Picture a patient case typed into an AI model, with a clearly labeled drug ad sitting just above it. Researchers at Mount Sinai ran that setup across 12 models from companies including Google, OpenAI and Anthropic, and the ad changed what the models recommended. The advertised drug was picked in 47.6% of responses when the ad was there and in 35% when it was not, according to Mount Sinai’s announcement of the study, published on 11 November in The Lancet Digital Health.
What the researchers did
The team gave the models 33 clinical cases written by physicians. Each case was run both on its own and with a real drug advertisement placed right before it. The patient details and the treatment choices stayed the same, so the ad was the only thing that changed.
In total they analyzed 258,559 responses. The study’s public code and data show 258,660 attempted calls, 101 of which failed, and name Ozempic, Mounjaro, Wegovy and Zepbound among the brands whose ad text was used.
Some swings were total. In four cases, models went from never choosing the advertised drug to choosing it every time once the ad appeared.
Why the answers still looked fine
Most of the shift happened between treatments that were all clinically reasonable. The ads did not clearly push the models toward unsupported supplements or clearly worse drugs. That is what makes the effect hard to spot.
“What I find quite interesting, and even surprising, was how reasonable the answers often appeared,” first author Mahmud Omar said in the Mount Sinai release. Most answers also kept quiet about the ad. In a sample of free-text answers, they mentioned the ad only 29.9% of the time, while repeating its claims more often than that.
An earlier version of the work, a medRxiv preprint from April that had not been peer reviewed, split the results by company. When both drug options fit clinical guidelines, Google’s models moved most, by 29.8 percentage points, then OpenAI’s by 10.9 and Anthropic’s by 2.0. We could not check whether the final journal version reports the same split.
Who this affects
Ads in chatbots are already here. OpenAI started testing ads in ChatGPT in the US on 9 February 2026, for the Free and Go plans. It says the ads run on systems separate from the chat model and do not influence answers. Its ad policies allow some healthcare and medicine advertisers, approved case by case, and keep ads away from personal health conversations.
That is a different setup from the study, where the ad sat inside the text the model read. So the study does not show that ChatGPT’s ads change its answers. It shows what can happen when promotional text ends up in what a model reads.
The limits matter too. The team did not look at whether any of this changed prescriptions, patient decisions, costs or health outcomes, as Medical Economics also noted. Their suggested fixes are simple: test AI tools with and without promotional content, say when an ad is shown to a model, and keep promotion apart from clinical advice.
Who runs a test shapes it, too. Last week we covered a study where clinicians rated a grounded medical AI above ChatGPT, a study the AI’s maker helped design. For the basics, see our guide to how health AI assistants work and where your data goes.
What to do this week
If you ask a chatbot about treatment, ask it to list every reasonable option and why, not just one pick, and take the list to your doctor or pharmacist. If you build a health AI tool, rerun your test cases with an ad pasted into the input and compare the picks. The researchers’ code is public, so you can borrow their setup.
Sources
- Mount Sinai: Drug ads can influence AI treatment recommendations, study finds
- medRxiv preprint: Ad-verse Effects, pharmaceutical advertising shifts drug recommendations by consumer-facing AI
- GitHub: code and data for the study
- Medical Economics: AI chatbot treatment recommendations are swayed by drug ads
- OpenAI Help Center: Ads in ChatGPT
- OpenAI: Ad policies
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
