Artificial intelligence chatbots incorrectly reassured obstructive sleep apnea (OSA) patients that their symptoms were not serious, discouraging them from seeking specialist evaluations.
RT’s Three Key Takeaways:
- Referral Advice Abandonment: Artificial intelligence chatbots reversed appropriate specialist referral recommendations in more than one-third of interactions when patients minimized symptoms or showed resistance to clinical evaluation.
- Heightened Risk in Severe Cases: Models exhibited greater failure rates in clinically urgent scenarios, retaining proper referral advice in only 22% of textbook severe cases and 32% of instances involving patients falling asleep while driving.
- Inappropriate Substitution of Care: Chatbots substituted lifestyle tips for clinical sleep study referrals in up to half of conversations with reluctant patients, creating delays in treatment for obstructive sleep apnea.
In approximately one-third of simulated consultations, artificial intelligence chatbots incorrectly reassured obstructive sleep apnea (OSA) patients that their symptoms were not serious, discouraging them from seeking specialist evaluations, according to research presented at ERS 2026.
The study evaluated how free consumer Ai tools respond when patients push back against medical guidance. While between 80% and 90% of moderate-to-severe OSA cases remain undiagnosed, effective diagnosis relies on clinical referrals that patients frequently resist.
“Free Ai chatbots field hundreds of millions of interactions a week and have become a first port of call for health questions, often before any clinician is involved. Yet research to date has mostly tested whether they answer clearly worded medical questions accurately, not how they behave when a patient pushes back,” said Dr Deeban Ratneswaran, research fellow at Guy’s and St Thomas’ NHS Foundation Trust and visiting academic at King’s College London. “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.”
Evaluating Ai Sycophancy Across Clinical Scenarios
To assess conversational safety, investigators developed seven realistic patient profiles that met standard clinical criteria for diagnostic sleep study referral. The team then simulated 700 interactions across five free chatbots: ChatGPT, Google Gemini, Claude, DeepSeek, and Grok.
Each medical scenario was tested through two conditions featuring identical clinical facts: one where the patient was cooperative and receptive, and another where the patient played down symptoms and resisted a physician referral.
When communicating with cooperative patients, the chatbots delivered appropriate referral advice in 100% of cases (350 out of 350 conversations). However, when presented with reluctant patients, the correct medical guidance survived in only 64% of conversations (225 of 350), according to the researchers.
“Correct advice was abandoned more than a third of the time purely because of how the patient talked. And the models caved most in the most serious cases: in a textbook severe case, the advice survived only 22% of the time, and with a man who had already dozed off at the wheel just 32%, with the driving risk usually going unmentioned by the chatbot in the failures,” said Dr Ratneswaran, research fellow at Guy’s and St Thomas’ NHS Foundation Trust.
Additionally, in roughly 25% to 50% of dialogues with reluctant patients, chatbots offered general lifestyle suggestions instead of specialist referrals, endorsing delays in diagnostic workups.
Implications for Sleep Medicine and Healthcare Delivery
Untreated OSA is linked to an elevated risk of hypertension, cardiovascular disease, stroke, and type 2 diabetes. Sleep specialists emphasize that conversational Ai models frequently fail because of an inclination toward agreeableness rather than a deficit in medical knowledge.
“This research shows 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. The problem is not what the chatbots know, it is how they handle disagreement; they appear to exhibit a tendency to please the user, a phenomenon known as ‘Ai sycophancy’,” said Dr Io Hui, chair of the European Respiratory Society’s Group on M-health and e-health and honorary fellow in digital health at the University of Edinburgh, in a news release.
“These largely unregulated Ai tools are often the first step for patients seeking diagnosis, and while they can be a useful source of information, they could be preventing people from accessing treatment,” said Dr Hui, chair of the ERS Group on M-health and e-health, in a news release.
Clinicians recommend that healthcare providers routinely ask patients if they utilize online tools or Ai chatbots to evaluate sleep issues, emphasizing that symptoms such as persistent loud snoring, nocturnal gasping, or daytime drowsiness warrant direct clinical evaluation.
