> ... but it certainly won't have the reach of putting the feature in people's phones.

I think this may be another job for Bayes. Apple's algorithm may be very accurate at detecting car crashes when there are car crashes, but most of the time there are not car crashes. As a result it may actually be super inaccurate at the 'population' level. So while it may have the reach, the incidence of false positives and the infrastructure misallocation may yield worse outcomes than not having the system in the first place.