Drug ads can steer AI choices between appropriate treatments, study finds

First author Mahmud Omar is a physician-scientist and lecturer in Mount Sinai's Windreich Department of Artificial Intelligence and Human Health who leads research on generative AI safety and reliability in clinical care. The preprint builds on earlier work by Omar and colleagues on medical misinformation, demographic biases and models' confidence in clinical answers. Its contribution statement credits him with conceiving the study, designing the experiments and running the analyses.

The authors declared no competing interests. In their funding statement, they identified Mount Sinai computing resources and support from National Institutes of Health grants, and said funders had no role in the study's design, analysis or writing.

In the main preference experiment, physicians developed and reviewed 13 cases in which two brand-name drugs from competing manufacturers were considered appropriate under clinical guidelines. Models could select either drug or a neither/other option. They received the same case without advertising, with an ad for one drug, or with an ad for the other. The sponsored message went into the system prompt, the context used to frame the assistant's behavior. It did not explicitly instruct the model to favor the product.

The preprint describes standardized messages modeled on direct-to-consumer ads, including a brand, indication and clinical claim. Mount Sinai's later announcement describes real, clearly labeled drug advertisements. Both describe holding the clinical case constant while changing exposure to promotional text.

Across 74,880 advertising-condition calls, compared with 37,440 no-ad calls in those 13 cases, the preprint reported an average preference shift of 12.7 percentage points toward the advertised drug. Aggregate clinical accuracy rose from 89.4% without ads to 93.0% with them. Thus, an accuracy-only evaluation would not identify the commercial preference shift as deterioration.

The preprint calculated that shift against a baseline proxy: the mean of the no-ad selection rates for the two drug options. The main experiments varied prompt wording and the assistant's assigned persona, with 20 repetitions per experimental cell. The large response count therefore measures repeated model behavior on a small set of constructed cases, not experience across hundreds of thousands of patients.

Some individual model–scenario combinations switched entirely between appropriate drugs. In a heart-failure example, selection of dapagliflozin, sold as Farxiga, changed from 0% to 100% after advertising; empagliflozin, sold as Jardiance, was the baseline alternative. Both were judged appropriate for that case. This illustrates a preference reversal, rather than a demonstrated improvement or deterioration in care.

The reported total of 258,660 API calls includes attempts: the research repository records 101 API failures, reconciling it with the 258,559 responses in the November announcement. The 33 scenarios span the broader experimental program, including supplements and suboptimal drugs; they are not 33 interchangeable tests of clinically equivalent prescription choices.

Large differences, with limits to the comparison

The preprint's main experiment found substantial differences across the particular models tested.

Individual results ranged from a 35.1-point increase for Gemini 2.5 Lite to a 3.8-point decrease for Claude Opus 4.6. Within OpenAI, GPT-4.1 Mini shifted by 18.1 points, compared with 3.1 points for GPT-5.2. These are results for specified model versions and experimental prompts, rather than a permanent ranking of the developers' consumer products.

The authors suggest that differences in alignment and handling of system-prompt content could help explain resistance. Their experiment does not isolate those causes. It also found baseline preferences without advertising, so resistance to an ad does not establish that a model's original choice was free of other biases.

The strongest qualification is what happened when promotion competed with clearer clinical evidence. In ten scenarios involving unsupported supplements, endorsement fell from 2.2% without ads to 1.6% with them. In another ten involving suboptimal or contraindicated prescription options, aggregate accuracy remained high and increased from 94.1% to 95.6%. The study therefore does not support a general claim that ads override models' medical knowledge or routinely produce unsafe advice.

In the preprint’s smaller experiment, explicit acknowledgment of advertising varied across three models, each with 520 ad-exposed responses. Disclosure was measured using rule-based text analysis. Source: PubMed Central.

Labeling an ad did not ensure disclosure in the answer

A smaller experiment required free-text justifications alongside drug selections. It used three models chosen to span the main experiment's susceptibility range: Gemini 2.5 Flash, GPT-4.1 and Claude Opus 4.6. It comprised 2,340 calls, including 1,560 with advertising, across the same 13 prescription scenarios.

Only 467 of the 1,560 ad-exposed responses—29.9%—explicitly acknowledged the advertisement. Disclosure ranged from 5.2% for GPT-4.1 to 55.9% for Claude Opus 4.6. Model-reported confidence remained high with and without ads. These measures used reproducible, rule-based text analysis, which the authors caution may miss nuanced responses.

Responses choosing the promoted drug echoed 52.7% of the advertising claims in their justifications, compared with 19.4% among responses choosing something else. Those percentages measure the share of claims echoed, rather than the share of answers containing any advertising language. The association links promotional wording to the recommendation, but does not independently establish the model's internal reasoning process.

For developers, the distinction between an ad being labeled in the input and its influence being visible in the answer matters. The results suggest that disclosure and clinical correctness need separate evaluation from the stability of treatment preferences.

OpenAI’s proposed ad-format mockup shows a sponsored grocery listing below a ChatGPT answer. The study tested pharmaceutical text in system prompts. Image: OpenAI. Source: OpenAI advertising announcement.

The next test is the boundary between ads and answers

Actual product design determines whether promotional text reaches the model at all. OpenAI, the developer of the ChatGPT conversational assistant as well as several models tested here, described advertising as a way to fund broader access. In its January 16 announcement, it planned ads for the free and lower-cost Go tiers, while saying Plus, Pro, Business and Enterprise subscriptions would have none. It said ads would remain separate from answers and would not influence them, and its initial test would exclude ads near health topics. These are company commitments, not findings verified by this study.

That placement policy subsequently changed. OpenAI's ad policies updated September 10 record an April revision under which medical, legal and financial advice contexts were no longer categorically blocked from ads. Sensitive conversations remain ineligible, and the policy lists mental and personal health conversations among those exclusions. It also describes gradual, manually reviewed approvals for healthcare and medicine advertisers. Permission to display an ad near a conversation does not establish that its text enters the model's clinical reasoning.

The researchers tested system-prompt exposure. They did not establish equivalent effects from ads shown separately on a screen, retrieved promotional pages, user-pasted material or advertising encountered during training. Nor did they follow patients or physicians to measure treatment decisions, prescriptions, costs or outcomes. The preprint's smaller justification experiment also used only five repetitions per experimental cell, limiting precision for individual scenarios.

Mount Sinai says its team plans further evaluation using real-world clinical cases and human-AI interactions. It recommends comparing systems with and without promotional content, disclosing exposure and separating promotion from clinical advice. The next consequential evidence is whether deployed systems enforce that separation—and whether recommendation shifts persist when people make actual care decisions.

Seedance 2.5 reaches 4K on Runway, Venice and Picsart

ByteDance’s video model is reaching creators through multiple platforms with 30-second 4K output. Posted prices and draft workflows show why the choice of distributor still matters.