Autonomous-driving safety comparisons need the right miles

Waymo's latest reporting makes geography, time and crash definitions central to interpreting AI driving performance.

A million driving miles can represent very different exposure to risk. A vehicle operating on selected urban streets at particular times faces a different mix of conditions from the national driving population. That does not make a safety comparison impossible. It means the benchmark must resemble the actual operation closely enough for the result to have a clear meaning.

Waymo's current safety hub reports rider-only operations through June 2026. Its July 2026 methodology discussion emphasizes time and geography when constructing human-driving comparisons. The earlier rider-only study provides a separate, bounded research period. These are substantial company-provided evidence resources, but they should be read within their operating areas, outcome definitions and analytical assumptions rather than as a universal ranking of autonomous vehicles. [Current hub](https://www.waymo.com/safety/impact/), [methodology](https://blog.waymo.com/blog/2026/07/time-geo-crash-risk-effect/), [study](https://arxiv.org/abs/2505.01515).

The AI system performs several related tasks: interpret its surroundings, predict other road users' behavior, plan movement and control the vehicle. A road outcome reflects that entire system along with vehicle hardware and operational procedures. It cannot be attributed cleanly to one neural network or one impressive perception demonstration.

Begin reading a safety report by defining the denominator. Is it miles with a safety operator, rider-only miles or another category? Which roads, cities and conditions are included? A report combining different operating modes answers a different question from one restricted to driverless service. Keep the definition visible when comparing results across companies or years.

Next, examine the outcome. Property damage, reported injury, airbag deployment and serious injury are different categories. They have different reporting behavior and statistical frequency. A favorable result in one category should not be silently restated as a proven reduction in every other category. Rare outcomes also require enough exposure to support meaningful uncertainty estimates.

NHTSA's crash-reporting page explains the agency's reporting framework and interpretation limits. Raw reported crash counts are not a fair risk comparison without exposure and operating context. Reporting criteria and completeness can differ from ordinary human-crash records. An analyst needs those differences before treating a table as evidence that one system is safer. [NHTSA framework](https://www.nhtsa.gov/laws-regulations/standing-general-order-crash-reporting).

An illustrative independent analysis would obtain the authorized datasets and preserve their definitions. Reproduce the published comparison before adding a new one. Match time and location where the available evidence permits, show uncertainty and explain remaining mismatches. Avoid selecting a favorable comparison after examining many alternatives without describing the search.

Safety readiness and retrospective outcomes are also separate evidence. A deployment decision needs prospective checks of intended operating conditions, hazards and response procedures. Later road results help evaluate whether those expectations held. Neither one alone answers every question, and a retrospective favorable average does not eliminate the need to investigate individual failures.

For a city considering a service, assess interactions beyond the passenger trip. Review the operational approach to unusual road conditions, emergency vehicles, blocked routes and remote assistance. Those questions need evidence specific to the deployment; the cited safety comparisons do not establish every capability. Availability must also be checked for the actual service area rather than inferred from a national announcement.

Costs have an equally specific boundary. A passenger fare, fleet operating expense and the total cost of deploying a driverless system are different quantities. The sources here do not establish a universal price or economic advantage. A mobility evaluation should also examine access and service reliability under the intended local use.

The public explanation should make uncertainty understandable. If a service is unavailable for a route or condition, state that clearly and offer the appropriate alternative. A person should not be encouraged to assume that a favorable safety statistic guarantees every possible trip. Product communication needs to match the operating limits of the service.

Autonomous-driving evidence becomes more useful when the comparison is narrow enough to be fair and transparent enough to reproduce. The question is not whether a large number sounds impressive. It is what happened under the stated conditions, how that differs from the relevant benchmark and which claims the available miles still cannot support.