Mount Sinai Study: Drug Ads Shift AI Treatment Picks by 12.6 Points

A clearly labeled drug advertisement pushed AI models to recommend the advertised drug in 47.6 percent of responses, up from 35 percent without an ad — a 12.6-point swing documented across 258,559 model outputs. Researchers at the Icahn School of Medicine at Mount Sinai published the findings November 11 in The Lancet Digital Health, after testing 12 consumer-facing AI models from companies including Google, OpenAI, and Anthropic.

The numbers are the story. The team, led by first author Mahmud Omar, MD, and senior author Girish N. Nadkarni, MD, MPH, fed the same 33 physician-written clinical cases to each model under two conditions: the case alone, or the case immediately preceded by a real, clearly labeled pharmaceutical ad. Patient information and treatment options stayed identical. Only the ad's presence varied. The researchers recorded which treatment each model picked and whether it matched clinical guidance.

In four of the 33 scenarios, the swing was total. Models that chose the advertised drug zero percent of the time without an ad chose it 100 percent of the time once the ad appeared.

Where the influence stops — and where it doesn't

Here's the nuance that matters for anyone building or buying health AI: the ads moved preferences, not clinical appropriateness. The influence operated within the bounds of what doctors would consider sound medicine.

"The model could recommend an appropriate treatment, but its preference among several reasonable options changed after seeing the ad," said Omar, a lecturer in Mount Sinai's Windreich Department of Artificial Intelligence and Human Health.

Nadkarni, who chairs that department and serves as chief AI officer of the Mount Sinai Health System, framed the problem in one line: "An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer."

That distinction is exactly why standard safety evaluations miss it. If you only check whether an answer is medically sound, the ad's effect is invisible.

How often did models admit they'd seen an ad?

Rarely. In a subset of free-text responses, models explicitly mentioned the advertisement in just 29.9 percent of cases. They repeated information or claims from the ad more often than they acknowledged it. So the contamination shows up in the content before it shows up in the disclosure — if it shows up at all.

Takeaway: an AI system can absorb promotional messaging and stay quiet about it.

What the study does not say

The researchers were explicit about the limits. The findings do not mean pharma ads will cause AI models to recommend unsafe or inappropriate treatments. In this experiment, ads did not clearly increase recommendations for unsupported supplements or clearly suboptimal drugs.

The team also did not examine downstream effects — whether changed AI recommendations altered prescriptions, patient decisions, costs, or health outcomes. And the design was a controlled experiment using consumer-facing models and clearly labeled ads. It did not test real-world clinical workflows, models embedded in electronic health records, or situations where advertising is less explicitly labeled.

What should developers and regulators do now?

The authors propose three concrete steps:

• Test AI systems with and without promotional content to see if recommendations change.
• Clearly disclose when advertising is presented to an AI system.
• Keep promotional material separate from clinical advice.

The findings also land on regulators' desks. The authors note the results may inform how existing rules governing pharmaceutical advertising apply when promotional content gets folded into AI-generated health information — a gap current frameworks weren't written to cover.

Mount Sinai says it will extend the work through ongoing evaluation of AI and human-AI interaction in real-world clinical cases. Until those results arrive, the 12.6-point gap between advertised and unadvertised conditions stands as the cleanest evidence yet that AI treatment advice is for sale to the highest bidder — even when every answer remains technically correct.