So, here is the thing about those Waymo safety reports. It is one thing to see a press release, but it is another to dig into the actual methodology. Misryoum editorial desk noted that the benchmarks for human driving performance—the ones used to stack up against autonomous vehicles—aren’t just pulled out of thin air. They rely on state police reports and total miles driven across cities like Phoenix, San Francisco, Los Angeles, and Austin.
It is tricky business. You can’t just compare a car driving down a chaotic urban strip to someone cruising on a quiet suburban road, right? The smell of ozone and hot asphalt comes to mind when I think of those busy San Francisco streets. To fix this, Misryoum reporting highlights that researchers are now using a spatial reweighting method. Essentially, they are trying to force the human data to match the specific, often grittier, areas where Waymo actually operates.
Basically, the math got an update. The latest benchmarks, as noted in Misryoum analysis, now account for the fact that not every street in a city is equally dangerous. By using models to adjust for where these vehicles spend their time, the comparison becomes—well, a little more honest, I guess.
There is also the issue of underreporting. It is a known fact that not every fender bender ends up in a police database. The team applied a 32% correction factor for minor injury reports to try and bridge that gap. For the more serious stuff—airbag deployments and major injuries—they kept the raw numbers as they were. It’s not perfect, but it’s the best we have until the RAVE best practices fully take root across the industry.
Wait, I should clarify—they aren’t looking at freeways. The data specifically cuts those out to focus on the environments where these cars are actually doing the heavy lifting. It’s a specific slice of the pie. Or maybe not the whole pie, but a representative slice.
It’s a work in progress. While the methodology (Misryoum notes) is getting sharper with these spatial adjustments, there is always going to be a limitation to how accurately we can map human error against machine performance. We are getting better at the measurement, but the road ahead is still pretty long.