Millions of people now ask an AI chatbot before — or sometimes instead of — asking a doctor. They describe their symptoms and ask what treatment might help. The answer arrives in seconds, written in calm, confident language. But a new study from the Icahn School of Medicine at Mount Sinai suggests that what sits next to your question can quietly change the answer you get.
The study, published on 11 November 2026 in The Lancet Digital Health, tested what happens when a clearly labelled pharmaceutical advertisement appears alongside a clinical question. Researchers gave 12 leading AI models the same physician-written patient cases, sometimes alone and sometimes with a real drug ad placed immediately before the case. Nothing else changed.
The result was striking. With no ad, the models picked the advertised drug in 35 percent of responses. With the ad present, that rose to 47.6 percent. In four cases, a model went from never choosing the advertised drug to choosing it every single time. And this happened even though the ad was clearly labelled as advertising.
One point matters most before going further. This study does not prove that drug ads make AI recommend unsafe treatments. The researchers are explicit about that, and the data back them up. What it does prove is subtler: advertising can tilt an AI's choice among several reasonable options, and the final answer can still look perfectly sound.
A carefully controlled experiment
The strength of this study is its design. It was not a casual test of a few chatbots, nor was it based on real patients, where countless factors could muddy the picture. Instead, the Mount Sinai team built a controlled experiment that isolates one variable and shows exactly what it does.
The team tested 12 AI models from the three biggest AI companies — Google's Gemini, OpenAI's GPT, and Anthropic's Claude models — the kinds of systems that power the chatbots people actually use. Across 33 clinical scenarios, the researchers analysed 258,559 responses, a scale that makes the findings hard to dismiss as a fluke. Each call started a fresh conversation with no memory of previous questions, so the ad was the only difference between the two conditions.
The cases were written and reviewed by physicians using current clinical guidelines, and the models were asked to recommend one of four named drugs. In six scenarios, the researchers deliberately chose situations of genuine clinical equipoise — cases where two different drugs were both acceptable, with no single obviously right answer. It is exactly in those grey areas that an ad appears to have the most room to tip the balance. The team also checked whether ads could push models toward unsupported supplements, or toward drugs that were clearly suboptimal for the patient.
The headline numbers
The main finding is easy to state. In the first experiment, the advertised drug was selected in 26,150 of 74,810 baseline opportunities — 35.0 percent — and in 35,628 of 74,815 advertisement calls — 47.6 percent. That is an increase of 12.7 percentage points, with a 95 percent confidence interval of 9.1 to 16.2 and a p-value below 0.0001 — a large effect, very unlikely to have appeared by accident.
But the average hides the most dramatic cases. In four model-and-scenario combinations, the advertised drug went from being selected in zero percent of responses to 100 percent once the ad was introduced. One example named in the paper involved dapagliflozin, a drug used in heart failure, which jumped from never chosen to always chosen.
Just as important is what did not happen. The ads did not clearly increase recommendations for unsupported supplements — the change was minus 0.6 percentage points, statistically indistinguishable from zero. Nor did they push models toward advertised drugs that were clearly suboptimal, where the change was just 0.7 percentage points. The models were not being dragged into bad medicine. They were being steered within the bounds of acceptable medicine, and their overall agreement with clinical guidelines stayed high.
Why a reasonable-looking answer is the real problem
The most unsettling detail concerns what the models said — and did not say — in their own words. In a free-text experiment, the models mentioned the advertisement in only 29.9 percent of responses. Yet when they selected the advertised drug, they incorporated 52.7 percent of the advertising claims into their reasoning, compared with 19.4 percent in other responses.
This is why the study's lead author, Dr Mahmud Omar, a physician-scientist at Mount Sinai, described the results as both interesting and surprising. "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. Because the final answer could still appear medically sound, that influence could be easy to miss."
That observation cuts to the core of how we usually judge AI health advice. Most checks ask a simple question: was the answer correct? If the drug suggested is a legitimate treatment, the answer passes. But this study shows correctness is not enough. Two answers can both be clinically reasonable and yet differ in a way that matters — one drug may cost far more, or carry different side effects. When the difference is decided by an advertisement rather than by the patient's needs, something has gone wrong even though no rule of medicine was broken.
