Waymo Safety Claims: Read Past the Percentage

Key takeaways

• Crash comparisons need matching mileage measures and consistent definitions of what counts as a crash.
• Safety results apply to the roads, weather, and operating conditions studied.
• Serious injuries and deaths deserve separate attention from total crash counts.
• External review matters, but its value depends on what reviewers actually verified.

Before you climb into a driverless car, “safer than a human” is a reasonable bar to set. For a service like Waymo, a crash-reduction percentage might seem to settle the question. The real work starts with the comparison behind it.

Start with the denominator

A fleet that drives farther has more opportunities to get into crashes. Comparing raw crash totals tells you little unless you also know how much each group drove.

That makes crashes per mile a useful starting point. It puts driving exposure into the calculation.

But the counting rules need to match, too. If one dataset records minor bumps while another includes only police-reported crashes, the resulting rates measure different things. Insurance claims introduce another boundary: crashes that never generate a claim can be absent from the analysis.

So when you read “fewer accidents,” pause on the noun. Does it mean fewer collisions, fewer crashes involving injuries, or fewer insurance claims?

Each can tell you something useful. None is automatically interchangeable with the others.

Compare the same driving job

Driving through busy city streets and cruising along a quiet road are different assignments. So are a clear afternoon and a night of heavy rain.

An autonomous system’s intended operating conditions are called its operational design domain, or ODD. That can include geography, road types, and weather.

The term sounds bureaucratic. The underlying question is practical: where, and under what conditions, is this system designed to drive?

Comparing a robotaxi operating within a defined area with the national average for human drivers can mix two effects: driving performance and the difficulty of the environment. A stronger comparison uses human driving data from similar locations, road types, and conditions wherever possible.

Restricting service to conditions a system can handle can itself be a safety measure. The catch is how far you extend the conclusion. Results from one operating environment do not establish performance on snowy roads or in other conditions outside that environment.

The boundaries belong alongside the headline.

Count the harm, not just the crashes

A scraped bumper and a crash that seriously injures a pedestrian can each count as one collision. Their consequences are vastly different.

Total collision rates therefore need company: injury-crash rates and rates of crashes involving serious injuries or deaths. The assessment should cover people outside the vehicle, including pedestrians and cyclists, as well as passengers.

There is another distinction worth keeping clear: crash involvement versus fault. A driverless car struck from behind may appear in crash statistics even when another driver caused the collision. Fault-based measures and overall collision measures answer different questions.

Severe crashes also create a statistical problem because the observed numbers may be small. No deaths over a limited driving distance does not establish zero fatality risk.

Look for how much mileage supports the estimate and how uncertain that estimate remains. A reassuring number can still leave a wide range of possible outcomes.

Ask what the reviewers actually checked

A company’s own safety analysis can contain useful evidence. Assessing it starts with whether the methods, counting rules, and comparison group are disclosed.

Outside involvement deserves the same attention to detail.

If an article cites the Insurance Institute for Highway Safety, or IIHS, check what was evaluated. A test of a particular crash-avoidance feature provides different evidence from an analysis of a driverless service’s real-world crash rate.

The reviewer’s role matters, too. An outside team might analyze data supplied by the company, or it might obtain data independently. Those arrangements support different levels of independent verification.

Likewise, peer review does not by itself mean someone independently checked the underlying raw records. The label is useful; the scope of the review tells you more.

For a passenger deciding whether to use Waymo, the question is how much harm the service reduces under conditions that match the intended trip. A large reduction deserves attention, but so do the roads, weather, and evidence behind it. A safety claim should travel only as far as its evidence does.