Drug Ads Can Sway AI Treatment Picks, Study Finds

Estimated reading time: 8 minutes · Last updated: 11 November 2026

Pharmaceutical advertising can tilt which treatment an AI model recommends, researchers at the Icahn School of Medicine at Mount Sinai found in a study published in The Lancet Digital Health on 11 November 2026. Across 12 AI models, including systems from Google, OpenAI and Anthropic, the advertised drug was chosen in 47.6 percent of responses when a labelled ad came before a clinical case. Without an ad, it was chosen in 35 percent. The answers often still looked clinically defensible. That is the problem: the ad changed the preference among acceptable options without leaving an obvious trace. Models mentioned the ad outright in only 29.9 percent of a sample of free-text answers. For AI treatment recommendations, advertising becomes a hidden variable rather than a visible one.

As AI becomes part of how people get health information, we need to be able to distinguish clinical guidance from commercial influence and be transparent about when advertising is part of the interaction.

Girish N. Nadkarni, MD, MPH, Chief AI Officer of the Mount Sinai Health System

How the Mount Sinai Team Isolated the Effect of One Ad

The design was deliberately narrow. Each of the 12 models received the same physician-written clinical case twice. The first time the case stood on its own. The second time, a real drug advertisement, clearly labelled as one, sat directly in front of it. Nothing else moved. The patient details stayed fixed, and so did the list of treatments the model could choose from.

That setup means any change in the answer can be pinned on the ad itself. A reworded prompt or a different patient cannot explain it. The researchers ran the comparison across 33 clinical scenarios and collected 258,559 responses. They scored each response on two counts: which treatment the model picked, and whether that pick fit clinical guidance.

The models came from major technology companies, among them Google, OpenAI and Anthropic. Using genuine, labelled advertisements rather than invented ones is significant. It tests the version of the problem a consumer product would actually create: a sponsored message that announces itself as sponsored and still shapes the answer.

Why a 12.6-Point Swing Toward the Advertised Drug Matters

With an ad in place, the models selected the advertised drug in 47.6 percent of responses. Without one, the figure was 35 percent. That gap of about 12.6 percentage points is the study's headline effect. It appeared in cases where more than one treatment was clinically appropriate, so the ad was tipping a choice between acceptable options.

The averages hide the sharpest results. In four of the 33 cases, the models went from never choosing the advertised drug to choosing it every time once the ad was added. That is a move from zero percent to 100 percent on a single change to the prompt.

Mahmud Omar, MD, the first author, teaches as a lecturer at the Icahn School, where he belongs to its Windreich Department of Artificial Intelligence and Human Health. The finding surprised him for an unexpected reason. "What I find quite interesting, and even surprising, was how reasonable the answers often appeared," he said. "The model could recommend an appropriate treatment, but its preference among several reasonable options changed after seeing the ad."

Why the Ad's Influence Is Easy to Miss

The researchers also checked whether the models admitted the ad was there. In a subset of free-text answers, only 29.9 percent explicitly referred to it, so about 70.1 percent did not. The models echoed the ad's information and claims more often than they acknowledged that the ad existed.

That combination is what worries the authors. A user sees a sensible-looking recommendation, sometimes carrying claims lifted from promotional copy, with no flag that an advertisement shaped it. Omar said the influence could slip past typical research evaluations. He said it could do the same in everyday use by patients and physicians, because the final answer may still look medically sound.

Girish N. Nadkarni, MD, MPH, the senior author, framed the problem in terms of how the model reached its answer rather than whether the answer was right. "An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer," said Nadkarni. He chairs the school's Windreich Department, directs the Hasso Plattner Institute for Digital Health and is Chief AI Officer of the health system affiliated with the school.

What the Study Did Not Show

The investigators are careful about the limits of their result. Advertising did not clearly push the models toward unsupported supplements, and it did not clearly push them toward drugs that were plainly suboptimal. The effect they measured was a reshuffling among acceptable choices, not a slide into unsafe ones.

The study also stopped at the model's answer. The team did not test whether the shifted recommendations led to different prescriptions or changed what patients decided. It also did not measure effects on costs or health outcomes. Those are the links that would turn a laboratory finding into a measurable harm, and they remain unexamined.

That leaves a real but bounded result. A labelled ad placed next to a clinical question can change what an AI model prefers. Nothing here shows that anyone has been hurt. The authors say their findings do not mean advertising will necessarily lead models toward unsafe or inappropriate treatments.

What the Authors Want From AI Developers and Regulators

The team set out three practical steps for developers and health systems. The first is to run AI systems with and without promotional content and compare the recommendations. The second is to disclose clearly whenever advertising is fed to a model. The third is to keep promotional material apart from clinical advice.

Nadkarni argued that AI is becoming a routine route to health information. As that happens, he said, people need to be able to tell clinical guidance apart from commercial influence, and they need to know when advertising is part of an exchange. His proposal puts the burden on whoever builds or deploys the system, not on the patient asking the question.

Regulators may also have a part to play. Omar, Nadkarni and their co-authors say the results bear on how existing pharmaceutical advertising rules should apply once promotional content becomes part of AI-generated health answers. Next, the team intends to extend the work by examining how AI, and people working alongside it, handle real-world clinical cases.

Does the finding point to real-world harm?

The bottom line

The study by Omar, Nadkarni and their colleagues does not find AI models recommending dangerous drugs because of an ad. It finds something subtler and arguably harder to police. A labelled advertisement placed beside a clinical question moved the advertised drug's share of picks from 35 percent to 47.6 percent. In four cases it flipped the choice completely, while the answers still looked sound. The models named the ad in only 29.9 percent of sampled free-text replies, so a patient or physician would rarely know it was there. Whether the shift reaches prescriptions or health outcomes is untested. Until it is, the case for testing models with and without ads, and for disclosing when ads are present, rests on how the models behave.

What to watch

• Watch for results from the Mount Sinai team's continuing evaluation of AI and human-AI interaction on real-world clinical cases; no date has been set.
• Watch whether regulators take up how pharmaceutical advertising rules apply to promotional content inside AI health answers; no timetable has been given.
• Watch whether AI developers, including Google, OpenAI and Anthropic, publish with-and-without-ad testing of their models as the researchers recommend; no date has been set.

This article is information, not medical advice. Anyone acting on it should speak to a qualified professional.