AI chatbots give inaccurate health advice, Oxford study finds
Researchers at Oxford University tested AI chatbots including GPT-4o, Llama 3, and Command R+ on health scenarios with UK participants. The study found chatbots were no better than internet search engines at helping users identify health problems or determine whether medical care was needed.
What happened
A British-led team of researchers presented nearly 1,300 UK-based participants with 10 different health scenarios such as a headache after a night out drinking, a new mother feeling exhausted, or what having gallstones feels like. Participants were randomly assigned one of three AI chatbotsโOpenAI's GPT-4o, Meta's Llama 3, or Command R+โor a control group that used internet search engines.
People using AI chatbots identified their health problem around a third of the time, while only around 45 percent figured out the right course of action, no better than the control group using traditional internet search. The study, published in Nature Medicine, found a disparity between these results and how AI chatbots score highly on medical licensing exams.
Reported impact
Rebecca Payne from Oxford University stated that patients need to be aware that asking a large language model about their symptoms can be dangerous, giving wrong diagnoses and failing to recognise when urgent help is needed.
One out of every six US adults asks AI chatbots about health information at least once a month, with that number expected to increase as more people adopt the technology.
Index analysis
Researchers attributed the gap between AI chatbots' high performance on medical exams and their poor real-world performance to a communication breakdown. In simulated patient interactions used to test AI, humans typically provided all relevant information, whereas real humans often did not give chatbots all relevant information and sometimes struggled to interpret or misunderstood the chatbot's advice.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.