Medical Artificial Intelligence · global
Drug Ads Change AI Treatment Recommendations: Reasonable Answers May Still Be Influenced by Commercial Messages
A controlled study covering 12 AI models found that adding clearly labeled drug advertisements made models more likely to select the promoted drugs. When preferences shift among several appropriate treatments, checking only whether an answer is correct may miss commercial influence.
When asking AI about treatment options, a seemingly reasonable answer may also bear traces of advertising. A team at the Icahn School of Medicine at Mount Sinai in the United States found that even when patient information and available treatments were identical, placing a drug advertisement before a case increased the proportion of responses in which AI selected the promoted drug. This raises another question for medical AI evaluation: if a recommendation aligns with medical knowledge, is that enough to show that the judgment is free from commercial influence?
The study was published on November 11 in *The Lancet Digital Health*. Researchers tested 12 models from companies including Google, OpenAI, and Anthropic, using 33 physician-written clinical scenarios and analyzing a total of 258,559 responses. Each scenario was tested under two conditions: without advertising, and with a real, clearly labeled drug advertisement placed before the case. Patient information and treatment options remained unchanged to isolate the influence of the advertisement itself.
With advertising, the proportion of responses in which models selected the advertised drug was 47.6%, compared with 35% without advertising—a difference of 12.6 percentage points. The study also recorded four extreme cases in which adding an advertisement shifted the selection rate for the promoted drug from 0% to 100%. These figures reflect changes in responses during controlled testing and cannot be treated directly as the magnitude of an increase in actual prescribing.
The harder-to-detect influence emerged among treatments that were all clinically reasonable. Both Mount Sinai’s description of the study and reporting by *Medical Economics* noted that advertising could change models’ preferences among appropriate options without making their answers appear obviously wrong. As a result, evaluations that ask only whether recommendations comply with clinical guidelines may miss the commercial influence behind those choices.
Models also did not necessarily let readers know that an advertisement had appeared. In a subset of free-text responses examined in the study, only 29.9% explicitly mentioned the advertisement, while restating advertising information or claims was more common. However, mentioning an advertisement and accurately explaining how it influenced a judgment are two different things; this percentage alone cannot fully measure transparency in the recommendation process.
The limits of the findings are equally clear: the study did not find that advertising clearly increased models’ recommendations of unsupported supplements or clearly inferior drugs, nor did it track actual prescriptions, patient decisions, costs, or health outcomes. It shows that models are susceptible to influence under specific input conditions, but does not provide sufficient grounds to infer that all medical AI services will exhibit the same degree of bias.
The research team recommended that developers and healthcare institutions compare systems’ recommendations before and after exposure to promotional content, disclose whether advertisements enter the information read by models, and keep promotional materials separate from clinical recommendations. For regulators, the next question to clarify is how existing advertising rules should apply when drug promotion is absorbed into a seemingly neutral AI answer. The study suggests that trustworthiness in medical AI also depends on whether people can identify what information has influenced a recommendation.