Drug Ads Shift AI Treatment Picks by 12.6 Points, Lancet Study Finds

A pharmaceutical ad placed in the context of a clinical question changed what large language models recommended: 12 consumer-facing LLMs picked the advertised treatment 47.6% of the time, versus 35% when no ad was present. That 12.6-percentage-point lift, published this week in The Lancet Digital Health, is the core finding of one of the first systematic tests of advertising's pull on AI-generated medical advice.

The researchers ran a dozen models, including versions of Claude (Anthropic), Gemini (Google), and GPT (OpenAI), through 33 scenarios. Each scenario paired a clinical case description with multiple treatment options. In a subset of cases, the models also received an ad for one of those options — drug information and approved uses, with no explicit instruction to recommend it.

The ad still worked. That's the uncomfortable part. The shifts varied across the 12 models, but the direction was consistent: exposure to promotional content moved recommendations toward the promoted drug.

Did the ads make the AI wrong?

Mostly, no — and that's precisely why the finding matters. In most cases the models still selected one of the clinically appropriate options, even after the ad. The problem isn't obviously bad advice. It's unexplained advice.

The disclosure record makes it worse. The models mentioned the ad's presence in their output only about 30% of the time. A user reading the response has, in roughly seven cases out of ten, no signal that promotional content sat in the context window.

"An answer can be clinically reasonable and still raise an important question about why the AI arrived at that answer," study author Dr. Girish Nadkarni, professor of medicine and system chair of the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai's Icahn School of Medicine, said in a statement. "As AI becomes part of how people get health information, we need to be able to distinguish clinical guidance from commercial influence and be transparent about when advertising is part of the interaction."

How would an ad even reach a medical chatbot?

The study lays out four routes: user prompts, application-added context, retrieved online material, or inadvertent or malicious insertion. Think of a patient pasting a webpage that carries a drug banner, or a retrieval step pulling ad-laced content into the model's context.

The regulatory gap is structural. Television and print drug advertising fall under U.S. Food and Drug Administration oversight. The authors note that "AI-mediated promotion might not fit existing categories" — a framework built for 30-second spots doesn't map onto tokens in a context window.

The money and the law behind the exposure

The advertising substrate is enormous. Pharmaceutical companies spent more than $6 billion on TV advertising alone in 2024, according to an IQVIA report, which stated that direct-to-consumer advertising "directly influences patient requests and prescription rates, driving real action in healthcare." In the U.S., prescription drugs fill more than 30% of commercial time during major networks' evening news, per a 2025 Senate Health, Education, Labor and Pensions Committee release.

Nearly every rich country has already opted out. All high-income countries except the U.S. and New Zealand ban direct-to-consumer prescription drug ads. Last year, Senator Bernie Sanders, ranking member of the HELP committee, introduced the End Prescription Drug Ads Now Act — co-signed by six other senators — which would ban such advertising, social media included.

Why clinicians might not catch it

Physicians are getting more comfortable with AI, not less. In a recent Doximity survey of 3,379 U.S. physicians, 54% reported moderate or high trust in AI's ability to recommend treatment or management, and three-quarters said patients have grown more likely to propose an AI-suggested diagnosis or treatment plan in the past six months.

The study's authors flag the compounding risk: "Clinicians might defer to algorithmic recommendations even when they conflict with clinical judgement," and an appropriate but ad-influenced suggestion "might provide little reason to question it."

One limit is worth stating plainly. The study did not examine whether its measured effects translate into real-world changes in patient requests or prescriptions. What it establishes is the mechanism: promotional text in, shifted recommendations out, disclosure missing two-thirds of the time. Whether regulators will extend FDA-style oversight to AI-mediated promotion is now the open question the study leaves on the table.