HealthApril 18, 2026

AI Chatbots Deliver Wrong Health Advice in Real‑World Use, Accuracy Falls to 35%

Oxford study shows AI chatbots diagnose correctly only 35% of the time in real‑world use, despite 95% accuracy with full data. England’s CMO warns advice is often confident but wrong.

Health & Science Editor

TweetLinkedIn
AI Chatbots Deliver Wrong Health Advice in Real‑World Use, Accuracy Falls to 35%

TL;DR When people ask AI chatbots for a diagnosis, the tools are correct only about one‑third of the time, despite scoring 95% when given complete medical information. England’s top doctor says the advice is often confident but wrong.

**Context** Abi from Manchester uses ChatGPT to manage her health anxiety. She finds the chatbot gives tailored advice, but in January it told her she had punctured an organ and needed A&E. After three hours in emergency care she learned the advice was wrong. Her experience mirrors a growing trend of the public turning to AI for quick health answers.

**Key Facts** Oxford’s Reasoning with Machines Laboratory ran a controlled experiment. Doctors created realistic health scenarios ranging from minor issues to life‑threatening emergencies. When chatbots received the full medical picture, they were correct 95% of the time. In the second phase, 1,300 participants interacted with the chatbots to obtain a diagnosis; accuracy fell to 35%, meaning two‑thirds received incorrect advice. England’s Chief Medical Officer, Prof Sir Chris Whitty, stated that the answers are “not good enough” and often “both confident and wrong.” A separate analysis by The Lundquist Institute tested Gemini, DeepSeek, Meta AI, ChatGPT and Grok on topics such as cancer and vaccines; more than half of the responses were classified as problematic.

**What It Means** Users should treat chatbot output as a preliminary suggestion, not a definitive diagnosis. For urgent or unclear symptoms, seek a qualified clinician. The gap between full‑information performance and real‑world use shows that missing or ambiguous user input drives errors. Watch for upcoming studies that test AI‑assisted triage in clinical settings and for any regulatory guidance on medical AI chatbots.

**What to watch next** Upcoming pilot programs that integrate chatbots into NHS pathways and forthcoming policy statements from England’s Chief Medical Officer on AI safety in health advice.

TweetLinkedIn

Reader notes

Loading comments...