> I wonder if there is really enough signal going to the training processes for “a human understands what is going on”

Unverifiable, un-scalable, no.