[Someone]

On similar roads, in similar traffic, in similar weather.

If you look at fatalities or define “accident” as cases where somebody had to visit a doctor or something like that, you also need to compare to similar cars. Most self-driving cars are relatively new and relatively large, and both decrease those (probably even accident rate because they statistically have newer brakes and tires and have more active safety features)

It’s far from trivial to decide whether some program is a better driver than the average human driver if the program gets the newer bigger car and can decide to cop out in cases where we do not expect humans to do so.

[lurkingmba]

You can add, "at a similar time of day". Hours after midnight are very dangerous.

To my knowledge, Tesla has never put out data on how Autopilot is "safer" on accidents per mile that was close to being apples to apples.

[d_theorist]

I agree you need to control for other factors. But I don't think that is particularly beyond the wit of man.