In the study published in Nature Sensors, 623 people distinguished melanoma from nevi, while 153 primary care physicians performed differential diagnoses among four skin diseases. In addition, 320 medical students completed the same task as the physicians to assess the impact of medical training. Participants were randomly assigned to one of four AI-assisted methods. In the public group, AI increased diagnostic accuracy from 69.7% to 75.8% and reduced the performance gap between skin tones. Explanations from a multimodal large language model contributed the most to correct predictions, while also increasing confidence in incorrect predictions. Physicians’ performance improved significantly; they remained resistant to incorrect recommendations and were more resilient than medical students. Clinical expertise enabled them to use AI more critically.
Background
Human is not a new name in the FikirPilot archive: we have published 2 reports mentioning this name in the past 90 days; the most recent one is dated August 31, 2026.