ChatGPT health advice worse for free users, better for paying subscribers

OpenAI is expanding its presence in consumer health by rolling out 'Health in ChatGPT' to U.S. users aged 18 and older, a feature the company first unveiled and began testing in January. The feature allows users to connect Apple Health, medical records, and wellness apps directly to ChatGPT, enabling them to review lab results, prepare for doctor's appointments, and analyze sleep or activity data. OpenAI has stated that connected health data will not be used for model training or advertising purposes.

The key distinction lies in the quality of health advice users receive based on their subscription status. Free users are powered by GPT-5.5 Instant, while paying subscribers gain access to GPT-5.6 Sol, OpenAI's new flagship model. On OpenAI's HealthBench Professional benchmark, GPT-5.6 Sol substantially outperforms both GPT-5.5 Instant and the older GPT-4o model across every tested category. The largest gaps appear in completeness (88.0 percent for GPT-5.6 Sol versus 53.2 percent for GPT-5.5 Instant) and health decision helpfulness (83.0 percent versus 50.8 percent). Both models also beat physician-written answers on this benchmark—a fact OpenAI appears likely to cite as ethical justification for the two-tier system.

However, benchmark results have significant limitations. They measure knowledge in artificial test environments and do not capture what occurs during actual medical exams, where doctors examine patients in person, observe nonverbal cues, and draw on years of clinical experience. Physicians may also score lower on such tests due to time pressure, fatigue, or lack of access to tools like patient records or colleague input. OpenAI itself repeatedly emphasizes in its announcement that ChatGPT can make mistakes and cannot replace medical advice. The company notes that more than 260 physicians contributed to developing the Health features.

Early testing revealed a usability challenge: over 70 percent of participants asked health questions outside the dedicated Health section because switching to it was too cumbersome. In response, OpenAI has made Health available within any ChatGPT conversation while maintaining a separate Health section for managing data and accessing past health chats. The scale of demand for health guidance on the platform is substantial—OpenAI reports that more than 300 million people now ask ChatGPT health questions each week, up from 230 million in January.

The feature's geographic rollout is limited. OpenAI has not announced whether or when Health in ChatGPT will become available in Europe. When the feature was announced in January, OpenAI specifically excluded the European Economic Area, Switzerland, and the United Kingdom—a decision likely driven by stricter EU data privacy regulations and the possibility that the feature could be classified as high-risk under the EU AI Act. These regulatory concerns reflect real risks: recent research on radiology AI (RadLE 2.0) found that none of 16 AI models tested performed as well as human radiologists, often because chatbots delivered incorrect findings with high confidence instead of acknowledging uncertainty. Conversely, some AI systems for health records (like MIRA and AMIE) have performed comparably to primary care doctors in simulated consultations, suggesting AI can effectively support routine diagnostic work under physician oversight.

OpenAI's Health feature launch represents a deliberate two-tier approach to AI-assisted medical guidance, with the company explicitly restricting its most capable health model to paying subscribers. The benchmarks OpenAI uses show substantial performance gaps—GPT-5.6 Sol beats both physician-written answers and the free-tier model across every category tested. However, the company's own caveats reveal the limits of these results: benchmark tests measure knowledge in artificial environments and cannot replicate the diagnostic work of in-person medical exams, where doctors assess nonverbal cues and draw on clinical experience. OpenAI's disclosure that more than 260 physicians helped develop the feature suggests effort toward medical credibility, yet the company simultaneously warns that ChatGPT still makes mistakes and cannot replace medical advice.

The practical reality is more nuanced than the performance numbers suggest. Recent research on AI radiology systems (RadLE 2.0) found that none of 16 AI models tested performed as well as human radiologists, often because chatbots express incorrect findings with high confidence rather than acknowledging uncertainty. At the same time, other systems like MIRA and AMIE have demonstrated performance comparable to primary care doctors in simulated consultations, suggesting AI can support routine tasks under physician oversight. OpenAI's early testing revealed that over 70 percent of users asked health questions outside the dedicated Health section because switching between modes was cumbersome—a friction the company has now reduced by making Health available in any conversation. The feature's absence from Europe, tied to stricter data privacy and potential AI Act classification, underscores regulatory tensions the company faces globally. With over 300 million people now asking ChatGPT health questions each week (up from 230 million in January), the scale of this rollout and its tiered quality model will likely shape how millions approach preliminary health research.

No comments yet. Be the first to share your thoughts!