I think the "confident but inaccurate" problem is one of the biggest hurdles for most of the powerful usecases of these models. We don't seem too far from being able to automate filler text or stock images, which is certainly useful but only so valuable, but a long way off from more technical usecases like law or medicine. Being right 90% of the time but completely wrong 10%, without any way to distinguish the two, isn't going to get very far in many fields of work.