AI reaches for brand drugs, even when generics exist

Clinical takeaway: Before acting on an AI drug recommendation, check whether a lower-cost generic is available. Explicitly asking for generic options may broaden the recommendations.

About one in three US nonfederal acute care hospitals were running large language models inside their electronic health records in 2024, while many patients were using the same tools to seek health information on their own. Choosing a medication means weighing diagnosis, comorbidities, and patient preference, which makes drug recommendations an obvious use case. Most work on bias in these models has examined race, gender, and income in triage and imaging decisions. 

Which product a model names within the same ingredient had drawn far less scrutiny. Models trained on web text can favor whatever is most publicized over a clinically equivalent alternative, and guidelines steer clinicians toward generics to hold down cost. This evaluation, built on simulated cases, puts numbers on how concentrated those product choices are and what predicts them. 

The models strongly favored brand-name products over generics. Brand status was the strongest predictor of a recommendation, increasing the odds roughly 20-fold. When asked about a condition and a specified active ingredient, the models recommended the same product 97.5% of the time on average, despite an average of 5.5 FDA-approved alternatives being available. The five models agreed almost perfectly on which product to choose. Prompting the models to favor the cheapest option changed the recommendation in 62.6% of comparisons, while prompting them to favor generics changed it in 77.9%. Most of those shifts were to generic drugs, but the models still leaned on a narrow set of products. 

Products with more Google searches or a dedicated Wikipedia page were recommended more often, while products approved later or prescribed more often in practice were recommended less. The most-prescribed product in national survey data matched the models’ first pick in 29.7% of cases. Manufacturer promotional spending was linked to slightly lower odds of a recommendation, not higher. 

Projected acquisition costs rose for 74.5% of the 51 active ingredients with complete pricing data, adding $55.6 million per year based on simulated pharmacy invoice prices. Between 12.4% and 14.7% of recommendations didn’t match FDA records, almost all for products no longer marketed, which the authors say doesn’t necessarily make a recommendation clinically inappropriate. 

Investigators ran 177,120 synthetic vignettes through five models: GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4, and MedGemma. Each vignette named a condition and an ingredient, and the models listed up to three FDA-approved products. The team compared the picks with prescribing in the Medical Expenditure Panel Survey from 2019 to 2023, and priced them using National Average Drug Acquisition Cost data. 

The authors note that the models were tested at the recommendation step only, so what patients receive after pharmacy substitution and dispensing is untested. The generic-leaning prompts may show the models following instructions rather than persistent bias, the authors say. The study advocated for reporting regulatory concordance by patient demographics and for pairing AI output with human verification. 

“Our findings highlight the need to evaluate healthcare AI for biases in product selection and the handling of cost and regulatory information,” Daiju Ueda, professor at Osaka Metropolitan University Graduate School of Medicine, said. “Combining AI systems with trusted drug information sources and human oversight may help support more appropriate medication choices.” 

Source: Sato J, et al. (2026 Oct 26) Nat Health. Large language models show product name convergence in drug recommendations