Senior author Dr Girish N. Nadkarni, who chairs Mount Sinai's Windreich Department of Artificial Intelligence and Human Health, added: "An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer."
What this study does not prove
Good science is as much about knowing its limits as its findings, and this study has clear ones. First, the experiment was controlled, not real. The ads were placed inside the message the model read, immediately before the clinical case. That is not how advertising works in the chatbots people use today, where ads typically appear as sponsored placements near or within an answer. The study did not test any live commercial product, so its exact effect size in a real product is unknown.
Second, the study measured what the models recommended, not what happened next. It did not examine whether changes in AI recommendations affected real prescriptions, patient decisions, costs, or health outcomes. The chain from an ad to a worse outcome for a real person has not been shown — only the first link in that chain.
Third, the models tested are general-purpose systems, not dedicated medical tools, and the scenarios were written cases rather than real patients. None of this weakens the central finding. It simply defines it honestly: advertising can shift AI treatment recommendations in a controlled setting, among reasonable options, often without disclosure.
Why this matters now — and what safeguards could look like
The timing is no accident. About one in five adults in the United States has used an AI chatbot for health information, according to research cited in the paper, and two in three US physicians report using healthcare AI. Health systems are weaving these models into patient-facing services, and the companies behind them are introducing advertising and health features at the same time.
For patients, the stakes are practical. Many people turn to AI precisely when they are anxious, in pain, or facing a difficult decision — exactly the moments when a confident-sounding answer carries the most weight. If that answer has been quietly tilted toward one brand, the patient may walk into a doctor's appointment already convinced of a particular drug. The same dynamic applies to other health tools reaching patients directly: our report on a new blood test that spots ovarian cancer earlier showed how quickly diagnostic technology is reaching patients, and AI advice is following the same path.
The researchers suggest several practical steps. First, AI systems should be tested with and without promotional content to see whether their recommendations change — and companies could publish those results just as they publish accuracy scores. Second, advertising should be clearly disclosed whenever it is part of an interaction with an AI system. But labelling alone is not enough: the study's ads were clearly labelled, and the shift happened anyway. Third, promotional material should be kept separate from clinical advice. An ad should never sit inside the same message or reasoning chain as a medical question.
Regulators have a role too. Current rules govern prescription-drug advertisements shown to consumers, but they do not specifically address promotional content that enters an AI model's input. Future evaluations of medical AI should also test stability — does the model give the same recommendation when the commercial context changes? A system whose medical judgement wobbles with the advertising around it is not a reliable clinical tool, however high its accuracy score.
What you can do when you ask AI about medicine
None of this means you should never ask an AI chatbot a health question. Used well, these tools can help you understand a diagnosis, prepare questions for your doctor, and make sense of medical language. But the study suggests a few habits worth adopting. Ask the same question in more than one way, and if possible in more than one tool, and compare the answers. Ask the chatbot to explain why it recommended a particular drug, and what the alternatives are, including their costs and side effects.
Most importantly, treat AI as a starting point, not a verdict. A chatbot cannot examine you, cannot see your full medical history, and cannot weigh the trade-offs that matter to your life. Those are decisions for you and your doctor together. The same principle applies to every new health technology: as our look at a DNA test that changes health habits showed, new tools are most useful when they inform a conversation with a professional rather than replacing it. And treatment choices are rarely black and white — as our examination of modern HRT and heart risk, based on data from 52,000 women, showed, the right choice depends on the individual.
The Mount Sinai study is, by the authors' own account, the first controlled evaluation of labelled pharmaceutical advertising in the context of general-purpose AI models. If these tools are to be trusted advisers, they must be free to give the answer the evidence supports — not the answer an advertiser paid to promote. A system that quietly favours one drug over an equally good one is not lying, but it is not neutral either. And in medicine, neutrality is not a luxury. Patients deserve to know which kind of advice they are getting.
Disclaimer
This article is general information only and is not medical advice. It describes a research study and offers general guidance about using AI tools for health questions. It is not a substitute for professional medical advice, diagnosis, or treatment. If you have questions about a medicine or a treatment, please consult a qualified healthcare professional, such as your doctor or pharmacist, before making any decision.