Third fatal Tesla Autopilot crash renews questions about the system

Imo the cases where autopilot fails should be a subset of where humans fail. If it's failing in places where humans wouldn't then there's an issue.

the problem with stats is that incident rates bunch together extremely heterogeneous driver, vehicles and driving environments, so for example the "40% safer" figure from the ntsb years ago is extremely misleading

since the autopilot has a very limited field of applicability the only way to determine if it's safer is to compare it with actual driving data details between the tesla car fleet within similar condition and I think the only data we can get as outsiders it's "number of death in tesla on highways" and "number of death in tesla on highway with autopilot", but deprived of relative miles driven in each mode information it's going to be quite meaningless, and even with the data at hand it might very well be that the datapoint are too few to reach a significant confidence about the findings.