Healthcare AI Safety Gap · Don't break patients
    Peer-reviewed / official11 Nov 2026·US

    A drug ad in the prompt nudged AI models toward recommending that drug

    The Lancet Digital Health: across 12 consumer AI models and 33 clinical scenarios, adding an ad for one treatment raised the share of answers recommending it from 35% to 47.6%, and the answers mentioned the ad only about 30% of the time.

    47.6% vs 35%
    of AI answers choosing a treatment with versus without an ad for it in the prompt

    What the research found

    Mount Sinai researchers gave 12 consumer-facing AI models 33 clinical scenarios, each with several treatment options, and in some runs added an advertisement for one option that described the drug without telling the model to recommend it. With the ad present, models chose the advertised treatment in 47.6% of responses against 35% without it, and disclosed the ad in about 30% of answers; the size of the shift varied by model. In most cases the models still chose one of the appropriate options, and the study did not test effects on real prescribing.

    Why it matters for providers

    As AI assistants start carrying ads, providers and patients need to know when a recommendation has been shaped by promotion rather than by the case in front of it.

    Original source
    The Lancet Digital Health (Icahn School of Medicine at Mount Sinai)
    Read the full report ↗
    Peer-reviewed or official source, the most reliable tier.