OpenAI is rolling out its "Health in ChatGPT" feature to users in the United States aged 18 and older, following a testing period that began in January. The feature allows users to connect Apple Health data, medical records, and wellness applications to review lab results, prepare for doctor's appointments, and analyze sleep or activity data. The move builds on years of consumer-tech expansion into personal health data — Apple's HealthKit and Google Fit have aggregated wearable and app data since 2014 — but OpenAI goes further by attempting to unify medical records, lab interpretations, and conversational health guidance within a single AI assistant already used by hundreds of millions. OpenAI states it will not use connected health data for model training or advertising purposes. In the United States, medical records held by providers and insurers are protected under HIPAA, though health data routed through consumer apps and AI platforms can fall into regulatory gaps where HIPAA does not apply — a distinction the Federal Trade Commission has increasingly flagged in enforcement actions against health app developers.

The announcement, detailed on OpenAI's official website, comes as the company reports that more than 300 million people now ask ChatGPT health questions each week — a figure up from 230 million in January. OpenAI says more than 260 physicians participated in developing the Health features.

Tiered Model Access Creates Quality Gap

The feature operates on a two-tier system. Users on the free version of ChatGPT receive health guidance powered by GPT-5.5 Instant, which scores lower on health benchmarks than the new flagship model, GPT-5.6 Sol, which is reserved for paying subscribers.

OpenAI is expected to defend this structure by noting that both models outperformed doctors' answers on the HealthBench Professional test. However, benchmark results are drawn from artificial test environments designed to measure knowledge. Doctors may score lower for several reasons, including time pressure, fatigue, or taking the test without access to patient records or input from colleagues. Benchmarks also cannot replicate the dynamics of an in-person medical exam, where physicians examine patients directly, observe nonverbal cues, and draw on years of clinical experience. The tiered approach also means that users who cannot afford a subscription receive quantifiably lower-quality health guidance — a dynamic that raises access questions as AI assistants increasingly become a first stop for medical information, particularly among younger and lower-income demographics who are more likely to use free tiers.

OpenAI itself repeatedly cautions in its announcement that ChatGPT can still make errors and cannot replace professional medical advice.

Usability Improvements Following Early Testing

OpenAI's early testing revealed that more than 70 percent of participants asked health questions outside the dedicated Health section because switching to it was too cumbersome. In response, the company has made Health available in any conversation while retaining a separate Health section for managing connected data and accessing past health chats.

European Launch Remains Uncertain

The company has not indicated whether or when the Health feature will be available in Europe. When OpenAI first announced the feature in January, it specifically excluded the European Economic Area, Switzerland, and the United Kingdom. Stricter EU data privacy rules and the possibility that the feature could be classified as high-risk under the EU AI Act are cited as likely reasons for the exclusion. GDPR Article 9 designates health data as a special category requiring explicit consent and additional safeguards, and the EU AI Act's high-risk classification for certain health-related AI systems could trigger conformity assessments and ongoing compliance obligations that do not apply in the US market.

AI Health Risks Are Not Hypothetical

The limitations of AI in medical contexts are well documented. In the latest radiology benchmark, RadLE 2.0, none of the 16 AI models tested performed as well as human radiologists. The primary issue identified was that chatbots delivered incorrect findings with high confidence rather than acknowledging uncertainty — an area where human radiologists performed significantly better.

That combination of overconfidence, persuasion, and sycophancy — where chatbots validate users' assumptions rather than challenge false ones — has also been linked to serious mental health harms.

At the same time, some studies show AI identifying patterns in health data that medical professionals miss, occasionally with striking results. MIRA, a system designed for electronic health records, and AMIE both performed approximately as well as primary care doctors in simulated consultations. One researcher compared AI agents like these to an airplane's autopilot: "These systems can support and relieve medical professionals by taking over routine tasks, but ultimate responsibility will always remain with the physicians." The FDA has been developing a regulatory framework for AI and machine learning–based Software as a Medical Device, but consumer-facing chatbots that provide general health information may operate outside that framework, particularly when they are not marketed as diagnostic tools — leaving a evolving boundary between helpful guidance and unregulated medical advice.