Physicians in Latin America were randomly assigned to answer four simulated clinical cases, each with four open-ended questions, either supported by an evidence-traceable conversational AI system or using their usual search practice. Two blinded specialists per area scored every response on a six-dimension rubric, and the main outcome was a composite validity score per physician. Of 202 physicians who started, 71 completed the exercise and were analyzed (26 supported, 45 usual practice), for 1,136 responses.
Supported physicians scored higher, with a median of 2.83 versus 2.46 out of 3 (difference 0.38; 95% CI 0.17 to 0.54; P < .001), and were more likely to meet the validity threshold after adjusting for academic degree (odds ratio 3.61; 95% CI 1.16 to 11.17). The effect held for the four criteria with the highest interrater agreement but not for accuracy alone, which an external evaluator also found non-significant; response time and searches did not differ, and acceptability was high (2.92 of 3). The sample is volunteers who finished, so the authors call the findings exploratory; this is a preprint not yet peer reviewed.
