The null experiment is the human driver of that particular Tesla, not an average of all drivers
This particular driver isn't what makes the story special or newsworthy. The nature of the incident is. Our choice of statistical cohort should reflect this data selection effect, ie "all Autopilot miles driven," not "all miles driven by <person's name>."
I'm a partner in an auto insurance company, and most of our no-accident records will stay that way, statistically speaking.
If this driver had a one in a billion miles chance of crashing and Autopilot (with an inattentve driver, as they all tend to be) had a one in a million miles chance, then Autopilot is decreasing safety.
At a certain level of safety, Autopilot would make everyone safer, but it's definitely not there yet.
Fatal human accidents are on the order of 1-10 every billion miles of driving in the US, yes. Tesla Autopilot crashes seem to be much, much, much more common.