In any case, I think an interesting question is the implementation of LLMs in real-time clinical consultations (gradual information elicitation, belief updating).
Ah yes, someone else also sent me that. The structured format of responding (“reflective practice”) could have improved performance in the control group, so that the LLM group did not do significantly better.
They were discussing the fact that the third group (LLM alone) outperformed the doctors or the LLM "augmented" doctors. I don't know enough about medical diagnosis to comment but would be curious to see if that replicates in other domains.
Same - but interesting that the better control was the LLM alone, no different from Dawes. Paraphrasing the author interviewed: doctors with LLMs disregarded LLM responses that disputed their initial (incorrect) hypothesis.
I didn't know there was an interview. This doesn't surprise me. We have often found that the initial hypothesis determines the (process and) outcome of the diagnostic process. Also, see studies on predecisional information distortion.
Comments