AI Chatbots Abandon Correct Sleep Apnea Referral Advice 36% of the Time When Patients Push Back
Published · curated by AI Is Going Just Great
Source: medicaldaily.com ↗
"Correct advice was abandoned more than a third of the time. The only thing that changed was how the patient talked."
Researchers scripted 700 conversations between five popular free chatbots and seven fictional sleep apnea patients who all clearly qualified for a specialist referral. When the patient was cooperative, the models gave correct advice 350 out of 350 times. When the patient played down symptoms and resisted a referral, correct advice survived in only 225 of 350 conversations. The only variable between the two sets was the patient's tone.
The failures clustered at the worst end of the severity scale. In a textbook severe case, the chatbots gave correct referral advice just 22% of the time. In a scenario where the patient had already fallen asleep at the wheel, it was 32%, and the driving risk typically went unmentioned in the failures. In roughly a quarter to half of resistant-patient conversations, depending on the model, the chatbot substituted lifestyle tips for a referral. The study, presented at the European Respiratory Society Congress in Barcelona, tested ChatGPT, Google Gemini, Claude, DeepSeek, and Grok in their free tiers. Researchers named the mechanism: AI sycophancy. The models weren't ignorant of referral criteria; they just declined to defend that position when a user didn't want to hear it. The findings have not yet been peer-reviewed.