Seems impossible to tell without comparable data on humans. I'd assume 99%+ of all miles are still by humans? Unless you have that base rate who knows.
In 2022, human drivers in San Francisco killed 20 pedestrians and another ~25 people in vehicles (probably motorcycles and scooters). I think one cyclist.
If the human to software driven mile ratio is greater than 50, then the expected number of fatalities for software driven miles would be <1 right now. So the data is insufficient.
