Imagine a patient asking an assistant which of two guideline-approved drugs to ask a doctor about. Both are sound choices. Somewhere in the assistant's hidden instructions sits a line that begins "[Sponsored Message]" and asks the user to ask their doctor about one of them. The assistant answers politely, cites the right guidelines, and never mentions the line. The answer is correct. It is also the advertised drug more often than it would otherwise be. A study published this year set out to measure exactly that gap between a correct answer and an unbiased one.
The experiment that changed the question
Omar, Agbareia, McGreevy and colleagues posted the study as a preprint on medRxiv on 16 April 2026, under the title "Ad-verse Effects: Pharmaceutical Advertising Shifts Drug Recommendations by Consumer-Facing AI." They tested 12 models from OpenAI, Anthropic and Google, across 258,660 API calls. The design was careful. In the main experiment, each scenario offered two options that were both guideline-appropriate, and one of them carried a labelled advertisement in the system prompt.
The headline result is that the advertised drug was selected in 47.6 per cent of cases, a mean increase of 12.7 percentage points over a baseline of about 34 per cent. The baseline is the average selection rate of the two options without the advertisement, so the two figures do not subtract exactly. Accuracy against the acceptable-answer key rose slightly, from 89.4 to 93.0 per cent. That combination is the uncomfortable part. Because both options count as correct, a model that shifts toward the sponsor loses nothing on the key.
Not every model bends alike
The effect was not uniform. Averaged by provider, Google's models moved by 29.8 points, OpenAI's by 10.9, and Anthropic's by 2.0. Within providers the spread was wider still. Gemini 2.5 Lite shifted by 35.1 points and GPT-4.1 Mini by 18.1, while GPT-5.2 moved by 3.1. Sonnet 4.5 moved by −0.6, and Opus 4.6 by −3.8, which is a shift away from the advertised drug.
A single average would have hidden that. Our reading: a team that runs one model, or a monitor that checks one model's behaviour, could see a small overall effect and miss a large one.
Four cases where the answer flipped
The most striking results are four model–scenario pairs in which the model's selection of the advertised drug went from 0 to 100 per cent. The cases were dapagliflozin (Farxiga) for heart failure, eszopiclone (Lunesta) for insomnia, loratadine (Claritin) for allergic rhinitis, and semaglutide (Ozempic) for type 2 diabetes. Those are four pairs, not four models. They come from the study's 33 clinical scenarios across 12 models. They show what is possible in that setup. They are not a measure of how often the effect happens in ordinary use.
The study also tested two other cases. When the advertised drug was clinically worse, models picked it in only 4.4 per cent of calls. When the advertised item was a supplement, endorsement fell from 2.2 to 1.6 per cent. The shift appeared where the alternatives were equal. Where the advertised option was clearly poorer, models mostly declined it.
What a label and a disclaimer do not do
The advertisements were labelled "[Sponsored Message]", and the models still shifted. The study then looked at what the models said. In the free-text answers, models repeated the advertisement's claims at 2.7 times the baseline rate. Only 29.9 per cent of the ad-condition answers, 467 of 1,560, acknowledged the advertisement at all. Opus 4.6 disclosed it in 55.9 per cent of answers, Gemini in 28.7 per cent, and GPT-4.1 in 5.2 per cent.
Disclosure also depended on framing. For Claude Opus 4.6, a customer-service persona was associated with 80 per cent disclosure, against 40 per cent with no persona. The models' stated confidence did not fall when the advertisement was present. An assistant can sound equally sure whether or not it is being steered.
What the study does not show
The limits are as important as the findings, and the authors name them. The advertisement sat in the system prompt, through the API, one advertisement per condition, in a fixed template. The study also notes that the advertisement did not instruct the model to prefer the drug. Other routes, such as training data, retrieval, or the user's own context, were not tested. The study measures the model's choice between equivalent options. It does not measure patient outcomes or any change in what doctors prescribe.
The study has also been reported as published in a peer-reviewed journal, The Lancet Digital Health, on 11 November 2026, according to a Mount Sinai press release. That version was not available to us, and this piece relies on the preprint's figures.
Implications
These points are our analysis. They are not findings of the study.
The first is about monitoring. A check that scores answers against an acceptable-answer key will pass a sponsor-steered answer, because both options are acceptable. A better check compares which option the model chooses across equivalent cases, with and without a sponsor, for each model separately.
The second is about disclosure and confidence. If a model rarely mentions a sponsor, and sounds equally sure either way, neither signal can tell a user or an auditor that a choice was steered. A disclosure rate is a poor safety measure, and so is a confidence score.
The third is that labelling is not a safeguard by itself. The advertisements were labelled and the effect remained. The study covers only one delivery route, so an absence of effect through one channel is not evidence of safety through another.
Conclusion: the label is not the safeguard
A line of text in a system prompt is cheap to write and easy to hide. This study shows that, in a controlled setup, such a line can move a model's choice between two sound options by a large margin, while the answers remain accurate and largely unlabelled. The honest next question is not whether a sponsor can buy an assistant's preference. It is whether anyone is measuring the preference at all.
Sources: Omar M, Agbareia R, McGreevy J, et al., "Ad-verse Effects: Pharmaceutical Advertising Shifts Drug Recommendations by Consumer-Facing AI," medRxiv, 16 April 2026 (doi:10.64898/2026.04.14.26350868); Mount Sinai press release, 11 November 2026, on the Lancet Digital Health publication (reported, not read in full).