El cuello de botella eres tú: la paradoja de la experiencia en la IA clínica.

In this episode of Memorandum, I analyze a recently published study in Radiology (Schramm et al.) and its editorial (McMillan). Twelve readers with three levels of experience interpreted 40 brain MRIs with and without the assistance of language models. The finding: AI accuracy was higher when fed by expert neuroradiologists (≈79–84%) and lower with resident-led interpretations (≈63–67%). But the diagnostic gain was inversely related to experience: +19.4% for the less experienced readers, and only +4.4% for the specialists. The lesson, valid for all of medicine: the model doesn't see the patient, it sees your description. The ability to observe and describe well remains the scarce asset—and the risk of overconfidence and detraining is real, especially in training. In the episode, I also discuss what this means for Mexico and Latin America: language, the expert gap, and the risk of bypassing the radiologist.