Led by Harvard medical student Arya Rao, a research team published in JAMA Network Open this week the results of a study that examined 21 leading off-the-shelf AI models in 29 standardized clinical vignettes. The bots all did fairly well when provided a full portfolio of medical information and asked to make a final diagnosis, with leading models correct 91 percent of the time. Early differential diagnosis, where clinicians try to rule out certain conditions while weighing various possibilities, is where that more than 80 percent failure rate comes in.
“Every model we tested failed on the vast majority of cases,” Rao told The Register in an email. “That’s the stage where uncertainty matters most, and it’s where these systems are weakest.”
In other words, it’s the midnight anxiety-fueled WebMD rabbit hole of yesterday all over again, just supercharged with AI that’s probably even more likely to get things wrong than you are without it.
This is commonly done for the purposes of replicability, but is not at all how these models are deployed in practice.
Larger institutions, especially those with strict data privacy requirements, are deploying locally hosted models permanently RAGed to their own internally vetted documentation.
It would’ve been much more interesting to see how much RAG setups fail, contrary to their marketed promises.
From experience, RAGs do help reduce hallucinations, but LLMs still do dumb things, like jumble up numbers. There were many cases where the LLM confidently presented some numerical results, but the number existed somewhere else entirely, like a footnote on the same page.