I agree that the demographic skew probably accounts for some of the difference. Probably also that Teslas need less maintenance (esp brakes, due to regenerative braking), so are less likely to fail for mechanical reasons — although I don't think most crashes are due to mechanical failure, it should show up to some degree in the stats.

I think the argument that the Autopilot numbers are essentially fake because the true crashes are concentrated in the Autopilot-off scenarios is hard to make a case for though, given the stats on Autopilot-off driving being so comparatively good. You would need incredibly good demographic skew to account for that if the crash rate is concentrated — you don't need to just equal the average after correcting for demographic skew, you need to be considerably worse than it. So while it's not a perfect metric, I would be much more surprised if Autopilot+human was more dangerous than human alone.

I 100% agree with you that this is only an argument for Autopilot+human though. Current Autopilot without humans, at least anecdotally (I have a Model 3 with Autopilot), does not seem safe. However, I think the concern among some that Autopilot is unsafe as it currently is typically operated — i.e. with a human in the loop — is largely contradicted by the evidence.

My personal anecdote is that I feel much less fatigued by driving with Autopilot, especially on longer drives. It's imperfect, but it actually generally helps improve my alertness because I don't have to constantly fiddle cruise control settings based on which car is in front of me or make micro wheel adjustments to stay centered in a lane; I usually take over proactively whenever there looks like a sketchy situation is coming up like multi-lane merges with trucks for example. And when those situations happen, I'm able to stay more alert and focused because I haven't been spending my energy on the simple stuff that Autopilot is good at, so I think I end up being safer overall even when it's disengaged. I notice a pretty large cognitive difference — which was unexpected for me when I first got it, because I thought I probably wouldn't use or like Autopilot, and initially was quite mistrustful of it.

Obviously this is just a personal anecdote, and not data! But what data we have, while imperfect, seems to support that conclusion much more than it supports the opposite.