Raging against the Machine: The Human Side of the Story

Large language models (LLMs) such as Generative Pretrained Transformer (GPT) 4 are increasingly recognized for their potential in clinical diagnosis, yet their performance alongside experienced physicians in real time has been less examined. In September 2023 at Harvard Medical Grand Rounds, the author, an internist, and GPT-4 simultaneously evaluated a diagnostic unknown before a live and remote audience. GPT-4 demonstrated notable diagnostic accuracy and often paralleled the author's reasoning, but also displayed diagnostic rigidity and anchoring reminiscent of human cognitive biases. The author posits that human diagnosticians retain advantages in contextual interpretation and in detecting subtle bedside cues that LLMs cannot, highlighting the promise of LLMs as cognitive adjuncts, while affirming the enduring importance of observation, narrative synthesis, and clinical judgment in the diagnostic process.