Should You Trust ChatGPT With Your Health?

Summary

• OpenAI says more than 300 million people a week use ChatGPT for health questions — the vast majority on the free tier, and most outside the US. Karan Singhal frames the health push as the “make upside happen” leg of the AGI mission: “raise the floor of human health” with “trustworthy, abundant, reliable health intelligence for free or nearly free,” and raise the ceiling with capabilities that didn’t exist before.
• The cost curve is the tradeable claim: GPT-6 Luna, OpenAI’s cheapest frontier API model, beat 5.6 Sol — “literally our best model a month ago” — on HealthBench Professional “despite being thirty-four times cheaper.” Singhal’s explicit bet is “commoditization of health intelligence” (“we don’t wanna hoard that health intelligence”), while conceding health AI “has not yet had a macro effect” on the $5.3T US spend — compression requires integration and workflow deployment, not just cheap tokens.
• Singhal makes the case for specialized training inside a general model, based on his experience building specialized models: a healthcare-only model “would fall behind the frontier relatively quickly.” OpenAI adds specialized healthcare data into each phase of training across its ChatGPT models shipped over the past year — “specialized training in a general model… the best of both worlds,” “hitching a ride on that scaling frontier.”
• Trust is built through expert aggregation: “if you ask five physicians a question, you’ll get, like, eight different answers,” so OpenAI has a global network of hundreds of physicians across dozens of specialties and countries. When user comfort conflicts with clinical reality, “it’s basically a prioritization of safety” — the model must say “you should seek care” even when the user wants reassurance. The system-level opportunity: clinical guidelines can take “five to 17 years” to penetrate practice, and point-of-care AI could help compress that lag to “weeks or days or even hours.”
• OpenAI says AI can’t provide all of medical care: “I don’t think AI can or will provide all of that.” The strategy is to make “the last mile of actually having impact on patients just look like the last hundred meters for all of our partners” — ChatGPT as “a quarterback for your health,” not a replacement for care providers. His advice to healthcare builders: “bet on the models getting better. That’s probably the most important thing.”
• Enterprise traction is concrete: Epic integration, UCSF using it for chart review, and AdventHealth reporting “roughly eighty percent reduction” in administrative workflow costs. On agents, Singhal argues healthcare’s legacy stack — EHRs running on VMs, no “MCP connector tomorrow” — is ripe for computer-use agents, predicting traditional barriers “will dissolve in the coming months.” Rughani cites a New York Times article and a Blue Cross Blue Shield Association analysis, saying billing and paperwork bots increased spend by “a billion dollars or something like that”; Singhal calls the arms race “a bit of a short-term distraction.”
• Singhal’s biggest-impact five-to-ten-year prediction is new treatments from AI-guided target and molecule generation plus AI-guided prioritization and validation of clinical trials, where candidate molecules already outnumber trial slots. Nearer term, he’s looking for a “move thirty-seven for health” — predictions from longitudinal data that may seem brilliant only in hindsight, because “a human can’t integrate all that data in their head.”

Deep dive

1. Health is OpenAI’s “make upside happen” bet — at 300M users a week

• Singhal’s decomposition of the mission: build and deploy AGI, prevent downside, and “make upside happen” — health is a concrete upside play, with twin goals of raising the floor (“trustworthy, abundant, reliable health intelligence for free or nearly free”) and raising the ceiling. The scale claim: more than 300M weekly health and wellness users, “the vast majority” free users and most outside the US — deployment to the free tier was deliberate from the start.
• The gap being filled, per both speakers: weeks-to-months waits for specialist appointments even in the Bay Area, then fifteen minutes with a doctor and “six more questions that I wanted to ask” afterward. AI’s natural home is “in between these visits” — Rughani’s own dermatology follow-up is four months out despite writing down his questions beforehand.
• The specimen story Singhal returns to: Nathan Labenz (host of The Cognitive Revolution), whose young son had fatigue and irritability. AI helped them navigate testing and repeatedly said, “Hey, take this seriously,” even as doctors were less sure and downplayed it. The son was diagnosed with an aggressive form of leukemia, and Labenz used GPT-5 Pro throughout the care journey to understand lab reports before specialists had time to help interpret them.

2. Specialized training inside a general model

• Singhal built specialized medical models in a previous role, which informs his view: a healthcare-only model would “fall behind the frontier relatively quickly” because frontier progress is so fast. OpenAI instead adds specialized healthcare data to each phase of training across its ChatGPT models shipped in the last year.
• The strategic logic in his words: “if you bet on scale… things will get better,” so the health team is “hitching a ride on that scaling frontier” while retaining the ability to integrate health-specific evaluations and training data — “the best of both worlds.”

