If they only considered the parts where Autopilot is enabled, it's obviously biased since people are going to enable Autopilot only where it is most likely to w

If they only considered the parts where Autopilot is enabled, it's obviously biased since people are going to enable Autopilot only where it is most likely to work and not kill them, and in general where accidents are less likely overall. If they instead considered drivers who never use Autopilot vs those who sometimes use Autopilot, it would make more sense, but then it might be biased in other ways due to self-selection. The only reasonable way to do this test is to randomly deny the use of Autopilot to some drivers who would otherwise use it and compare accident rates between those two groups (this still only applies to people who would want to use Autopilot, but it would be unethical to force people to use it since it's dangerous).