A big challenge is that most doctors can't operate outside of the most common conditions and differential diagnoses of their specialty and have to punt patients.

Of course, to the doctor's defense, patients are not an easy population to work with, "Why can't I eat candy and cake 3 meals day and my diabetes is getting worse?"

There have been clinical decision support systems that were pretty decent at doing DDX for decades, well before the current crop of AI LLMs. The issue is getting relevant data about patient history, symptoms, signs, etc into the system and then convincing doctors to actually use it. For most simple cases, using any form of clinical decision support (whether it's based on AI or simpler deterministic algorithms) just slows them down and they don't have time to waste.