It actually would be really hard to get an unbiased estimate of safety given the current systems, because the data is inherently cherry picked by drivers who can switch the feature on/off depending on how complex the driving task is. What a simple number like crashes per mile really measures is really how likely FSD drivers are to overestimate the system’s ability plus some unknown base rate of unavoidable accidents.

Probably the only way to control for this is looking at cars that are fully autonomous door to door and aren’t limited to pre-selected roads/areas. I don’t know that anyone is even doing that sort of testing.