New research has found that an AI chatbot outperformed human doctors in some tasks but also created safety risks and amplified social inequities.
Human physicians in China were evaluated against ERNIE Bot—China’s most widely used AI chatbot—and two other advanced AI systems, ChatGPT-4o and DeepSeek R1. The large simulation study—conducted by an international research team including scientists from the Yale School of Public Health—involved 384 simulated patient consultations for unstable angina and asthma. Heart and respiratory conditions are non-communicable chronic diseases, one of the leading causes of death and illness worldwide.
Such conditions are responsible for 41 million deaths annually, with 77% occurring in low- and middle-income countries. The promise of AI bots is that they could bridge crucial gaps in low- and middle-income countries, where many patients remain undiagnosed or poorly managed due to a shortage of qualified health care providers.
In the study, ERNIE (Enhanced Representation through kNowledge IntEgration) Bot achieved a 77.3% diagnostic accuracy rate and 94.3% accuracy in prescribing correct medications, far outperforming frontline doctors in China’s primary care system, who gave correct diagnoses only 25% of the time and correct prescriptions only 10% of the time.
The chatbot also ordered unnecessary lab tests in 91.9% of the cases and prescribed potentially inappropriate or harmful medications to 57.8% of the simulated patients. In addition, older patients (65 vs. 55) and wealthier patients received more accurate diagnoses and more intensive—but often excessive—treatment than their younger or poorer counterparts.
The ERNIE Bot also followed only 14.5% of standard full diagnostic checklists and 20.3% of essential steps, raising concerns about its thoroughness and interest in higher efficiency.
Despite the mixed results, the study’s authors, an international team from Australia, China, and the United States, wrote that the chatbots “hold promise in alleviating the burden of non-communicable chronic diseases by extending diagnostic and treatment capabilities in settings where resources are scarce.” But they cautioned that “our findings also emphasize balancing AI’s potential with necessary safeguards.”