Opinion: STAT+: Why AI like Grok isn’t ready for the radiology big leagues
I tested out Grok, my company’s diagnostic AI, and several others to see which would do best at diagnosing from an X-ray. The results were surprising.
Recently, a bunch of people at my local basketball club watched me swagger onto a court and sink a 35-footer. Nothing but net. I soaked in the adulation and gave the requisite high-fives. What I did not do — at any point in time — was believe that shot entitled me to be on the floor in the crunch time of an actual NBA game.
In January, a user on X posted about how Grok, X’s artificial intelligence tool, diagnosed their daughter’s broken wrist from an X-ray that her care team had misread. The post now has more than 14 million views and, in the ultimate X win, got a shout-out from Elon Musk, who declared, “Grok can diagnose medical injuries.”
Certainly Grok and other vision large-language models (LLMs) have massive potential, and when they occasionally get stuff right it looks magical, just like my 35-foot net burner. But they are not ready for game time, at least not yet.
