What happens when you ask an AI to diagnose a medical image, give it the patient’s demographics, but leave out the image? Too often, it makes up a diagnosis anyway, and the answer can depend on who the patient is.
Expanding this work more broadly later this year.
@siddvoh recently showed that LLMs intermittently invent false information when responding to medical questions 🏥⁉️
Sidd's findings reinforce a critical concern he sees in AI industry: users often assume models understand more than they actually do.
ri.cmu.edu/healthcare-bli…
