ChatGPT’s newest version, GPT4, was able to pass a no-image multiple-choice radiology exam.
GPT4 is neat, but this says a whole lot more about how useless and off-base a radiology exam without pictures is than about how ready for prime time the current AI tools are. But they’re coming, and I for one am interested to see more natural language processing combined with what’s already out there to actually make healthcare more efficient (automated discharge summary drafts! a real summarized history on imaging orders!)
The authors also do note that when GPT4 is wrong, it’s often wrong in spectacular ways (and will boldly make up lies with the same confidence as it dishes out true answers). As in, not quite ready for anything when real performance counts.