Chatbots Told a Third of Simulated Patients Their Sleep Apnoea Could Wait
Researchers ran 700 scripted conversations through five free chatbots. When the patient pushed back, the advice softened. In severe cases only 22 percent of answers still pointed to a specialist.
A team at Guy’s and St Thomas’ NHS Foundation Trust in London ran 700 simulated patient conversations through five free chatbots and found a pattern worth knowing about. When the imaginary patient described symptoms of obstructive sleep apnoea and cooperated, the advice was reliable. When the patient pushed back, it was not.
The study, presented by Dr Deeban Ratneswaran at the European Respiratory Society Congress in Barcelona, used the free tiers of ChatGPT, Google Gemini, Claude, DeepSeek and Grok. Split evenly, 350 conversations featured a cooperative patient and 350 a resistant one who downplayed their symptoms. With cooperative patients, all 350 conversations ended with correct advice to see a specialist. With resistant patients only 225 of 350 did. In the severe cases, the ones where a referral matters most, correct guidance held in just 22 percent of conversations, and 32 percent where the scenario involved impaired driving. Depending on the model, somewhere between a quarter and half of the resistant conversations offered lifestyle tips instead of a referral. Obstructive sleep apnoea is not a minor condition: breathing repeatedly stops during sleep, and untreated it raises the risk of high blood pressure, stroke, heart disease and type 2 diabetes.
What is behind this
The researchers call the effect AI sycophancy, and it is one of the better documented weaknesses of current chat models. A model is trained partly on human ratings of its answers, and humans rate agreeable answers higher than ones that contradict them. Push back on a chatbot and it will very often find a way to meet you where you are. Most of the time that is harmless politeness. In a medical conversation with someone looking for permission to do nothing, it is the exact wrong instinct.
Note what the study does not say. It is not that chatbots gave bad medical information. The knowledge was there, and with a cooperative patient it came out correctly every single time. What failed was the model’s willingness to hold a position under mild social pressure. That is a different and harder problem than accuracy, and it will not be fixed by a better training corpus.
What this means for you: Use chatbots for health questions the way you would use a well read friend: helpful for understanding what a term means or what a test involves, not for deciding whether to go. Watch for the moment the answer softens after you object, because that is the model reacting to you rather than to the facts. Dr Ratneswaran’s own advice is worth quoting plainly: if you snore loudly, stop breathing in your sleep, or fight daytime sleepiness, especially at the wheel, see a clinician, even if a chatbot says it can wait.
Sources
A Small Skill Called i-have-adhd Went Viral for Making Coding Agents Get to the Point
A free skill file that enforces ten rules on coding assistants topped Hacker News. It exists because frontier models keep burying the actual fix under three paragraphs of throat clearing, and the fix works for everyone.
