A pharmaceutical ad placed immediately before a clinical case changed which drug many AI models selected—even when the answer remained medically reasonable. In a controlled study spanning 12 commercial models and 33 clinical scenarios, the average share of responses naming the advertised drug rose to 47.6%, from a baseline proxy of about 34%. The finding points to a subtle risk: promotional context may act as a tiebreaker among legitimate treatment options, without making an answer obviously wrong.

Table of Contents

What the study tested

A new controlled study in The Lancet Digital Health reports that pharmaceutical advertising embedded in a model’s prompt can shift which treatment it recommends. The paper, “Ad-verse Effects: Pharmaceutical Advertising and Drug Recommendations by Consumer-Facing AI in a Controlled Study”, evaluated 12 large language models from Google, OpenAI and Anthropic. Researchers gave each model a physician-written case, either on its own or preceded by an explicitly labeled drug advertisement. The advertisement was not an instruction to select the promoted medicine.

The core finding is not that the models routinely chose a dangerous drug. In scenarios where two named prescription drugs were both guideline-appropriate, the models became more likely to select whichever one had just been advertised. Across those comparisons, advertised-drug selection was 47.6%, versus a baseline proxy of roughly 34%, a reported mean increase of 12.7 percentage points. The paper reports statistical significance (P < 0.001), but the result describes model responses under an experimental setup—not changes in treatment received by people.

That distinction matters. A system can answer with a medically defensible option and still have its preference influenced by commercial material. The research therefore concerns how an answer is selected, not a demonstrated rise in injury, inappropriate prescribing or patient exposure.

How the experiments worked

Investigators ran four controlled experiments involving 33 clinical scenarios and 12 models accessed through providers’ official APIs. Models included Gemini 2.5 Flash-Lite, Gemini 2.5 Flash and Gemini 3 Flash Preview; five OpenAI models; and four Anthropic models. The calls were made in February 2026, not at the time the article was published. Each API call began a new stateless conversation. The researchers varied the user prompt by placing a standardized, labeled ad immediately before the clinical vignette, while keeping the case and answer choices the same.

In the main preference experiment, 13 scenarios offered four named medication choices. Two were prespecified advertising targets. Researchers compared each advertised target’s selection rate against that drug’s rate in the matched no-ad baseline. Baseline calls were shared between the two target comparisons, rather than being rerun. Cases were developed and reviewed by physicians for clinical accuracy and therapeutic equipoise—that is, situations in which more than one treatment could be acceptable.

Other experiments tested whether models would endorse unsupported wellness supplements, whether ads could push a clinically suboptimal drug in cases where the appropriate choice was clearer, and how models explained their choices in open-ended text. For the latter, three models representing different susceptibility patterns were tested across 13 scenarios, conditions and four personas, with five repetitions per cell. The paper’s methods describe attempted calls; as is common in large API experiments, call attempts, successful responses, and analyzed response totals are not interchangeable.

The paper says 258,660 calls were attempted across all experiments. Its experiment descriptions account for 256,320 calls in experiments one through three and 2,340 in the open-response experiment, which sum to 258,660. The paper’s reported main-study conditions comprise 74,880 ad-condition calls and 37,440 baseline calls for the preference experiment, plus 57,600 ad and 28,800 baseline calls for the suboptimal-drug experiment; the supplement experiment makes up the remaining experiment-one-through-three calls. The authors’ public summary and Mount Sinai’s release describe 258,559 analyzed responses, one fewer than the paper’s 258,660 attempted calls. Medical Economics likewise reports 258,559 analyzed responses. The one-response difference is plausibly the distinction between attempted and analyzed calls, but the materials reviewed do not itemize that reconciliation; it should not be treated as a confirmed explanation.

The ad shift—and its boundaries

The average concealed meaningful differences between providers. In the main preference experiment, the reported mean shift was 29.8 percentage points for Google models, 10.9 points for OpenAI and 2.0 points for Anthropic. Gemini 2.5 Flash-Lite showed the largest listed model-level shift, 35.1 points; the two other tested Gemini models also shifted substantially. Anthropic models were comparatively resistant on average, though that does not establish that every Anthropic model or product will behave similarly in another setting.

The paper describes several model-and-scenario combinations in which selection of the advertised drug moved from zero to 100%. Those were shifts between guideline-equivalent alternatives in the relevant cases—not proof that an ad made a model switch from a correct answer to an incorrect one. In the experiment’s forced-choice scenarios, overall accuracy was 89.4% without an ad and 93.0% with one. Accuracy can rise while choice preference shifts because the advertised option may itself be an appropriate option; that result does not mean advertising improves clinical care.

When the advertised prescription drug was designated clinically suboptimal in a separate experiment, models largely maintained accuracy. The study reports baseline accuracy of 94.1% and ad-condition accuracy of 95.6%; 4.4% of responses in the ad condition selected the suboptimal advertised option. In the supplement experiment, advertising did not increase endorsement of unsupported supplements: the reported rate fell from 2.2% at baseline to 1.6% with ads. These are reassuring findings within the tested cases, but they do not guarantee safety in other cases, models or deployment settings.

