IANAD [doctor] and missed that in the sea of jargon - but I think your question actually clarifies how the LLM got confused. I wouldn't necessarily say "more ge

IANAD [doctor] and missed that in the sea of jargon - but I think your question actually clarifies how the LLM got confused. I wouldn't necessarily say "more general than vision," but more severe. There's a big difference between "unaware of objects" and "difficulty seeing or distinguishing objects." It looks like GPT-4 disregarded "unaware of objects on their left side" in favor of more clinical (but less specific) information later in the question: "two hours after the accident, there was a left hemianopia." According to Dr. Wikipedia, "hemianopia" is a vague term that could indicate partial loss of vision, or could only affect one eye. But the patient's informal description sure doesn't sound like a partial loss of vision, nor did he say "left eye." The information of "left hemianopia" needs to be considered alongside what the patient self-reported - it confirms and clarifies what the patient said - but the LLM didn't seem to do this. There's other implicit information that the LLM would have trouble leveraging: - the problem is so serious that a 35-year-old with a 20-year history of migraines immediately went to the hospital, so I would be very reluctant to dismiss it as a migraine related to the "stress of the accident." - when I hear "unaware of objects on their left side" my gut reaction as someone with a brain is "that sounds like brain damage, you need to go to the hospital." - Even with its limitations, GPT-4 should have realized "two hours after the accident" meant that it wasn't a temporary migraine-related hemianopia. Apparently it didn't. But I think this is amenable to specific training. The other two things are not, but they are the kind of thinking doctors rely on.