I don’t think it’s as black-and-white as you claim. Suppose a driver is inattentive at some moment with probability x, and conditions where inattentiveness would lead to a crash exist with probability y, then a probability of a crash at that moment is xy. Probably, both x and y are > 0 for all drivers, but xy could be low enough that an accident never occurs.

The driver has no reliable way of estimating x or y and therefore can’t judge the risk of their lack of attention causing a crash. Maybe they “shouldn’t” be driving according to this formula, but they don’t know that, and autosteer could save them from an accident.

Most likely, the presence of autopilot increases x and decreases y. We can’t tell which affect dominates by arguing on the internet. We need data.