I'm optimistic for this area of tech and research in general, but agree we need to stop benchmarking against average human crash rates.
Although anyone can be hit at any time, the distribution of human crashes is not purely random. People who drive compromised, for example, are way overrepresented in those events.
So, theoretically, the tech could get to a point with a lower than expected crash rate for humans generally, but still increase your personal crash likelihood.
If you believe you are a better than average driver when manually driving, why couldn't you also be better than average at intervening when autopilot is driving?
Humans don't have forward facing radar nor 360 degree always active vision, so in some cases, cars can see hazards that even a perfect human driver cannot.
It's a natural human tendency to get distracted, especially when we are not engaged with a task (as with autopilot). Driving manually physically engages your body in the task, making it harder to get distracted. On the other hand intervening is subject to distraction and longer response times.
Although anyone can be hit at any time, the distribution of human crashes is not purely random. People who drive compromised, for example, are way overrepresented in those events.
So, theoretically, the tech could get to a point with a lower than expected crash rate for humans generally, but still increase your personal crash likelihood.