Raw LLM Responses

Similarly, I heard of efforts to estimate chances of short term survival for trauma patients in the ER. When the first AI came back with a pretty strong accuracy (I forget the exact numbers, but it was in the 80% area iirc) people where pretty stoked about how good it was. But when they "cracked open" the AI and started trying to find out how it was doing it, they noticed that it didn't look at the patient at all. Instead, it looked at the type of gurney that was used during the scan. The regular gurney got a high chance of survival, the heavy-duty, bells-and-whistles gurney got a low chance, as that gurney is used for patients with heavy trauma. Another one I heard did something similar (I forget the goal completely), but it based its predictions on the text in the corner of the image, mainly it learned to read the date of birth and make predictions based on that.