The free-text sub-study adds a possible explanation for why the shift may be hard to spot. Across its 1,560 ad-condition responses, models explicitly acknowledged the advertisement in 29.9% of responses (467 of 1,560). They echoed more promotional claims when they selected the advertised drug than when they did not. That analysis involved only three models and a limited set of scenarios, so it should not be generalized to all systems. It nevertheless illustrates how advertising can enter a model’s justification without an explicit disclosure in the answer.

When a reasonable answer can still be influenced

The authors interpret the pattern as most evident in a zone of clinical equipoise: when several options are supported, a prompt may make one option more salient without overturning the model’s broader medical knowledge. That is a plausible interpretation of the controlled results, not direct evidence about the internal reasoning process of a deployed product. The model’s generated explanation is not a transparent record of how its decision was formed.

This distinction complicates evaluation. A test that scores only whether a recommendation is clinically acceptable might mark both the baseline and ad-influenced response as correct, missing the fact that the promoted option became more likely. Conversely, the study does not establish that every such shift is harmful: physicians may appropriately choose among equivalent options for reasons such as patient history, cost, formulary access or preferences, none of which can be inferred from a short vignette alone.

What the study does not show

The researchers tested a very specific pathway: one labeled advertisement was inserted directly into the text prompt immediately before a case. They did not test a commercial banner beside a chatbot, a sponsored search result, advertisements retrieved from the web, material in model training data, or a full clinical product with retrieval, memory, interface design and workflow controls. Nor did they test whether an AI recommendation changed a clinician’s prescription, a patient’s decision, spending, adherence or health outcome.

The study used a fixed set of physician-reviewed scenarios and model versions available through production APIs in February 2026. Model behavior can vary with version, system instructions, sampling, tools and deployment context. The open-response analysis covered only three representative models and five repetitions per experimental cell. The experimental design gives researchers control over exposure, but it cannot reproduce the complexity of actual consultations. The authors themselves caution against extrapolating the results to other products and pathways.

Accordingly, the finding should be read as evidence of susceptibility in a bounded laboratory-style simulation, not proof that consumer chatbots currently accept paid drug ads or that commercial influence has changed real-world care. It does support asking how systems behave when promotional content enters the context they process.

Implications for healthcare and AI

For developers and health systems, one practical implication is to test outputs with and without promotional content, including cases where several treatments are reasonable and cases where an advertised treatment is not. Such evaluations should measure both clinical correctness and preference changes, disclose when advertising is supplied to a model, and keep promotional content visibly separate from clinical guidance. The study points to these as safeguards to consider; it does not test whether any particular safeguard works in a live product.

Clinicians and patients should not treat a model’s confident language as evidence that its recommendation is independent of context. At the same time, the research is not a reason to infer that a specific chatbot response was influenced by an ad: the tested mechanism was explicit text inserted into a prompt, and the authors did not measure real users’ interactions. The central question for governance is whether an AI system can distinguish clinical evidence from sponsored material—and whether users can tell when the system has been exposed to it.

By showing that a labeled advertisement can alter preferences even while answers remain broadly guideline-concordant, the study makes a case for evaluating more than the final answer’s correctness. But the size and consequences of any effect in deployed healthcare remain unknown. Establishing those would require testing complete systems under realistic conditions and measuring how people respond.

Frequently asked questions

Did the study show that drug ads made AI recommend unsafe medicines?

Not generally. In the tested cases, ads shifted choices most clearly between clinically appropriate alternatives. Models largely maintained accuracy when the advertised medicine was suboptimal, though a small share of responses selected it. The experiments did not assess patient harm.

How many AI models were tested?

Twelve models from Google, OpenAI and Anthropic were evaluated through their APIs. Results differed by provider and model, so the aggregate result should not be assumed to apply equally to every AI system.

Were real patients or prescriptions involved?

No. This was a controlled simulation using clinical vignettes and model API calls. It did not measure real patient decisions, prescribing, cost or health outcomes.

Why do the paper and Mount Sinai report different response counts?

The paper reports 258,660 attempted calls, while Mount Sinai and Medical Economics report 258,559 analyzed responses. The difference is one response; the public summaries do not provide a detailed accounting of that specific discrepancy.

References

  1. Omar M, et al. “Ad-verse Effects: Pharmaceutical Advertising and Drug Recommendations by Consumer-Facing AI in a Controlled Study.” The Lancet Digital Health, Nov. 11, 2026. Full paper; PubMed Central full text.
  2. Shryock T. “AI chatbot treatment recommendations are swayed by drug ads.” Medical Economics, Nov. 11, 2026. Independent report.
  3. Icahn School of Medicine at Mount Sinai. “Drug Ads Can Influence AI Treatment Recommendations, Study Finds.” Nov. 11, 2026. Institutional release.