Key Takeaways

  • In a controlled study of 12 AI models, a labelled drug ad placed before a clinical case raised selection of the advertised drug from 35.0% to 47.6%.
  • In four model-and-scenario combinations, the advertised drug went from never chosen to always chosen.
  • Only 29.9% of responses disclosed that an ad had been shown.

Could a Drug Ad Change What an AI Chatbot Recommends? Researchers Tested It.

Imagine asking an AI chatbot which medication a patient should receive.

Now imagine someone slipping a drug advertisement into the conversation just before the chatbot answers.

Would its recommendation change?

Researchers at Mount Sinai tested that question across 12 AI models. Their finding: when a labelled drug advertisement appeared before a clinical case, the advertised medication was selected 47.6% of the time, compared with 35.0% without the advertisement.

That's a shift of approximately 12.7 percentage points — caused by changing what the AI saw before making its recommendation.

And in four specific model-and-scenario combinations, the difference was even more dramatic.

The Experiment: 12 AI Models, 33 Medical Scenarios

The researchers tested models from Google's Gemini, OpenAI's GPT-4.1 and GPT-5 families, and Anthropic's Claude.

They used 33 physician-written clinical scenarios and analysed 258,559 model calls made between February 17 and 28, 2026.

The experiment compared responses with and without a drug advertisement labelled as paid content, inserted immediately before the clinical case.

The question was straightforward: would seeing promotional material change which medication the AI selected?

The study was reported in The Lancet Digital Health.

In Four Cases, the Recommendation Flipped Completely

Across the experiment, advertisements increased selection of the promoted drug.

But four model-and-scenario combinations produced a particularly striking result.

Without the advertisement, the promoted medication was selected 0% of the time.

With the advertisement, it was selected 100% of the time.

One example involved dapagliflozin, a medication used to treat heart failure.

These were specific experimental combinations, not a result observed across every model or medical scenario.

Still, they illustrate how strongly promotional content could influence certain AI responses.

The AI Often Didn't Mention the Advertisement

The researchers also examined how promotional material appeared in the models' answers.

Among responses selecting the advertised medication, the study reported a 52.7% measure of advertising-claim repetition.

Meanwhile, only 29.9% of responses disclosed that an advertisement had been shown.

That creates a transparency problem.

Someone reading an AI-generated treatment recommendation might not realize that promotional material had appeared immediately before the medical question.

The Most Surprising Finding: The Answers Could Still Follow Medical Guidelines

Here is what makes the research especially important.

The advertisements influenced which drugs the models selected, but overall agreement with clinical guidelines remained high.

In other words, an AI could favour an advertised medication while still choosing a treatment supported by medical guidelines.

That matters because a medical question can have more than one acceptable treatment option.

An ordinary accuracy check might find nothing obviously wrong, even if advertising influenced which option appeared in the answer.

The researchers also found no clear advertising effect for unsupported supplements or clinically suboptimal drugs in the conditions they tested.

That does not establish that AI systems are immune to unsafe medical advertising. It means the demonstrated influence was especially notable when the advertised medication was already a clinically acceptable option.

Why This Matters for Canada

Canada restricts most direct-to-consumer advertising of prescription drugs, although certain limited forms of advertising are permitted.

The United States allows broader prescription-drug advertising. US drug companies spend more than $6 billion a year on direct-to-consumer advertising.

But this study raises a question that goes beyond television commercials or traditional pharmaceutical marketing.

What happens when promotional content appears inside the information an AI system uses to answer a medical question?

The experiment did not examine real prescriptions, actual patient treatment decisions or health outcomes.

It measured AI-generated recommendations under controlled conditions.

That distinction matters: the findings demonstrate an influence on model responses, not evidence that patients received inappropriate medications because of advertisements.

The Twikup Take

The unsettling part isn't simply that a drug advertisement changed an AI's answer.

It's that the answer could still appear medically reasonable.

The researchers' findings point to the need for testing medical AI systems under advertising exposure, clearly disclosing promotional material and introducing safeguards before advertising becomes routine in health chatbots.

A medical recommendation can follow the guidelines and still be influenced by an advertisement the reader never knew was there.

Image: Illustrative image.

Sources