What the research found
Analysts at NYC Health + Hospitals, the largest US municipal safety-net system, evaluated large language models in their electronic health record over the past year. AI-generated chart summaries omitted A1c information for nearly 25% of patients and included wrong information for 15%, and omissions were three times as common for patients with five or more comorbidities as for those with diabetes alone; they did not see significant variation by race or ethnicity. AI-drafted portal replies were written at a 9th-grade reading level against 7th grade for the system's providers.
Patients with the most complex conditions were the ones most often missing key results in AI summaries, so providers should test summary tools on their sickest patients before relying on them.