Oxford researchers find subtle errors and oversimplification in the advice
-
Researchers at Oxford University warn that popular AI chatbots can deliver medical advice that sounds confident but may be incomplete, misleading, or unsafe.
-
The study found that even when AI systems cite credible sources, they can misinterpret guidelines or fail to account for patient-specific factors.
-
Experts say the findings highlight the growing need for regulation, transparency, and human oversight as AI tools become more common in health care.
A new study from the University of Oxford is raising new concerns about the reliability of medical advice generated by artificial intelligence, warning that widely used AI platforms may inadvertently put users at risk.
The research, conducted by a multidisciplinary team of clinicians and computer scientists, examined how large language models respond to common health-related questions. According to the study, AI systems frequently produced answers that appeared authoritative and well-structured, yet contained subtle errors, oversimplifications, or advice that conflicted with established medical guidance.
One of the most troubling findings was the tendency of AI platforms to generalize. Researchers found that chatbots often failed to distinguish between symptoms that require urgent medical attention and those that can be managed at home.
In some cases, the systems offered reassurance where caution was warranted, while in others they suggested unnecessary alarm.
How useful is the advice?
Dr Adam Mahdi, senior author on the study, said that while AI is able to give medical information, people "struggle to get useful advice from it."
"People share information gradually,"he told the BBC. "They leave things out, they don't mention everything. So, in our study, when the AI listed three possible conditions, people were left to guess which of those can fit.
The study also highlighted issues with sourcing. While AI-generated responses sometimes referenced reputable organizations or clinical guidelines, the models did not always apply those sources correctly. Advice could be outdated, taken out of context, or mismatched to the users situationparticularly for people with chronic conditions, multiple medications, or atypical symptoms.
Problematic design
The researchers stressed that the problem is not malicious intent, but design. Most AI platforms are not built to practice medicine, yet they are increasingly used that way as consumers turn to chatbots for quick answers about symptoms, medications, and treatments.
Oxfords team is calling for clearer warnings to users, improved training data, and stronger collaboration between AI developers and medical professionals. The team also emphasized that AI tools should supplement, not replace, qualified health care providers.
Posted: 2026-02-10 18:50:59


















