If we’re supposed to be telling the AI what’s right, why do we so often get it “wrong?”