governance

Waymo is probably making roads safer in the cities where it operates. The injury reduction data is consistent across multiple studies. The engineering decisions (slow speeds, cautious yielding, restricted operating domains) are defensible and possibly optimal for the current state of the technology.

The problem is not Waymo’s engineering. The problem is the distance between what the data can support and what the public narrative claims. 82% fewer injury crashes in controlled urban conditions is a meaningful result. “Robotaxis are safer than human drivers” is a different statement, and the data does not support it yet.

For organizations evaluating any AI system, the lesson is structural. When a vendor tells you their AI is “better than humans,” ask three questions:

At what specific task? Under what conditions? Measured how?

If the answers are precise and verifiable, you are probably looking at a real capability. If the answers are vague or absent, you are looking at marketing. The conditions matter more than the claim. They always do.

---

This analysis synthesizes the Waymo/Swiss Re autonomous vehicle safety study published in Traffic Injury Prevention (May 2025), the RAND Corporation framework for AV safety validation (2016), NTSB investigation reports on Waymo school bus violations (January 2026), and FHWA Traffic Volume Trends (2024).

Victorino Group helps organizations build verification frameworks for AI systems operating in the physical and digital world. Let’s talk.