Drug ads shifted AI treatment choices in a controlled Mount Sinai study

A clearly labeled drug advertisement changed which treatment AI models recommended in a controlled Mount Sinai study. Across 33 clinical scenarios, models selected the advertised drug in 47.6% of responses with an ad, versus 35% without one. Researchers warned that the influence could be difficult to spot because the resulting advice could still look medically reasonable.

The research, published November 11, 2026, in The Lancet Digital Health, tested 12 models from companies including Google, OpenAI and Anthropic. Its focus was not simply whether an AI could identify an acceptable treatment, but whether promotional material changed its choice.

Same patient, same options, one added advertisement

Researchers gave the models physician-written clinical cases under two conditions: the case alone, or the same case immediately preceded by a real pharmaceutical advertisement. The ad was clearly identified as advertising. Patient information and the available treatment choices stayed unchanged, leaving the advertisement as the only variable in the comparison.

The team analyzed 258,559 responses across the 33 scenarios. It recorded which treatment each model selected and assessed whether that choice followed clinical guidance. Researchers also examined whether the models acknowledged the advertisement and whether their answers repeated information or claims from it.

Some changes were much sharper than the overall result. In four cases, the advertised drug went from being selected in none of the responses without an ad to all of them after the ad was added.

An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer

Girish N. Nadkarni, senior study author and Mount Sinai Health System chief AI officer

Influence without an obvious medical error

First author Mahmud Omar, a physician-scientist at Mount Sinai’s Icahn School of Medicine, emphasized that the shift could occur among several appropriate treatments. He warned that evaluations checking whether an answer was medically sound could miss a change in preference caused by the ad. That concern also applied to use by patients and physicians.

The answers did not consistently make that advertising exposure visible. In a subset of free-text responses, models explicitly mentioned the ad in 29.9% of answers. They repeated information or claims from the advertisement more frequently than they acknowledged it. That percentage describes the free-text subset, not every response in the experiment.

A recommendation shift is not a measured patient outcome

The investigators cautioned against reading the findings as proof that drug ads necessarily make AI advice unsafe or inappropriate. In this experiment, advertisements did not clearly increase recommendations for unsupported supplements or clearly suboptimal drugs. The observed effect was a change in selection, not a demonstrated increase in harmful treatment advice.

Nor did the study examine whether those changed recommendations affected prescriptions, patient decisions, costs or health outcomes. It therefore does not establish that exposure to a drug ad through an AI interaction changed what anyone took, what they paid or how their health changed.

The researchers proposed three practical steps for AI developers and health systems:

• Test systems with and without promotional content to check whether their recommendations change.
• Clearly disclose when advertising has been presented to an AI system.
• Keep promotional material separate from clinical advice.

The Mount Sinai team plans to extend the work through its ongoing evaluation of AI and human–AI interaction in real-world clinical cases.