Five AI models favored branded drugs in a study, raising projected costs

Five AI models repeatedly selected a narrow set of drug products, often favoring brands, in a Nature Health study published October 26. Their choices differed sharply from actual prescribing and produced higher projected acquisition costs. The experiment tested which products models recommend after a disease and active ingredient have already been specified—not whether they choose the right treatment.

Researchers led by Daiju Ueda at Osaka Metropolitan University, working with the University of Texas MD Anderson Cancer Center, tested GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4 and MedGemma. They submitted 177,120 synthetic patient scenarios spanning 123 disease–drug pairs and 1,440 combinations of social and demographic characteristics.

Each prompt requested up to three FDA-approved product names, ranked by preference. Four physicians reviewed the outputs to standardize names and check their agreement with FDA approval information. That design let the researchers examine commercial neutrality: which specific products an AI favors when the active ingredient is held fixed.

Different patient profiles, much the same product

Across disease–drug pairs and models, the most frequently selected product captured an average of 97.5% of first-choice recommendations. That was despite an average of 5.5 FDA-approved products being available. The percentage describes the dominant product within each tested pair; it does not mean one medicine dominated recommendations across all diseases.

The models also largely agreed with one another. Pooling their answers, the leading product accounted for more than 90% of first choices in 96 of the 123 disease–drug pairs. Changing the model therefore often left the preferred product unchanged, rather than opening a substantially different set of options.

Product visibility was associated with selection. Drugs with more Google search results or a dedicated Wikipedia page were more likely to be recommended, and the models favored originator products. Products approved later by the FDA, and those more commonly prescribed in practice, were less likely to appear. These were associations, not a demonstration that online visibility caused a recommendation.

To compare those answers with practice, the researchers used 2019–2023 records from the Medical Expenditure Panel Survey, a nationally representative US survey. Across the models, the most commonly prescribed product matched the first-choice recommendation in only 29.7% of cases on average. It appeared among their top three choices in 58.8%.

Some outputs also conflicted with regulatory information, including recommendations for discontinued products. This was a separate problem from favoring a small number of available brands: the recommended product itself could be inconsistent with current regulatory records.

Turning product choices into a cost projection

The team then applied pharmacy acquisition prices to the two product distributions. Those prices reflect average invoice costs for retail community pharmacies, not patients’ out-of-pocket bills. Researchers calculated per-unit cost differences and multiplied them by annual prescription volumes to estimate what would happen if prescribing followed the model-generated product mix.

The cost analysis covered 51 active ingredients. Of the original 101 unique ingredients, 18 lacked corresponding prescription records and 32 lacked complete pricing for relevant products. Within the included group, 38 ingredients—74.5%—had higher projected acquisition costs under the AI recommendations. The increase was therefore common, but not universal.

$55.6 millionProjected additional annual cost

Ueda put the estimated increase at approximately $55.6 million, or 1,019%, for the prescriptions in the US scenario.

Those figures are simulated projections, not observed spending increases. The study did not model downstream prescribing, pharmacy substitution, dispensing or purchasing. Its estimate measures the implications of the recommended product mix, rather than predicting every step between an AI answer and a medicine actually bought.

Cheaper prompts changed the answers, not the concentration

When instructed to prioritize lower-cost or generic products, the models shifted their recommendations in that direction. Their choices nevertheless remained concentrated on a limited set of products. Ueda called for evaluating product-selection bias alongside cost and regulatory accuracy, and suggested combining AI with trusted drug information sources and human oversight.