An ad can change the drug an AI chooses
Can an ad change the treatment an AI recommends to you? In an experiment published in The Lancet Digital Health, yes. With the ad in front of it, out of every 100 answers, about 13 more chose the advertised drug.
An ad can change the drug an AI chooses
Experiment with 12 AI models, no real patients. Researchers at Mount Sinai tested models from Google, OpenAI and Anthropic with clinical cases written by doctors. In some queries, right before the case, they placed an ad for a real drug, labelled as an ad. Everything else was identical.
Across the 13 prescribing scenarios, the AI chose the advertised drug in 47.6% of answers, compared with 35.0% without the ad. The effect varied a lot by case, from a drop of 7 points in major depression to a rise of 40.5 points in type 2 diabetes with heart failure. In free-text answers, which were only tested with three of the models, it mentioned the ad in 3 out of 10.
In those scenarios, every option was valid according to the guidelines. In other tests, with an unsupported supplement or a worse drug as the ad, no clear change was seen, although that does not prove it cannot happen. The ad mostly shifted the choice between correct options, which is why it is hard to spot.
What is still unresolved. They tested a single ad inside the prompt, in a single query and with models from February 2026. The authors warn that it cannot be extrapolated to the apps you use, and the study does not measure real prescriptions.
What does this mean for you? If an AI gives you a specific brand, check the alternatives against the clinical guideline and the indications against the summary of product characteristics, not against the AI itself. And be careful with what you paste into it. The ad in the study was inside the prompt, just like a leaflet you copy and paste.
Read the Mount Sinai press release and the study in The Lancet Digital Health.
This week's radar
The Valencian Community tenders an AI to detect skin cancer from primary care. This is a public tender, not a study. The regional health department wants an AI that analyses photos of skin lesions taken at the health centre, estimates their risk and speeds up referral to Dermatology. It would reach 244 health centres and 567 local clinics, about 5.4 million people. Base budget, 1,452,000 euros. Bids are open until 9 December. The regional government's press release (in Spanish).
5.4 millionpeople, approximately, that it would reach
Apple studies how we use health chatbots. This is the announcement of a study that is just starting, with no results yet. Apple and OpenEvidence are looking for up to 200,000 adults in the US who will ask health questions to an AI app and can share data from their iPhone. They will measure trust, the quality of the answers and what people do afterwards. Only for people aged 21 or over in the US. Apple's press release.
200,000adults at most in the study, US only
ChatGPT changes its voice mode on 11 December. OpenAI is retiring the Advanced and Standard voices. If you have one selected, you will automatically move to Live voice, which listens and speaks at the same time, like a conversation. Live video and screen sharing will stay on Advanced voice for a while. Your past conversations are not lost. OpenAI's help article.
11 DecemberOpenAI retires the Advanced and Standard voices
This week's reflection
Cervantes tells us that Alonso Quijano slept so little and read so much that his brain dried up. An AI does not lack sleep, but it reads whatever we put in front of it. We know advertising works on people. What is unsettling is that it also works on machines that learned to write by reading us. An assistant that reads everything is more useful and, at the same time, easier to nudge, and it does so with such good manners that you do not notice. Perhaps the question is not whether we can trust AI, but whether we know what it has read before answering us.
IAal día
Your weekly dose of artificial intelligence applied to healthcare. News, tools and real cases directly in your inbox.
With every edition, a free resource as a gift: a guide, a report or a tool we have picked for you (in Spanish).
Resources from recent editions
- Ten consensus recommendations for bringing AI into hospitals
- Tutorial de Anthropic para crear una skill hablando con Claude
- Estudio completo de AMIE
Subscribe for free
Every Wednesday in your inbox, in Spanish. No spam.
Related articles
25IA al día · No. 2528 oct 20265 min readIn Spanish
Una IA acierta el 83 % de los casos clínicos difíciles. Esta es la letra pequeña
Lo esencialEstudio en Science: IA de Microsoft ante 304 casos clínicos difíciles.
Regalo: Tutorial de Anthropic para crear una skill hablando con Claude
24IA al día · No. 2414 oct 20264 min readIn Spanish
Si usas Claude con pacientes, respondes tú y se lo tienes que decir
Lo esencialNuevas normas de uso de Anthropic para Claude en recomendaciones clínicas.
Regalo: Estudio completo de AMIE
23IA al día · No. 237 oct 20264 min readIn Spanish
Tu GPT o tu Gem tiene fecha de caducidad, y esto es lo que cambia
Lo esencialAnuncios de Google y OpenAI sobre el fin de Gems y GPTs.
Regalo: Grabación del curso «Crea tu primer agente con Microsoft Copilot»