Study finds clinical chatbots lag behind general AI models in accuracy, sparking debate among doctors

Hundreds of thousands of U.S. physicians are using clinical large language model tools marketed by firms such as OpenEvidence, Doximity and UpToDate as safer alternatives to generic AI chatbots.

A research team from NYU Langone Health evaluated both general purpose and clinical focused models on three sets of medical questions, and the results were published in Nature Medicine in June.

The study showed the clinical specific AI performed worse than the general models, a finding that surprised many and prompted a strong reaction on social media, including a comment from Kaiser Permanente's vice president of AI.

Developers of the clinical tools argue that current benchmarking methods do not accurately capture safety and usefulness, highlighting ongoing debate over how to assess AI in healthcare.