Instead of anchoring on "it does work, unless a human intervenes to the downside", consider "it *isn't* working, unless a human can show evidence to the contrary."
There are many more ways for things not to work than not. Better to capture the signals that give us confidence systems are working as desired.