OpenAI is now rolling out “Health in ChatGPT” to U.S. adults, letting you connect Apple Health, medical records, and wellness apps for help reviewing lab results or prepping for a doctor’s visit. More than 300 million people already ask the chatbot health questions every week. But here’s the catch that should make you pause: the quality of the advice you get depends entirely on your subscription.
Free users are served by GPT-5.5 Instant, a model that scores lower on health benchmarks. Paying subscribers get the stronger GPT-5.6 Sol. OpenAI will likely frame this two-tier system as responsible, pointing out that both models beat human doctors on the HealthBench Professional test. That defense crumbles under scrutiny. Benchmarks are artificial environments. Doctors in those tests might be fatigued or working without the patient records and colleague input they’d normally have. The tests can’t measure a physician’s ability to read nonverbal cues or draw on years of hands-on experience. OpenAI’s own announcement is littered with warnings that ChatGPT makes mistakes and can’t replace a doctor.
The feature’s design has already hit real-world friction. Early tests found over 70% of participants just asked health questions in the regular chat window because switching to the dedicated Health section was too clunky. OpenAI has since made health queries possible anywhere, while keeping the separate area for managing data. They’re also steering clear of Europe for now. When the feature was first announced in January, the EEA, Switzerland, and the UK were explicitly excluded—likely because of stricter GDPR rules and the risk of being classified as high-risk under the EU AI Act.
This all lands at a moment when the gap between AI’s confidence and its competence is dangerously wide. The latest radiology benchmark, RadLE 2.0, saw none of 16 AI models match human radiologists. The core problem wasn’t just wrong answers. It was chatbots delivering incorrect findings with unshakeable confidence instead of admitting uncertainty. That combination of overconfidence and sycophancy has already contributed to serious mental health harms. One researcher compared AI to an airplane’s autopilot—useful for routine tasks, but with ultimate responsibility staying firmly with the human in charge. That’s a fair analogy, as long as we remember autopilots don’t put their best safety features behind a subscription.
