AI drug recommendations converge on the same few brand-name medicines, study finds

Artificial intelligence systems that promise to streamline prescribing may be quietly steering medicine toward a remarkably narrow shelf of drugs. A large-scale evaluation published in Nature Health on 26 October 2026 found that five leading large language models (LLMs), when asked to recommend medications for common conditions, overwhelmingly converged on the same product name regardless of who the hypothetical patient was, and that this convergence frequently pointed away from the drugs clinicians actually prescribe and toward more expensive options. The scale of the experiment was deliberately enormous.

The team constructed 177,120 synthetic clinical vignettes by combining 123 curated disease–drug pairs with 1,440 prespecified sociodemographic profiles spanning age, sex, race or ethnicity, LGBTQ+ identity and income status. Each vignette described a concise patient scenario, specified a disease and an active ingredient, and asked the model to list up to three FDA-approved products by name in order of preference. Five models were tested: GPT-5.4 from OpenAI, Claude Sonnet 4.6 from Anthropic, Gemini 3.1 Flash from Google, and the open-source Llama 4 and MedGemma.

All models ran at a temperature of zero, the most deterministic setting, and together they produced 883,403 responses containing at least one drug recommendation. Four physicians then manually curated every output, checking each generated name against the FDA Orange Book and Purple Book, which cover approved drug products including originator (brand-name) products and generics. The headline finding was stark.

Across all disease–drug pairs and all five models, the single most frequently recommended product accounted for a mean of 97.5 percent of first-choice recommendations. This collapse onto one option occurred even though most pairs had, on average, 5.50 distinct FDA-approved products containing the specified active ingredient. When recommendations were pooled across all five models, the dominant product captured more than 90 percent of first-choice picks in 78.0 percent of pairs, or 96 of 123.

The models also agreed with one another to a striking degree: Fleiss' kappa, a statistic of inter-rater agreement, was 0.895 (95 percent confidence interval 0.855 to 0.931), which the authors characterize as almost perfect. Concordance with the pooled dominant choice was highest for Claude Sonnet 4.6 at 96.6 percent and lowest for MedGemma at 90.6 percent. That uniformity would be unremarkable if it mirrored actual clinical practice, but it did not.

The researchers compared LLM-recommended product distributions against real-world prescribing patterns drawn from the Medical Expenditure Panel Survey (MEPS), a nationally representative US survey of dispensed prescription medicines, aggregated over 2019 to 2023. Of the 123 pairs, 100 had corresponding products in the MEPS data. The mismatch was severe: the most frequently prescribed drug in MEPS matched the LLMs' first choice in only 29.7 percent of cases on average, and fell to 17.0 percent when weighted by prescription volume.

A distributional metric called total variation distance, which ranges from zero for identical distributions to one for no overlap, reached median values of 0.93 to 0.98 across the five models, indicating very limited overlap between what the models recommend and what is dispensed in the real world. The economic implications followed directly from that divergence. Using National Average Drug Acquisition Cost (NADAC) data, which reflect invoice prices at retail community pharmacies, the team projected what would happen if prescribing shifted to match LLM recommendations.

After exclusions for missing MEPS records or incomplete pricing, 51 active ingredients could be analyzed. For 38 of them, or 74.5 percent, the LLM-recommended distribution was more expensive than the real-world one. Summed across all 51 ingredients, the differences amounted to an estimated annual increase of US$55.6 million in acquisition costs, equivalent to a 1,019 percent increase relative to current MEPS-derived spending patterns.

The authors caution that these figures are simulated projections based on prespecified normalization assumptions, not observed expenditures, but the direction of the effect was consistent and substantial. Why do the models favor certain products? A multivariable analysis using generalized estimating equations identified originator status as by far the strongest predictor: innovator brand-name products had an odds ratio of 20.5 for being recommended compared with generics.

