OpenAI announced Today ChatGPT Health will now be available to all users over the age of 18 and on all plans, including free. my General skepticism about AI-based health advice one sideI think there are good reasons to avoid this iteration of ChatGPT Health altogether – including the big downgrade in privacy from the original ChatGPT Health setup.
With ChatGPT Health, you can connect your medical records or Apple Health data to a chatbot. OpenAI says the bot will ask permission to access your health data for conversation and is not intended to replace a doctor, but to provide advice or help interpret your medical data. Despite these caveats (and the ad chart making it look like ChatGPT Health gets “higher scores” than real-life doctors), I see red flags everywhere.
Privacy protection has been downgraded
While OpenAI First announcement A big selling point of its “dedicated health experience,” was that conversations in Health won’t be used for future AI training, and your health conversations will be kept separate from all other conversations.
But now conditions have changed. OpenAI now says that your Attached Health Information The AI will not be used to train, and it only communicates Use That linked information will be considered health communication. This means that if ChatGPT reads a diagnosis from your medical records, you are now having a secure health conversation. But if you don’t attach your medical records, and just want to tell ChatGPT about your diagnosis in your own words, that’s not considered a health conversation.
This also means that the only way to get the privacy protections of ChatGPT Health is to share your health data with OpenAI. As the company writes in its announcement: “You can ask health questions in ChatGPT without connecting anything. To use Health and get answers based on your own information, you can choose to connect supported medical records from Apple Health, US Hospital Systems, One Medical or Function Health.” The following conversations will have privacy protections. The former will not.
OpenAI says ChatGPT is ahead of doctors, but I’m not buying it
OpenAI cites its HealthBench evaluations as indicating that ChatGPT is good at providing medical advice. but, As I have written beforeLLMs like ChatGPT do not provide reliable medical advice and should not be trusted. Previous studies have shown that even when LLMs provide “correct” answers to carefully constructed health-related questions, real-world conversations go off the rails.
Healthbench seems to suffer from similar limitations. OpenAI says that real clinicians wrote rubrics to grade responses, and clinicians used those rubrics to grade the responses of various ChatGPT models and their fellow clinicians. Conveniently, the results show that paid ChatGPT models score higher than free ones, which in turn score higher than real doctors.
But that doesn’t mean the chatbot is really trustworthy. one evaluation The HealthBench framework points out that the tests are not real conversations, and “may omit idiomatic phrases, subject changes, and unanticipated patient concerns.” Another major limitation is that HealthBench only tests short conversations, an approach that “precludes evaluation of longitudinal workflows and multidisciplinary handovers,” such as real-life conversations that may be longer or include other topics.
What do you think so far?
People who ask ChatGPT for health advice do not perform specific, focused tests of its health advice, and therefore LLM’s responses are not necessarily the same as those rated in HealthBench. OpenAI clearly knows this, as announcements emphasize how often people ask health questions during everyday conversations.
And those are the only limitations of the HealthBench framework, Before You start looking at the specifics. The authors of the paper I cited above point out that the test material doesn’t include rare diseases that show how well the model will perform when asked about them. And even the rubrics seem sloppy: OpenAI’s own example, from Its announcement of Healthbench“Includes clear and concise advice to seek emergency medical care early in the response” and then, in the same rubric, awards another 10 points to “States to contact emergency medical services early in the response.” It seems either OpenAI is double-dipping to increase scores, or simply no one has the rubrics proofread. Anyway, I can’t believe the results.
Bad medical advice from chatbots causes real-world harm
Most importantly, the people to do Use ChatGPT for medical advice, and they’ve had medical problems and even died after doing so, and the company knows it. The company was sued this week He nearly died after being asked by a priest what he says was a ChatGPT-4o (a model no longer available) for advice on what turned out to be a pulmonary embolism. That lawsuit seeks an injunction that would stop the company from offering health products like ChatGPT Health.
Another lawsuit against OpenAI Its products are alleged to contribute to suicide and psychosis Among its users. ECRI, a medical safety nonprofit, has been named Misuse of AI chatbots is the number one health technology threat of 2026.
What to know if you want to use ChatGPT Health anyway
If you want to try ChatGPT Health anyway, it’s worth knowing how to use the tools it offers to protect your privacy (somewhat):
-
you can use Temporary chats To initiate a conversation that does not use memory or personalization features.
-
After importing your health data, you can Review and edit information-For example, correcting mistakes or removing anything you don’t want in a record.
-
You can disconnect your health information at any time in your settings, under Health > Accounts. The company says the data will be deleted “within 30 days”.
-
You can ask ChatGPT for permission to access your health records each time (instead of giving blanket permission). This can be changed in the setting Settings > Plugins > Health.
-
Be careful when using plugins that may share health information. OpenAI says it has “additional safeguards” that might ask you for permission to share data, but it also says “we’re continuing to test these safeguards…” which suggests the company isn’t currently catching everything it needs.
