药物广告可影响人工智能治疗推荐：《柳叶刀数字健康》11月11日研究测12个模型与33个临床情境，有广告时被宣传药物入选比例47.6%
纽约报道 — 西奈山伊坎医学院研究人员的一项研究11月11日在线刊于《柳叶刀数字健康》（The Lancet Digital Health），发现清晰标注的药物广告可以改变人工智能模型给出的治疗推荐，使其更倾向于选择被宣传的药物。本稿只报道研究事实，不作个体诊疗建议。
研究测试了包括谷歌、OpenAI 与 Anthropic 等主要科技公司的12个面向消费者的人工智能模型。研究人员把医生撰写的临床病例交予模型，病例之前或放置一则真实药物广告、或不放置，患者信息与治疗选项保持不变，只改变广告是否存在。在33个临床情境中，有广告时被宣传药物被选中的比例为47.6%，无广告时为35%。在4个情境中，模型在无广告时从未选中被宣传药物（0%），加入广告后则每次都选中（100%）。研究共分析258559条人工智能回复，并评估所选治疗是否符合临床指引。
研究人员同时发现，在一部分自由文本回复中，模型仅在29.9%的回复中明确提及广告的存在，却更频繁地重复广告中的信息或说法。研究第一作者、伊坎医学院讲师马哈茂德·奥马尔（Mahmud Omar）说，模型给出的答案往往看起来合理，其在几个合理选项之间的偏好在看过广告后发生了变化，这种影响容易被忽略。研究资深作者吉里什·纳德卡尼（Girish N. Nadkarni）说，答案即便临床上合理，也需要追问人工智能为何得出该答案，并区分临床指引与商业影响。
研究人员同时指出，研究结果并不意味着药物广告必然导致人工智能推荐不安全或不恰当的治疗：在本研究中，广告并未明确增加对无依据补充剂或明显次优药物的推荐。研究也未考察人工智能推荐的变化最终是否影响处方、患者决定、费用或健康结局。
Drug Advertisements Can Sway AI Treatment Recommendations: Lancet Digital Health Study on November 11 Tests 12 Models Across 33 Scenarios, With the Advertised Drug Chosen in 47.6 Percent of Responses
NEW YORK — A study by researchers at the Icahn School of Medicine at Mount Sinai, reported in the November 11 online issue of The Lancet Digital Health, found that a clearly labeled pharmaceutical advertisement could shift which treatment an artificial intelligence model recommended, making it more likely to select the advertised drug. This item reports the study's facts only and offers no individual medical advice.
Researchers tested 12 consumer-facing AI models from major technology companies, including Google, OpenAI and Anthropic. They gave the models physician-written clinical cases, either alone or preceded by a clearly labeled real drug advertisement, with the patient information and treatment choices unchanged. Across 33 clinical scenarios, the advertised drug was selected in 47.6 percent of responses when an ad was present, compared with 35 percent without an ad. In four cases, models shifted from selecting the advertised drug zero percent of the time to selecting it 100 percent of the time when the ad was introduced. The researchers analyzed 258,559 AI responses and assessed both the treatment selected and whether it was consistent with clinical guidance.
In a subset of free-text responses, models explicitly mentioned the ad in only 29.9 percent of responses, while they repeated information or claims contained in the advertisement more frequently. First author Mahmud Omar said the answers often appeared reasonable, but the model's preference among several reasonable options changed after seeing the ad, an influence that could be easy to miss. Senior author Girish N. Nadkarni said an answer can be clinically reasonable and still raise the question of why the AI arrived at it, and that clinical guidance must be distinguished from commercial influence. The investigators noted that the advertisements did not clearly increase recommendations for unsupported supplements or clearly suboptimal drugs, and that the study did not examine whether changes in AI recommendations ultimately affected prescriptions, patient decisions, costs or health outcomes.
来源：《柳叶刀数字健康》
发布：GSM·美国（发布链接）
图片来源：图：医院大楼外观（资料图）。经Wikimedia Commons。 / Photo: a hospital building exterior (file photo). Via Wikimedia Commons.
GSM Global · 本篇英文版 English edition of this story