Supportive AI in Primary Care and Internal Medicine Evidence, Limitations, and Governance for Latin America
Abstract
Primary care and internal medicine involve multimorbidity, longitudinal follow-up, coordination, extensive documentation, and decisions under uncertainty. Recent evidence suggests that AI can support synthesis, documentation, clinical search, messaging, administrative prioritization, and risk prediction, but does not justify presenting it as an independent diagnostic or therapeutic substitute. A systematic review identified 519 studies of large language models in health care, of which only 5% used real clinical-care data. In a study integrated into electronic health records (EHRs), message drafts increased reading time by 21.8%, did not significantly reduce response time, and increased response length by 17.9%. A randomized trial likewise found no significant improvement in physicians' diagnostic reasoning with access to a large language model. Direct regional evidence is narrow: a Brazilian scoping review included 27 studies and found mainly machine-learning and deep-learning applications, alongside data-availability and access barriers. For Latin America, the implication is to strengthen local evaluation, interoperability, equity, and oversight without treating this agenda as proof of clinical effectiveness in the region.
