A new study evaluating five large language models (GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4, and MedGemma) on recommending FDA-approved cancer drug products across simulated scenarios finds a bias toward brand-name, costly treatments over cheaper or generic options.

The research was led by Professor Daiju Ueda of Osaka Metropolitan University’s Graduate School of Medicine, in collaboration with MD Anderson Cancer Center, and published in Nature Health (Sato et al., 2026).

Model agreement was very high (Fleiss’ kappa about 0.895), indicating near-perfect concordance among the models on top choices, yet this did not match real-world prescribing patterns from MEPS data, where the first-choice match averaged 29.7%. There is a gap between model output and actual practice.

Some outputs conflicted with regulatory information, including recommendations for discontinued products.

The originator (brand-name) status emerged as the strongest predictor for recommendations, with odds about 20.5 to 1 for originators versus generics; public visibility factors like Google search counts and Wikipedia presence also increased likelihood of being recommended.

The study tested 177,120 synthetic clinical vignettes to assess drug recommendations for common conditions.

Exploratory disease-only analysis showed convergence within models but variation across diseases, suggesting prompt design can nudge toward cost-conscious choices but cannot eliminate concentration bias.

Researchers warn that relying on AI chatbots for medical guidance may lead to suboptimal or unsafe treatment choices due to biases in recommendations.

Limitations include the use of synthetic vignettes that don’t capture full clinical complexity, MEPS data constraints, missing cost data, and evaluation limited to the recommendation stage without considering dispensing or insurance factors.

A National Academy of Sciences survey notes teenage boys spent about 40 minutes outdoors daily in 2023, down from roughly an hour in 2010, while girls’ outdoor time remained unchanged.

When pooling recommendations, a dominant product accounted for over 90% of first-choice picks in 78% of disease–drug pairs (96 of 123), signaling a strong bias toward certain products.

Factors reducing the likelihood of being recommended included later FDA approval, higher real-world prescription volume, higher manufacturer promotional spend, and longer drug names.