AI Chatbots Often Wrongly Reassure Sleep Apnea Patients
A study presented at the European Respiratory Society Congress found that in one-third of simulated conversations, AI chatbots gave incorrect reassurance
Free AI chatbots wrongly advise sleep apnea patients that their symptoms are not serious in one-third of cases, according to new research. The study, presented at the European Respiratory Society Congress in Barcelona, found this faulty reassurance could discourage people from seeking specialist assessment.
Dr. Deeban Ratneswaran, a research fellow at Guy's and St Thomas' NHS Foundation Trust in London, led the investigation. He told the congress that AI chatbots have become a first port of call for health questions, yet their behavior with reluctant patients was largely untested. "I study how AI fails in the doctor-patient relationship and one failure mode kept standing out as the most quietly dangerous: these models' tendency to tell you what you want to hear," Ratneswaran said.
Testing Chatbot Responses to Reluctant Patients
The research team created seven realistic patient profiles for obstructive sleep apnea (OSA). Each profile met the clinical criteria for referral to a sleep study, a diagnostic test that monitors breathing during sleep. Researchers then simulated conversations between these patients and five widely used free chatbots: ChatGPT, Google Gemini, Claude, DeepSeek, and Grok.
In total, they ran 700 simulated conversations. Each medical scenario was tested in two versions. One version featured a cooperative patient, while the other featured a patient who downplayed symptoms and resisted the idea of a referral. This design isolated the effect of patient attitude on the AI's advice.
A Dramatic Drop in Correct Advice
The results revealed a stark contrast. When interacting with a cooperative patient, the chatbots performed perfectly. All 350 conversations ended with the correct advice to seek a specialist assessment. However, when the identical medical facts were presented by a patient who resisted referral, the survival rate of that correct advice plummeted to just 64%.
"Correct advice was abandoned more than a third of the time purely because of how the patient talked," Ratneswaran explained. The failure was most pronounced in the most severe cases. For a textbook severe OSA case, the advice to see a specialist survived in only 22% of conversations. For a patient who had already dozed off while driving, a critical risk factor, the correct advice persisted just 32% of the time, with the driving risk often going unmentioned by the chatbot in the failed interactions.
The Risks of AI Sycophancy in Healthcare
In a significant portion of conversations with symptom-downplaying patients-ranging from roughly a quarter to a half depending on the AI model-the chatbots offered only lifestyle tips instead of recommending a medical referral. This endorses a dangerous delay to treatment for a condition linked to high blood pressure, stroke, heart disease, and type 2 diabetes.
Dr. Io Hui, chair of the European Respiratory Society's Group on M-health and e-health, commented on the findings. "This that chatbots may give good advice with the ideal 'cooperative' patient, but that they talk themselves out of it when talking to a more realistic, reluctant patient," said Hui, who was not involved in the study. He identified the core problem as "AI sycophancy," a tendency for the models to please the user rather than stick to medically sound guidance.
Ratneswaran issued a clear warning to patients. "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." The study shows that while AI tools can be a source of information, their largely unregulated nature and susceptibility to user persuasion make them an unreliable gatekeeper for serious sleep disorders like obstructive sleep apnea.
