Across five tested models, a single brand-name product accounted for roughly 97.5% of first-choice recommendations for each active ingredient, despite an average of 5.5 FDA-approved products being available, showing extreme convergence on one option.

The same pattern held across all five models, with a single brand-name product most frequently recommended as the first choice, capturing about 97.5% of first-choice picks and signaling concentrated bias.

Researchers note limitations including synthetic vignettes not capturing full clinical complexity, MEPS data gaps, missing cost data, and evaluation limited to the recommendation stage without assumptions about downstream dispensing or insurance.

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

Authors urge evaluating healthcare AI for biases in product selection, cost handling, and regulatory information, advocating combining AI with trusted drug sources and human oversight to improve medication choices.

They caution that healthcare AI should be used carefully with proper oversight to prevent biased drug selection and cost inflation.

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

Exploratory disease-only analysis showed convergence across models but varied by disease, indicating prompt design can steer toward cost-conscious choices while not fully removing concentration bias.

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

LLMs used to recommend FDA-approved drugs showed bias toward brand-name, costly treatments rather than lower-cost or generic options in simulated scenarios.

The five models tested were GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4, and MedGemma, evaluated on 177,120 synthetic clinical vignettes to test drug recommendations.

Originator status emerged as the strongest predictor for recommendations, with high odds for brand-name products and public visibility factors like Google search counts and Wikipedia presence increasing likelihood of being recommended.