Oxford Study Finds AI Chatbots Give Inconsistent and Inaccurate Medical Advice
TL;DR: A University of Oxford study involving 1,300 people found that AI chatbots give inconsistent medical advice that varies depending on how questions are worded. Participants often did not know what to ask, and the chatbots listed possible conditions without helping users determine which applied to them.
Researchers gave participants scenarios including a severe headache and new-mother exhaustion, then had them seek advice from AI chatbots. The results showed a consistent problem: people share medical information gradually, leave things out, and phrase questions differently each time — and the AI responses shift accordingly.
The Problem With Asking
Dr Adam Mahdi, one of the researchers, said people “struggle to get useful advice” from AI chatbots because the models respond to what they are told rather than probing for what has been left out. A doctor in a consulting room will ask follow-up questions, notice what a patient avoids mentioning, and weigh symptoms against experience. Chatbots list possible conditions, but users cannot easily distinguish which one fits their situation.
Dr Rebecca Payne described the findings as “dangerous,” particularly given how many people now turn to AI for health guidance. Mental Health UK polling from November 2025 found that one in three UK residents have used AI for mental health support — a figure that has almost certainly risen since.
Baked-In Bias
Dr Amber Childs at Yale noted that chatbots also repeat biases “baked into medical practices for decades.” Training data reflects historical patterns in diagnosis and treatment that have long been criticised for underserving certain populations. When those patterns are reproduced at scale through AI chatbots used by millions, the reach of those biases expands considerably.
Looking Forward
Dr Bertalan Mesko noted that both OpenAI and Anthropic have released health-dedicated chatbot versions, suggesting the companies recognise the problem and are working on improvements. The broader goal, according to researchers, is to keep improving AI medical tools alongside “national regulations, regulatory guardrails and medical guidelines” that ensure chatbots meet a minimum standard of accuracy before people rely on them for health decisions. Until then, the gap between what people expect from AI medical advice and what it actually delivers remains wide.