Mount Sinai Study Finds Labeled Drug Ads Shift Which Treatment AI Models Recommend

When someone types a medical question into a consumer chatbot, the reply usually arrives in the flat, confident register of a clinical reference. Behind that answer, the model is picking among options that often all make sense. New research suggests that the commercial context around a model, and not only its medical knowledge, can decide which of those defensible options it prefers.

A controlled study published on November 11, 2026 in The Lancet Digital Health by researchers at the Icahn School of Medicine at Mount Sinai set out to test a narrow question: can a clearly labeled pharmaceutical advertisement change what a consumer AI model recommends? The team tested 12 AI models built by major technology companies, including Google, OpenAI and Anthropic, and gave each one the identical physician-written clinical case, either on its own or immediately preceded by a real, clearly labeled advertisement for a drug.

The setup matters more than it might appear. Nothing about the patient changed between the two arms. No symptom, lab value, history item or list of treatment choices was altered. The only variable was whether an advertisement appeared in front of the case. That design turns a messy question about marketing and machine behavior into something closer to a laboratory comparison.

The results, announced in a Mount Sinai Health System release, were consistent enough to be uncomfortable. Across 33 clinical scenarios and 258,559 analyzed model responses, the advertised drug was selected in 47.6 percent of responses when an ad was present, compared with 35 percent when it was not. In four of the scenarios the shift was absolute: models went from choosing the advertised drug zero percent of the time to choosing it 100 percent of the time once the advertisement was introduced.

Key Facts

The study's scale is one of its stronger features. Twelve models from major technology companies, 33 physician-written clinical vignettes, and 258,559 individual responses, each assessed both for which treatment was selected and for whether that selection was consistent with clinical guidance. The design isolated advertising as completely as a study of this kind can, holding patient information and the menu of treatment choices constant while varying only the presence of a labeled advertisement placed directly before the case.

The headline effect is the movement from 35 percent to 47.6 percent. Equally notable is the concentration of that effect. Four scenarios flipped from never selecting the advertised drug to always selecting it. Averages hide that kind of clustering, and the four-case flip is the detail that makes the aggregate look less like noise and more like a mechanism.

Medical Economics reported on November 11 that the shift often occurred among treatment options that were all clinically appropriate, which is precisely what makes it hard to detect. A wrong answer is easy to catch. A preference for one reasonable drug over another reasonable drug is not. The researchers also examined whether models were candid about what they had seen. In a subset of free-text responses, models explicitly mentioned the advertisement in only 29.9 percent of responses, while they repeated information or claims contained in the ad more frequently. So the ad was rarely named, but its language traveled.

Fierce Pharma reported on November 11 that Google's models were the most susceptible to the ad effect and that Anthropic's barely moved. The same report noted that the advertisements did not push models toward unsupported supplements or clearly inferior drugs. That asymmetry matters: the vulnerability appears to live in the ranking of plausible options, not in a willingness to endorse something the evidence does not support.

The authors were careful about the boundaries of their own work. The investigators said the findings do not mean pharmaceutical advertising will necessarily cause AI models to recommend unsafe or inappropriate treatments, and the study did not examine effects on prescriptions, patient decisions, costs or health outcomes. The analysis was supported in part by grants from the National Center for Advancing Translational Sciences and the National Institutes of Health.

Analysis

What this really means is that the measurable risk is not a chatbot inventing bad medicine. It is a chatbot quietly reordering good medicine. That distinction changes how the problem should be tested and governed. Standard evaluations ask whether a model's answer matches clinical guidance, and on that test a model nudged toward an advertised drug that is already clinically appropriate will pass. The answer can be defensible and still be the product of an advertisement.

Mahmud Omar, MD, a lecturer in the Windreich Department of Artificial Intelligence and Human Health and the study's first author, described exactly this trap. The model could recommend an appropriate treatment, he said, but its preference among several reasonable options changed after seeing the ad. Because the final answer could still appear medically sound, that influence could be easy to miss. That observation is the study's real finding, more than any single percentage.

Girish N. Nadkarni, MD, MPH, chair of the department and chief AI officer of the Mount Sinai Health System, framed the issue as one of transparency rather than competence. An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer, he said. As AI becomes part of how people get health information, he added, users need to be able to distinguish clinical guidance from commercial influence and to be told when advertising is part of the interaction.

The bigger picture here is that consumer AI has arrived at the same structural fork that search and social media reached years ago, where the thing being recommended and the thing being paid for are not always the same. Search engines eventually separated paid placements with visible labels. The Mount Sinai result suggests that labeling alone may not be enough for a language model, because a label can be present, followed, and then effectively ignored in the output. The ad was disclosed in the prompt and mentioned in fewer than a third of sampled free-text answers. What the models did instead was absorb and restate the ad's claims.

Why It Matters

The regulatory question is now concrete. News-Medical reported on November 11 that the findings may have implications for regulators applying existing pharmaceutical advertising rules to AI-generated health information. Those rules were written for media in which an advertisement and the editorial content around it occupy visibly different space. In a chatbot, both arrive in the same block of text, produced by the same model, in the same voice. Fierce Pharma reported that the results may matter as AI companies add advertising to consumer products, which turns an academic finding into a product design decision.

Medical Xpress reported on November 11 that the investigators recommended three practical steps: testing AI systems with and without promotional content, disclosing when advertising is presented to a model, and keeping promotional material separate from clinical advice. None of these is technically exotic. The third is the most consequential, because it accepts that a sufficiently capable model may treat an in-context advertisement as relevant clinical context rather than as decoration.

For patients, the practical lesson concerns the shape of the risk. A recommendation that names a familiar, appropriate drug is not evidence that the recommendation was uncommercial. The authors stress that the ads did not clearly increase recommendations for unsupported supplements or clearly suboptimal drugs, so this is not a story about dangerous advice. It is a story about preference formation under commercial pressure, occurring in a channel that millions of people already treat as a neutral second opinion.

Next Up

The obvious next step is replication outside a controlled vignette. The Mount Sinai team varied only the presence of a clearly labeled advertisement, which is the right way to isolate a cause but not the way real consumer products surface promotional material. Whether the effect survives different placements, formats and multi-turn conversations is an open question, and so is the gap between Google's models and Anthropic's. A large difference between vendors is a lead worth chasing, because it implies the behavior is engineerable rather than intrinsic.

The researchers also left the downstream questions untouched: prescriptions, patient decisions, costs and health outcomes. Until those are measured, the honest conclusion is that a labeled ad can move a model's choice among reasonable treatments, that the models seldom say so, and that testing AI systems with and without promotional content should become routine rather than exceptional.