Drug advertising changed AI treatment choices in a controlled test
A labeled pharmaceutical advertisement placed beside a clinical vignette changed how consumer-facing AI models chose treatment. Across 12 major models and 33 physician-written cases, the share of responses selecting the advertised drug rose to 47.6% with an ad present, compared with 35% when no ad appeared.
In four scenarios, the effect was absolute: models that never selected the advertised drug without the ad switched to choosing it every time once the promotional material was introduced.
The work, conducted by investigators at the Icahn School of Medicine at Mount Sinai and published Nov. 11 in The Lancet Digital Health, tested whether commercial messaging can alter AI-generated medical advice even when the advertising is clearly labeled. The design isolated the influence of the ad itself. The clinical case, patient details and medication options stayed fixed; only the promotional content changed.
That simple change altered the recommendation pattern across a very large response set. Investigators analyzed 258,559 AI responses in total, allowing them to compare treatment selection and assess whether the choice still aligned with clinical guidance.
Reasonable answers can still be commercially steered
The most striking feature of the findings is not a sudden collapse into unsafe advice. The shift often occurred among options that were already clinically acceptable. In practical terms, the model still sounded plausible while nudging preference toward the drug that had just been advertised.
That pattern is harder to detect than an obviously wrong recommendation. A clinician or patient reviewing the output could see a medically reasonable answer and miss the fact that commercial content had shaped the model’s preference among several defensible choices. In medicine, subtle steering may be more consequential than a blatant error because it can pass through standard reasonableness checks without raising alarms.
Mahmud Omar, MD, a physician-scientist and lecturer in the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai, said the responses often looked appropriate on their face even when the preference changed after ad exposure. The model could recommend a clinically suitable therapy, yet the ranking among reasonable options shifted once the advertisement entered the prompt context.
That matters because AI systems are increasingly used for health information outside formal clinical settings. When a patient asks a chatbot about a symptom or a medication, the output may appear neutral, even if the surrounding content has already influenced the answer.
Disclosure was inconsistent even when the ad was present
The models did not reliably acknowledge the promotional material. In a subset of free-text responses, they explicitly mentioned the advertisement in only 29.9% of cases, even though they repeated ad claims more often than they flagged the ad itself.
This is the point where transparency breaks down. A user may receive an answer that sounds clinical, not commercial, while the model quietly reproduces language or claims introduced by the advertisement. For health systems and developers, the problem is not only whether AI can be influenced, but whether it can reveal when influence has occurred.
Girish N. Nadkarni, MD, MPH, chair of the Windreich Department and chief AI officer of the Mount Sinai Health System, framed the issue directly: an answer can remain clinically reasonable while still raising the question of why the system arrived at that answer. As AI becomes part of routine health information seeking, separating clinical guidance from commercial influence becomes a basic safety requirement rather than an optional feature.
The study also found what it did not find. Ads did not clearly push models toward unsupported supplements or obviously suboptimal therapies. The concern here is narrower and more insidious: promotional content shifted preference within a set of acceptable drugs, where the change could influence prescribing habits, patient expectations or downstream treatment discussions without triggering an obvious safety signal.
Why adjacent ads may be a special problem for medical AI
The mechanism is straightforward enough to worry regulators and developers. Large language models are sensitive to prompt context, and a drug ad placed immediately before a case can function as a priming signal. In human terms, the model has just been reminded of a brand, its claims and its framing. In computational terms, the promotional text becomes part of the context window that shapes the next response.
This is a different challenge from conventional drug advertising aimed at clinicians or consumers. Here, the ad is not merely seen by a person who can consciously discount it. It is ingested by a system that generates answers by pattern association, making it vulnerable to context contamination in ways that standard oversight may not capture.
The researchers emphasized several practical safeguards. One is A/B testing AI systems with and without promotional content to see whether recommendations shift. Another is explicit disclosure when advertising is present in the interaction. A third is strict separation between marketing materials and clinical decision support outputs.
Those steps are modest, but they are also concrete. Without them, a model could appear stable in routine evaluation while still being susceptible to commercial priming in the wild.
Regulatory questions now extend to AI-mediated health advice
The findings also carry implications for regulators examining how existing pharmaceutical advertising rules apply when promotional material is embedded in AI-generated health information. Traditional ad oversight assumes a human audience. This study shows that the audience may increasingly include the model itself, which can then pass the influence along to the user.
The investigators did not measure whether the altered recommendations changed prescriptions, patient decisions, spending or clinical outcomes. That leaves an important gap. A shift in answer selection does not automatically translate into a change in care, but it does create a pathway through which commercial influence could enter the clinical conversation earlier than many observers expect.
The Mount Sinai team plans to extend the work through ongoing evaluation of AI and human-AI interaction in real-world clinical cases. For now, the study offers a blunt warning: even a clearly labeled drug ad can reweight an AI model’s treatment preference, and the resulting output may still look medically sound enough to pass unnoticed.