Mount Sinai study: labeled drug ads raised AI selection of advertised treatments from 35% to 47.6%

Researchers at the Icahn School of Medicine at Mount Sinai tested whether a pharmaceutical advertisement could change which treatment a generative model recommended. They gave 12 models from major technology companies, including Google, OpenAI, and Anthropic, the same physician-written clinical cases, either alone or immediately after a clearly labeled real drug ad. Patient details and the list of treatment options stayed fixed. Only the advertisement varied. The team scored 258,559 responses across 33 scenarios for the drug selected and for concordance with clinical guidance.

With an advertisement present, models selected the advertised drug in 47.6% of responses, compared with 35% without an ad: an absolute rise of 12.6 percentage points, about a 36% relative increase over that base rate. In four of 33 scenarios, selection of the advertised drug moved from 0% to 100% after the ad appeared. In a free-text subset, models explicitly mentioned the ad in only 29.9% of answers, while repeating advertising claims more often. Investigators reported that ads did not clearly increase picks of unsupported supplements or clearly suboptimal drugs.

The pattern resembles human pharmaceutical marketing, where influence often appears as a shift among acceptable options rather than an obvious error. Safety checks that only flag unsafe answers would miss it. Girish N. Nadkarni, MD, MPH, the senior author, said a reasonable recommendation can still leave open why one drug was preferred. If a chatbot is treated as a neutral adviser, an ad in the same prompt is commercial context the model may not name. This experiment did not test record-embedded decision support, and it did not measure prescriptions, cost, or health outcomes, so a patient-safety effect remains a hypothesis.

Developers and health systems, the authors said, should test systems with and without ads in the prompt and disclose when advertising is part of the interaction. Funding was not stated in the report. Evidence calibration: this is a controlled experiment on model outputs, not an observational study of real patient chats and not a clinical trial. It can show that a labeled ad is associated with a preference shift under these conditions. It cannot show effects on prescribing, cost, or health. A follow-up needs deployment logs, prespecified guideline-concordance endpoints, and a link from recommendation to what is actually ordered.