The research team tested five models: GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4, and MedGemma. Each was asked to recommend up to three FDA-approved drug products across a range of simulated clinical scenarios. The researchers then analyzed which products the models chose and why. The recommendations clustered around brand-name drugs, products that rank highly in Google searches, and treatments with dedicated Wikipedia pages. In other words, the models appeared to favor drugs with a strong digital footprint, a pattern the authors describe as product name convergence.

The bias extended beyond visibility. The models tended not to recommend products that had received later FDA approval, and they also overlooked drugs that are more commonly prescribed in actual physician practice, indicating that their choices did not mirror real-world prescribing behavior. More concerning, the study identified recommendations that were inconsistent with regulatory information, including suggestions of products that have been discontinued. Because members of the public increasingly turn to chatbots for health information, such errors could mislead users about treatments that are no longer available or approved.

To estimate the financial consequences, the researchers used US data to model a scenario in which physicians’ prescribing behavior aligned with the AI recommendations. They found that annual acquisition costs would increase by approximately US$55.6 million, a 1,019% increase in the cost of purchasing the affected prescriptions. The projected rise stems largely from the models’ preference for brand-name products over cheaper alternatives that were equally approved and available for the same active ingredients.

The team also examined whether the bias could be corrected. When the models were explicitly instructed to prioritize lower-cost or generic products, their recommendations did shift accordingly, showing that the systems can respond to cost-related prompts. However, even under these instructions, the recommendations remained concentrated on a limited set of products rather than spreading across the full range of approved options. The convergence pattern persisted, suggesting that the underlying preference for well-known names is not easily eliminated by a single instruction.

Professor Ueda emphasized the practical implications of the findings. “The recommendations were mostly for brand-name drugs, those highly visible in Google searches, and treatments with dedicated Wikipedia pages,” he said. This suggests that the training data and information ecosystem surrounding a drug, including its search ranking and online encyclopedic coverage, may act as hidden filters on what the models propose, regardless of clinical equivalence among approved products.

The cost projections highlight a systemic risk if AI outputs were adopted at scale in healthcare settings. While the study modeled a hypothetical alignment of prescribing with AI suggestions rather than observed changes in clinical practice, the magnitude of the estimated increase, roughly US$55.6 million annually for the affected prescriptions, indicates how even modest per-decision biases can compound across a national prescribing volume. The authors note that the analysis was based on US acquisition cost data and statistical modeling, and the figures represent potential rather than realized spending.

“Our findings highlight the need to evaluate healthcare AI for biases in product selection and the handling of cost and regulatory information,” Professor Ueda said. “Combining AI systems with trusted drug information sources and human oversight may help support more appropriate medication choices.” The study, titled “Large language models show product name convergence in drug recommendations,” was published on 26 October 2026, and the authors declare no competing interests. As AI tools continue to mediate access to treatment information, the results argue for systematic evaluation of how these models weigh visibility, cost, and regulatory status before their recommendations influence medication decisions.

Subject of Research: Bias in large language model drug recommendations and its potential cost impact

Article Title: AI’s preference for drugs with a strong online presence may cost millions

Article References: AI’s preference for drugs with a strong online presence may cost millions. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, drug recommendations, AI bias, brand-name drugs, FDA-approved products, healthcare costs, prescription drugs, generics, regulatory information, ChatGPT, health information, Osaka Metropolitan University