OpenAI announced that ChatGPT Health will now be accessible to all users over 18, spanning across all plans including free ones. Putting aside general skepticism toward AI-based health advice, there are compelling reasons to avoid this iteration of ChatGPT Health completely, including a major downgrade in privacy protections compared to the original setup. Through ChatGPT Health, users can connect their medical records or Apple Health data directly to the chatbot. OpenAI states that the bot will request permission before accessing health data for conversations and is not designed to replace a physician, but rather to offer guidance or help interpret medical data. Despite these stated caveats, red flags persist regarding privacy and reliability.
Significant Downgrade in Privacy Protections
When OpenAI initially unveiled its dedicated health experience, a major selling point was that conversations within Health would remain excluded from future AI training and stay completely separated from all other chats. However, the terms have now shifted. OpenAI clarifies that while connected health information will not be utilized to train AI, only conversations directly leveraging that connected data are officially classified as Health conversations.
This structural change means that if ChatGPT reads about a diagnosis originating from uploaded medical records, the exchange qualifies as a protected Health conversation. Conversely, if users choose not to connect medical records and instead describe a diagnosis using their own words, that interaction is not categorized as a Health conversation. Consequently, the sole method to secure the privacy protections of ChatGPT Health is by sharing personal health data directly with OpenAI. As the company notes in its announcement, users can ask health questions without connecting anything, but utilizing Health with responses grounded in personal data requires connecting Apple Health, supported U.S. hospital medical records, One Medical, or Function Health. The latter category receives privacy protections, whereas the former does not.
Questioning the Reliability of Chatbot Medical Advice
OpenAI points to its HealthBench evaluations to suggest that ChatGPT excels at delivering medical advice. However, large language models do not provide reliable medical guidance and warrant profound caution. Previous studies indicate that even when an LLM produces correct answers to structured health inquiries, real-world user conversations frequently encounter complications. HealthBench appears encumbered by similar inherent limitations.
The company notes that practicing physicians authored rubrics to grade responses generated by various ChatGPT models as well as by fellow doctors. Conveniently, the outcomes demonstrate that paid ChatGPT tiers score higher than free versions, which subsequently outperform human physicians. Yet, high benchmark scores do not equate to genuine trustworthiness. An independent evaluation of the HealthBench framework highlights that these tests do not replicate real conversations and may omit idiomatic phrasing, topic shifts, and unanticipated patient concerns. Another critical limitation is that HealthBench exclusively tests brief conversations, precluding any assessment of longitudinal workflows or multidisciplinary handovers typical of real-world medical consultations.
Flaws in the HealthBench Testing Framework
Individuals turning to ChatGPT for health guidance do not conduct structured, isolated tests, meaning model responses will rarely mirror those evaluated in HealthBench. OpenAI clearly recognizes this reality, frequently emphasizing how often everyday users casually ask health-related questions during standard conversations. Furthermore, the testing material omitted a sufficient variety of rare diseases to evaluate performance accurately in complex scenarios. The grading rubrics also appear careless; an official OpenAI example awards 9 points for advising emergency medical care early in a response and subsequently awards another 10 points for instructing users to contact emergency services at the beginning. This discrepancy suggests either double-dipping to inflate scores or a lack of thorough proofreading, making the results difficult to trust.
Real-World Harms Stemming from Chatbot Medical Advice
Most critically, people actively utilize ChatGPT for medical guidance and have experienced severe health complications and even fatalities as a result, a risk the company acknowledges. A lawsuit filed this week by a pastor alleges he nearly died after consulting ChatGPT-4o, a discontinued model, regarding symptoms that turned out to be a pulmonary embolism. The legal action seeks an injunction compelling the company to halt health-focused products like ChatGPT Health. Additional lawsuits allege that company products contributed to user suicide and psychosis. Furthermore, the medical safety nonprofit ECRI designated the misuse of AI chatbots as the top health technology hazard for 2026.
Essential Safety Practices for Using ChatGPT Health
For those intending to experiment with ChatGPT Health despite these concerns, understanding how to navigate available privacy tools is essential
- Users can initiate temporary chats to conduct conversations without activating memory or personalization features.
- After importing health data, individuals can review and edit the stored information to correct errors or delete unwanted records.
- Connected health data can be disconnected at any time through settings under Health and Accounts, with the company stating data deletion occurs within 30 days.
- Users can configure ChatGPT to request explicit permission prior to accessing health records for every query by adjusting settings under Settings, Plugins, and Health.
- Caution should be exercised when utilizing plugins that might share health information, as OpenAI notes that additional safeguards are still undergoing testing, indicating that current detection systems may not catch every violation.
