“What does a raised TSH level mean?”, “Is this rash dangerous?” – questions like these are now often put to an AI chatbot first. According to OpenAI, more than 300 million people use ChatGPT alone for health questions every week. But how good are the answers, and how should they be used properly?
Exam knowledge is not the same as help in everyday life
Large language models now often pass medical exam questions with very good results. However, a study by the University of Oxford, published in Nature Medicine in February 2026, reveals an important gap: around 1,300 participants were asked to work out, in case scenarios developed by doctors, which condition was likely and whether they should, for example, see their GP or go to the emergency department. Those who used a chatbot for this did not make better decisions than the comparison group, who searched the internet or relied on their own judgement.
The reasons lie in the interaction: people often provide incomplete information, ask questions differently from those in test datasets and do not always pick out the key points from longer answers. The study makes it clear that medical AI must be tested under realistic conditions with real people.
Health features and data protection
Providers are specifically expanding their chatbots for health topics. In 2026, OpenAI introduced “ChatGPT Health” in the USA, which can incorporate test results and data from fitness apps on request. In Germany and the EU, the feature is not available at the time of writing – partly because of the stricter rules on health data. Anyone who copies test results or doctors’ letters into a chatbot should be aware that health data is among the most sensitive personal data. It is worth removing names, dates of birth and other identifying details and checking the service’s privacy settings.
What chatbots can be useful for
- Understanding medical terms: having a doctor’s letter or lab report translated into plain language.
- Preparing for appointments: collecting questions, describing symptoms in a structured way, keeping a symptom diary.
- General knowledge: how does a medicine work in principle? What happens during a particular examination?
Where caution is needed
- Diagnoses and treatment decisions: a chatbot cannot examine you, knows your medical history only as far as it has been entered, and can make convincing-sounding errors (“hallucinations”).
- Doses and interactions: pharmacies and doctors’ practices are the right people to ask here.
- Emergencies: with chest pain, shortness of breath, signs of paralysis, speech problems, heavy bleeding or impaired consciousness, do not waste time with a chatbot: call the emergency number 112.
- Mental health crises: chatbots are no substitute for help in an acute crisis. In Germany, the TelefonSeelsorge helpline can be reached around the clock on 0800 111 0 111 and 0800 111 0 222.
Five tips for using chatbots
- Treat the answer as a starting point, not as a result.
- Ask for sources and check them yourself, for example at gesundheitsinformation.de or gesund.bund.de.
- Explicitly ask the chatbot which warning signs require a medical assessment.
- Do not stop or start any medicines without consulting a professional.
- Take the result with you to your doctor’s appointment and discuss it openly.
Conclusion
AI chatbots are a powerful new information tool – with the same limitations as any health information from the internet, plus a few more. Used correctly, they can help you go into a consultation better informed. They do not make a diagnosis.
More on this topic: AI in Medicine 2026 – an overview.
Note: This article is for general information only and does not replace advice from a doctor or pharmacist. As of: October 2026.
Video on this article
Watch on YouTube · Subscribe to the channel
Sources
- University of Oxford: New study warns of risks in AI chatbots giving medical advice (February 2026)
- Bean AM et al.: Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine 2026; 32: 609–615
- Notebookcheck: ChatGPT liest jetzt Gesundheitsdaten – nur nicht in Deutschland (2026)
- IQWiG: gesundheitsinformation.de
Leave a Reply