A Methodology for Normalizing Safety Statistics of Partially Automated Vehicles

This paper prove that the Tesla's comparison of raw Autopilot crash data to average human driver crash data is deceptive because it doesn't adjust for the safety of the operating domain (among other factors).

It creating an apples-to-apples comparison of crash data based on the operating domain, he also adjusted for owner age. The results effectively eliminate Tesla's claim that Autopilot meaningfully improves safety outcomes.