Bias and Inaccuracy in AI Chatbot Ophthalmologist Recommendations

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Ophthalmology Practice Management

Background:

• AI chatbots are increasingly used for information retrieval, including healthcare-related queries.
• The accuracy and potential biases of AI recommendations in specialized medical fields require thorough evaluation.
• Understanding AI performance in physician referrals is crucial for patient safety and equitable access to care.

Purpose of the Study:

• To assess the accuracy and identify biases in ophthalmologist recommendations from ChatGPT 3.5, Bing Chat, and Google Bard.
• To analyze chatbot performance across the 20 most populous U.S. cities.
• To compare AI-generated recommendations against national averages for physician demographics and practice types.

Main Methods:

• Generated 80 recommendations per chatbot using the prompt 'Find me four good ophthalmologists in (city)'.
• Collected physician characteristics: specialty, location, gender, practice type, and fellowship.
• Utilized one-proportion z-tests and Pearson's chi-squared tests to compare chatbot recommendations with national data (e.g., proportion of female ophthalmologists, academic practice prevalence).

Main Results:

• Bing Chat (1.61%) and Google Bard (8.0%) significantly underrepresented female ophthalmologists compared to the national average (27.2%).
• All chatbots demonstrated high inaccuracy rates: ChatGPT (73.8%), Bing Chat (67.5%), and Bard (62.5%).
• Chatbots disproportionately recommended ophthalmologists in academic medicine or combined practices, exceeding the national academic average (17%).

Conclusions:

• AI chatbots exhibit significant bias and inaccuracy in recommending ophthalmologists, often suggesting physicians outside the specialty or desired location.
• Bing Chat and Google Bard displayed a bias against recommending female ophthalmologists.
• All evaluated chatbots showed a preference for recommending ophthalmologists in academic settings.