AI models recommended an advertised drug in 47.6% of responses, compared with 35% when no advertisement appeared.
Advertising could shift recommendations among medically appropriate treatments, making its influence difficult to spot.
Researchers called for greater transparency but did not establish whether the changes affected prescriptions or patient health.
For years, the broadcast airwaves have been filled with ads for prescription drugs. Now, a new study has found that drug advertising is influencing the medical advice that AI platforms dispense.
The findings, gathered by the Icahn School of Medicine at Mount Sinai and reported in The Lancet Digital Health, raise questions about how consumers can distinguish medical guidance from commercial influence.
Researchers tested 12 AI models from Google, OpenAI and Anthropic in four controlled experiments. The tests examined how advertising affected drug selection, including situations in which multiple treatments were appropriate and others involving products with weaker medical justification.
The study
Across 33 clinical scenarios, researchers analyzed 258,559 responses. Models received the same patient information and treatment choices, with a clearly labeled drug advertisement placed before some cases.
The advertised medicine was selected in 47.6% of responses when advertising appeared, versus 35% without it. In four model-and-scenario combinations, selection jumped from zero to 100%.
The results suggest advertising can have particular influence when medical evidence leaves room for more than one reasonable choice. In the researchers’ earlier publicly available report, Google models showed the largest average shift toward advertised drugs, followed by OpenAI models. Anthropic models showed comparatively little change. Those comparisons describe the models and experimental conditions tested.
A challenge for AI medical advice
That creates a challenge for evaluating AI health advice: an answer can pass a check for medical appropriateness while still reflecting advertising.
First author Dr. Mahmud Omar said the recommendations often appeared reasonable. In a subset of written responses, models explicitly acknowledged the advertisement only 29.9% of the time.
The experiments also identified limits to advertising’s influence. Models generally resisted promotion when the advertised product lacked supporting evidence or represented a clinically inferior choice. The research therefore points to a vulnerability in how AI chooses among acceptable options, rather than a wholesale abandonment of medical knowledge.
Important questions remain unanswered. Researchers did not determine whether altered recommendations changed actual prescriptions, consumer spending or health outcomes.
They urged AI developers and health systems to test responses with and without advertising, disclose promotional content and keep it separate from clinical advice.