Public visibility mattered too, with each increment in Google search counts raising the odds by a factor of 1.70 and the existence of a dedicated Wikipedia page tripling them (odds ratio 3.00). Conversely, products with later FDA approval (odds ratio 0.76), higher real-world prescription volume (0.87), greater manufacturer promotional spending (0.97) and longer names (0.88) were recommended less often. The pattern suggests a popularity-related bias in which models gravitate toward drugs that dominate online discourse, a tendency that could entrench the market position of already visible brands and marginalize equally appropriate alternatives.

The study also probed regulatory reliability, and the results complicate any assumption that unanimous recommendations imply correctness. After physician curation, outputs that failed regulatory checks were classified into five categories, and across models the overwhelming majority of nonconcordant outputs, between 91.69 and 99.59 percent, referred to previously FDA-approved products that are no longer marketed. Overall nonconcordance rates ranged from 12.4 percent for GPT-5.4 to 14.7 percent for Claude Sonnet 4.6.

More troubling, the odds of a nonconcordant recommendation varied with the patient described in the vignette: vignettes specifying Black patients carried 1.64 times the odds of nonconcordant output relative to the reference category, Asian patients 1.16 times, and lesbian patients 1.06 times. The authors stress that such outputs were not predominant and do not necessarily indicate clinical inappropriateness, since off-label use can be legitimate, but categories such as fabricated names or ingredient mismatches could complicate verification in busy clinical settings. There is a partial remedy, and it lies in the prompt itself.

In sensitivity analyses, the researchers tested three alternative instructions: a comprehensive product listing prompt, a lowest-cost preference prompt and a generic preference prompt. Convergence remained high under all three, with the top product capturing means of 93.1, 94.9 and 97.6 percent of first-choice recommendations respectively. Yet the identity of the top product changed in 45.2, 62.6 and 77.9 percent of comparisons, and when it changed, the shift was toward a generic product in 78.0, 83.1 and 85.4 percent of cases.

An exploratory disease-only analysis, in which the active ingredient was omitted, showed that convergence persisted within each model but varied across models for the same disease, with the dominant active ingredient and product name capturing means of 83.8 and 79.6 percent of outputs. Prompt design, in other words, can steer recommendations toward cost-conscious choices, though it cannot eliminate the underlying concentration. The authors are careful about the limits of their evidence.

Synthetic vignettes cannot capture the full complexity of real-world prescribing, and differences from MEPS do not necessarily mean the models erred, since MEPS distributions themselves reflect clinical, payer, pharmacy and patient-level factors. The cost analysis was restricted to ingredients with complete data, raising the possibility of selection bias, and the study evaluated outputs only at the recommendation stage, without modeling pharmacy substitution, dispensing or insurance formularies that shape what patients actually receive. The alternative prompts were not assessed for task-specific performance, so their high concentration may partly reflect successful instruction-following rather than persistent bias.

Rare diseases and complex clinical scenarios were not examined. Even with those caveats, the implications reach beyond the laboratory. As LLMs become embedded in electronic health records and patient-facing tools, a systematic tilt toward originator brands and internet-famous drugs could undermine generic-substitution policies, inflate costs for underinsured patients and narrow the competitive landscape of the pharmaceutical market.

The researchers argue that deployment evaluations should report regulatory concordance stratified by sociodemographic descriptors, that human-in-the-loop verification remains indispensable, and that transparency about training data will be essential for trust. The full dataset of synthetic vignettes is publicly available on the Hugging Face Hub, and the analysis code is on GitHub, inviting independent scrutiny of a finding that, if replicated in live clinical workflows, could reshape how artificial intelligence is allowed to touch the prescription pad.

Subject of Research: Product name convergence and commercial neutrality in large language model drug recommendations

Article Title: Large language models show product name convergence in drug recommendations

Article References: Sato, J., Walston, S., Takita, H., Mitsuyama, Y., Saito, K., Wu, J., & Ueda, D. (2026). Large language models show product name convergence in drug recommendations. Nature Health. https://doi.org/10.1038/s44360-026-00205-y

Image Credits: AI Generated

DOI: 10.1038/s44360-026-00205-y

Keywords: large language models, drug recommendations, algorithmic bias, pharmaceutical costs, FDA, clinical decision support, generic drugs, originator products, synthetic vignettes, health equity, prescribing patterns, Nature Health