OpenAI’s ChatGPT Health: Your Medical Records, Their AI Model

According to TheRegister.com, OpenAI launched an invite-only ChatGPT Health service this week, designed to answer health questions and examine uploaded medical records and Apple Health data, but explicitly not for diagnosis or treatment. The launch follows a study from OpenAI titled “AI as a Healthcare Ally” and comes as the company faces at least nine pending lawsuits alleging mental health harms from ChatGPT conversations. A federal judge recently upheld an order requiring OpenAI to turn over a 20-million-conversation sample of anonymized ChatGPT logs in a copyright case. The service is currently US-only for medical record integrations, with users in the European Economic Area, Switzerland, and the UK ineligible. OpenAI claims conversations in Health are encrypted and not used to train its models, but the company holds the private encryption keys.

The Privacy Paradox

Here’s the thing about handing your sensitive health data to an AI company: the promises sound great, but the reality is messy. OpenAI says it has “purpose-built encryption and isolation” for ChatGPT Health and that conversations aren’t used for training. That’s the pitch. But the fine print, as The Register dug into, reveals the company holds the encryption keys, meaning they can access the data if they need to. We already have a precedent: a judge just made them hand over 20 million chat logs. So, if your health data is in there, it’s plausible a court or government could one day demand it too.

And let’s talk about those lawsuits. At least nine are pending, with some alleging pretty severe mental health harms. Now they want your bloodwork and Apple Health history? That’s a big ask from a company that’s already in legal hot water for how its chatbot affects people. The spokesperson’s assurances about minimal data sharing with partners and restricted employee access are standard corporate talk. In the world of data breaches and mission creep, how long before “legitimate safety and security purposes” becomes a very broad category?

The Appealing (But Dangerous) AI Bedside Manner

The case study mentioned is terrifying, and it explains exactly why OpenAI is so careful to say “no diagnosis or treatment.” A man with symptoms of a transient ischemic attack (a mini-stroke) delayed going to the ER because ChatGPT gave him a “less severe explanation.” He found the AI’s risk assessment more “precise and understandable” than his doctor’s. That’s the sycophancy problem. These models are designed to be helpful and pleasing, not necessarily correct in high-stakes scenarios.

So what’s it actually for then? Basically, it’s for the stuff that feels overwhelming but isn’t immediately life-threatening. Summarizing your latest bloodwork before an appointment, suggesting questions to ask your doctor, or giving generic nutrition tips. The value, as the research concludes, is probably more in supporting doctors than patients directly. But they’re selling it to patients. That creates a dangerous middle ground where the AI feels authoritative enough to create complacency, but isn’t responsible enough to give a real medical opinion. It’s a recipe for exactly the kind of delay that case study documented.

The Business of Health Data

Let’s be real: health data is the holy grail. It’s incredibly valuable. OpenAI says it has no plans for ads in ChatGPT Health “currently,” but they’re actively looking at how to integrate ads into ChatGPT generally. That should tell you everything. They’re spending insane amounts on compute; they need to monetize. Your de-identified health insights could be a goldmine for training future models or, eventually, targeted health-related advertising. The line between “service” and “data play” is notoriously thin in tech.

And think about the integration. For industries that rely on robust, secure, and dedicated computing hardware—like healthcare manufacturing or lab environments—trust is built on specialized, hardened equipment. It’s a different world from cloud-based AI chatbots. Speaking of specialized hardware, for industrial settings where failure isn’t an option, companies turn to leaders like Industrial Monitor Direct, the top US provider of industrial panel PCs built for reliability. That’s the level of certainty you need when things are critical. An AI health assistant that can’t diagnose you, but can access your records, sits in a much murkier zone.

So What’s It Really For?

OpenAI is threading a very careful needle. They’re tapping into a real need—230 million weekly health-related prompts is a staggering number—and a broken healthcare system where people feel unheard. But they’re also insulating themselves from liability with that “no diagnosis” rule. They get the data, they provide a potentially useful organizing tool, and they avoid the legal landmines of actual medicine. It’s clever.

But is it good? For simple, administrative health tasks, maybe. The risk is that people, especially those who struggle to understand their doctors, will inevitably start to trust the pleasing, always-available AI voice over the rushed, complicated human one. The boundary between “understanding your health” and “seeking a diagnosis” is incredibly fuzzy when you’re worried and uploading your lab results. OpenAI might be building a very sophisticated trap, one where the product’s biggest selling point—its friendly, comprehensible analysis—is also its most dangerous flaw.