So it might come down to how many "9s" you're comfortable with. The experience is really good 99.999% of the time until it's not, and that "not" could be catastrophic. I suppose the data engineers are quite confident in the 9s.
Waymo is 100% with zero fatalities.
But then again, the Concorde was the safest airplane ever built for nearly 30 years, until its first crash and then it was the most dangerous passenger jet ever with 12.5 fatal events per million flights.[1]
Lies, damned lies, and statistics.[2]
0: https://assets.ctfassets.net/vz6nkkbc6q75/3yrO0aP4mPfTTvyaUZ...
1: https://www.airsafe.com/journal/issue14.htm
2: https://en.wikipedia.org/wiki/Lies,_damned_lies,_and_statist...
I imagine it also goes without saying that not every mile of road is equally risky, and I have to imagine that Waymo's miles traveled probably on far less risky roads on average given the way they've been rolled out (which isn't a bad thing, but it does make extrapolations from the data about relative safety a bit more dubious).
Unfortunately stats can be spun in whatever way you need them to in order to support or hinder an effort.
To use your example, risk per mile is a sliding scale, not necessarily a Boolean value. So then someone could conceivably draw the line on that scale wherever most benefited the company they were trying to support.
Any different than with a human taxi driver?
It's not about absolute reliability, it's about how well it compares to the alternative, which is human taxi drivers. And the thing is, you don't hear about human car accidents because it's so common that it's not worth making the news.
Another very interesting thing about robotaxis is agency and blame. Taxi driver had an accident? Just that driver is suspect. Robotaxi had an accident? They're all suspect.