Waymo states that its driverless cars were involved in 94% fewer crashes with serious injury than human drivers in the same regions. The number is the company’s own, about its own product. On the same page where it publishes the number, it also makes available the files that allow the calculation to be redone.
That combination is what makes the claim readable. A safety percentage released by the party being measured is not a fact: it is a supplier’s claim. It becomes useful information when it comes with the base, the comparison, and what was left out, and it fades again when it is repeated alone, with none of that.
This post is an analysis of public sources, not an account of an operation or of a vehicle test. The numbers are presented as in the original publication, with a verification date, and the conclusions about how to read them are editorial interpretation.
The same indicator appears with two values in five months
In a check made on August 28, 2026, Waymo’s safety dashboard reported 220.6 million driverless miles driven, about 355 million kilometers. The period covered runs from September 2020 to March 2026, in five cities: Phoenix, San Francisco, Los Angeles, Austin, and Atlanta. On that base, the company reported 94% fewer crashes with serious injury or worse, 82% fewer crashes with airbag deployment, 82% fewer crashes with any injury, and 93% fewer pedestrian strikes with injury.
On March 19, 2026, a statement from the same company presented 170 million fully autonomous miles and, on them, 92% fewer serious or fatal crashes, 83% fewer airbag deployments, and 82% fewer crashes with any injury.
The two sets are consistent with what each publication states, and they describe bases named in different ways. What they show is something else: the percentage tracks the base it was calculated on and the moment of the cut. Repeated without those two data points, it ceases to be comparable, including with itself.
First question: what is the denominator
Every safety percentage is a ratio. The numerator usually shows up in the headline; the denominator almost never does.
In Waymo’s case, the denominator is published: driverless miles traveled, broken out by city. That is 80.551 million in Phoenix, 67.078 million in the San Francisco Bay Area, 51.816 million in Los Angeles, 15.789 million in Austin, and 5.379 million in Atlanta. The concentration matters. A fleet whose mileage is mostly in three cities in the southern and western United States does not describe traffic just anywhere, and the company itself does not claim that it does.
When the denominator is not published, there is nothing to ask afterward. The calculation cannot be redone, and comparison with any other number becomes impossible.
Second question: compared with which human driver
The human reference is also constructed, and the construction changes the result.
Waymo states that its reference comes from police crash records and from mileage data. They are restricted to the counties where it operates and consider only passenger vehicles, on the road types where the fleet drives. The spatial distribution is adjusted by a method published by Chen and others, in 2024.
This is more rigorous than comparing against the national average, and that is worth recognizing. It is also narrower: the human driver in the comparison is the driver in those counties, on those roads. It is not the Brazilian driver, nor the highway driver.
Third question: which piece of the world was left out
No system is evaluated under every condition, and the cut should be stated. In this case, it is.
The company states that the analysis considers only urban roads, because the fleet still has limited mileage on highways. The crashes and the mileage in the human reference exclude highways for the same reason. It also records that none of the cities where it operates has significant snow, so neither the fleet data nor the reference data include that kind of condition. And it acknowledges that time of day affects the crash rate, without enough granularity to adjust for that factor.
There are three stated exclusions: highway, snow, and time of day. They do not invalidate the result. They define where it holds, and that is exactly what is lost when the percentage travels alone into a headline.
Fourth question: which corrections went into the calculation
Statistical correction is legitimate and common. What cannot be missing is knowing which one went in, and where.
The reference for crashes with any injury receives a 32% correction for underreporting, based on national data, because minor crashes often do not reach a police record. The references for serious injury and for airbag deployment do not receive that correction: they use the observed crashes.
The practical consequence is that the four published percentages were not produced by the same procedure. Comparing them with one another as if they were interchangeable is a mistake that careful reading avoids.
What distinguishes this company is not the percentage
Waymo publishes four open-format data files: driverless mileage by location, crashes with the identifier of the mandatory report sent to the U.S. regulator and the outcome group they belong to, collision counts compared with the references, and the geographic distribution of mileage. It also publishes release notes that describe methodology changes between updates, plus a list of peer-reviewed studies.
A detail worth noting shows up only when you look at the files: three of them cover the period through March 2026, and the geographic-distribution one covers through March 2025. The lag is stated on the dashboard itself, and anyone who redoes the calculation needs to take it into account.
The identifier of the report to the regulator is what allows the strongest verification available. With it, a third party can link each crash the company counted to the corresponding record in a public database, instead of accepting the count as it arrives.
Publishing the percentage costs one line. Publishing the denominator, the construction of the reference, the exclusions, the corrections, the methodology notes, and the raw files costs exposure. Each of those pieces of information is a point through which the calculation can be contested. A safety claim that offers none of those points does not become more trustworthy for that reason. It only becomes less verifiable.
What the check does not resolve
Redoing the calculation proves arithmetic, not safety.
The verification confirms that the published numbers are consistent with the published data. It does not evaluate what was not measured, it does not cover the conditions excluded from the cut, and it does not say how the system behaves on roads, in climates, and at times of day outside the sample. It also does not replace a regulator’s assessment, which has access to information that is not public.
The gain from careful reading is more modest and more solid: separating the claim that can be checked from the one that can only be accepted.
What remains
A safety percentage alone tells almost nothing. With the denominator, the reference, the exclusions, and the corrections alongside it, it becomes a comparable number, and that is how it should be required, whoever it comes from.
The four questions do not depend on autonomous cars. They serve any performance claim or safety claim in which the party that publishes the result is also the party being measured.
Sources
- Safety dashboard with 220.6 million driverless miles between September 2020 and March 2026, the percentages by outcome type, the construction of the human reference, the exclusions for highway, snow, and time of day, the 32% correction for underreporting, and the four open-format data files — Waymo, Safety Impact, the company’s own source about its own product, verified on 2026-08-28.
- Statement of March 19, 2026, with 170 million fully autonomous miles and the percentages of 92%, 83%, and 82% — Waymo, safety update, the company’s own source, verified on 2026-08-28.
- List of peer-reviewed studies on fleet safety, including a 2026 publication in the journal Traffic Injury Prevention — Waymo Research, the company’s own source, verified on 2026-08-28.
- Proposed criteria for determining the deployment readiness of an automated driving system, authored by researchers at the company — Favaro and others, arXiv 2505.09880, open-access academic publication, verified on 2026-08-28.
