The old joke was that people should stop diagnosing themselves on Google.
That advice is becoming harder to apply. Health searches are no longer limited to a list of blue links, medical forums and hospital websites. Millions of people now ask conversational AI tools about symptoms, medication, diet, mental well-being, medical paperwork and how to navigate healthcare systems.
A 2026 paper published in Nature Health gives one of the clearest large-scale pictures yet of how this is happening in everyday life. The study, by Beatriz Costa-Gomes, Pavel Tolmachev, Eloise Taysom, Viknesh Sounderajah and colleagues at Microsoft AI, analysed health-related conversations with Microsoft Copilot from January 2026.
The dataset was substantial. After filtering and classification, the authors analysed 617,827 health conversations. The work used a privacy-preserving pipeline: personal information was automatically removed, summaries were generated by machine, and no human researcher accessed the raw conversation content. The team then used a hierarchical intent taxonomy with 12 primary categories, validated against expert human annotation, to understand what people were asking and when.
The largest category was general health information and education, accounting for 40.8% of conversations. These included questions about conditions, treatments, medication, nutrition and health concepts.
But the more revealing finding is that many interactions were more personal than a simple information search.
Nearly one in five conversations involved personal symptom assessment or condition discussion. One in seven of these personal health queries concerned someone other than the user, showing that AI is also becoming a caregiving tool. A parent, partner or adult child may be using a chatbot to understand what a symptom might mean, how serious something sounds, or what steps to consider before speaking to a professional.
The timing also mattered. Personal health queries rose in the evening and at night, when traditional healthcare access is often most limited. Mobile use leaned more towards personal concerns, while desktop use was more common for professional, academic and paperwork-related health tasks.
That pattern will feel familiar in South Africa. A person with private medical aid may still hesitate before calling a doctor after hours. Someone without easy access to care may use AI because it is immediate, free at the point of use and available on a phone. Even higher-income households face practical friction: specialist waiting times, unclear benefits, co-payments, emergency-room uncertainty and the simple anxiety of not knowing whether a symptom is serious.
The study does not claim that AI should replace doctors. In fact, its importance lies in showing why that risk must be managed carefully. People are already using general-purpose chatbots in health contexts, whether healthcare systems are ready for it or not.
This creates a design and safety challenge. A chatbot answering a general question about nutrition is not operating in the same risk category as one responding to chest pain, medication interactions, pregnancy concerns or mental health distress. The system needs to understand when to provide general information, when to encourage professional care, and when to avoid sounding more certain than the evidence allows.
The paper’s authors also note limitations. The analysis only covered Microsoft Copilot consumer logs, and the findings may not generalise perfectly to other platforms or clinical settings. The study characterised what people asked, not whether the answers they received were medically correct or helpful.
That last point is crucial.
The next phase of health AI research cannot only be about usage volume. It must examine answer quality, safety, escalation, misinformation risk and real-world outcomes. A tool that is widely used is not automatically safe. A response that sounds calm and authoritative is not automatically correct.
For consumers, the practical lesson is clear. AI can be useful for understanding terminology, preparing questions for a doctor, organising symptoms, learning about general health topics or navigating paperwork. It should not be treated as a diagnosis, a prescription, or a substitute for urgent care when symptoms are serious.
For healthcare providers, insurers and digital platforms, the message is more strategic. People want health guidance that is immediate, understandable and available outside office hours. If trusted institutions do not provide better digital pathways, consumers will keep filling the gap with general AI tools.
The shift from “Dr Google” to “Dr Chatbot” is already under way. The challenge now is not to pretend it can be stopped, but to make sure the tools people use in vulnerable moments are built with the seriousness health deserves.
Source Information
Study Title: Public use of a generalist LLM chatbot for health queries
Authors: Beatriz Costa-Gomes, Pavel Tolmachev, Eloise Taysom, Viknesh Sounderajah, Hannah Richardson, Philipp Schoenegger, Xiaoxuan Liu, Matthew M. Nour and colleagues
Journal: Nature Health
Year: 2026
