You're jumping to a quantitative comparison because that's simpler and more obvious, but the right way to think about it is qualitatively.
We have over half a century of understanding of modern human driving patterns. We know the ceiling on the number of accidents/deaths that humans will cause. We know that if it changes, it will change slowly.
AI works differently. An update to code or models wipes out all expectations we've learned from prior versions.
In short, the entire fleet of autonomous vehicles can change behavior overnight with a software update and become murder tanks.
Nothing will ever happen to all human drivers at once. Some small percentage will be drunk at any given time. We'll never all wake up with a new drunk version of ourselves at the same time.
I don't mean to say that autonomous technology could never work in principal. But we know that this SV in practice has too much hubris and too little oversight to deploy sensitive products responsibly, or with attention to any values above profit and growth.