3. Proactive assistants and agents against a fragmented system

• The doctor-in-your-pocket vision Rughani sketches — memory, personalization, and behavioral nudges — is one Singhal says OpenAI is excited about: “a proactive personal health assistant,” with the hedge intact: “I don’t think that’s any kind of replacement for actually having a doctor.” Connected health information in ChatGPT is intended to be used securely, not for training data or ad targeting.
• Rughani’s Greg Brockman anecdote carries the safety-net argument: a friend double-checked a medication in ChatGPT, realized from her medical history she might have an interaction or adverse event, and her doctor agreed they hadn’t had time to review the history. Singhal’s generalization: a “safety net for both patients and clinicians,” studied in clinician-copilot experiments with Penda Health, with a similar patient-side role in view — for example, flagging a medication review if a diabetic user’s A1C rises.
• On agentic products (Rughani cites Instinct, Muse, GrokBot, and OpenAI’s Dots): healthcare “shouldn’t expect everything to have an MCP connector tomorrow”; primary EHRs often run on virtual machines, which can make some things easier for AI agents. Singhal says healthcare is “absolutely ripe for agents and computer use” and predicts barriers between patients, clinicians, and healthcare data “will dissolve in the coming months.”

4. Trust is built by aggregating experts and prioritizing safety over comfort

• Singhal’s operating principle — health “moves at the speed of trust” — means evaluation and training are tailored by persona. The original HealthBench evaluation, released in May 2025, examined how models adapt technical language for lay users versus clinicians.
• Why not rely on a few in-house doctors to adjudicate hard cases? “If you ask five physicians a question, you’ll get, like, eight different answers.” OpenAI built a global physician network of hundreds of experts across dozens of specialties, languages, and countries of practice, to aggregate expert perspectives where there may be disagreement.
• When a user wants reassurance the clinical picture doesn’t support, the priority is safety: evaluations track cases where the model should say “this is unsafe. You should seek care,” even if the user doesn’t want to hear it or wants an alternative treatment. On politicized gray zones such as unapproved peptides and vaccines, Singhal says to rely on experts to define response standards. If a doctor disagrees with ChatGPT, his advice is to handle it like other disagreements and seek a second opinion.

5. The 5-to-17-year guideline lag, honest hedging, and Mental Health Bench

• Singhal’s systems claim: biomedical papers inform yearly clinical guidelines, which are roughly “flowcharts”; studies have found it can take five to 17 years for guidelines to reach everyday practice. “You don’t want to be practicing medicine five to 17 years behind.” ChatGPT for Healthcare and ChatGPT for Clinicians bring current literature to the point of care, with the aim of helping compress that lag to “weeks, days, or even hours.” His reframe: sometimes “there is medical consensus, but it’s not being practiced.”
• Where uncertainty is genuine, models should recognize it and say, “Hey, there just isn’t enough information.” Hedging under uncertainty is a public aspect of HealthBench; another response to unsettled science is accelerating the science itself.
• Mental Health Bench extends the method to a domain where “people definitely disagree even more here than they do in physical health”: a baseline safety floor, including safeguards on Dots, plus measurement of upside — “positive mental health experiences.” OpenAI developed it with licensed experts worldwide, aiming for a “shared ruler or measuring stick” across model developers, the healthcare industry, and mental health experts.

6. Enterprise distribution: Epic, UCSF, AdventHealth, and the Overton window

• The three-part approach: capable models, products that lower the “activation energy” of using intelligence to improve health, and partnerships and deployment. Singhal cites 1.2B weekly ChatGPT users and says OpenAI wants to make frontier health intelligence free. He also describes security, compliance, HIPAA, and privacy work to support healthcare organizations.
• The traction evidence: Epic can be used in ChatGPT by providers and ChatGPT can be used in an Epic sidebar; UCSF is an early adopter using it for chart review; and AdventHealth has been a leader in deploying agents, reporting “roughly an 80% reduction” in administrative workflow costs. Retention on ChatGPT for Clinicians has been “a very pleasant surprise.”
• On seeing ChatGPT open on a doctor’s screen: the Overton window “is shifting in real time,” and Singhal predicts both patients and clinicians may feel safer with AI in their care journey — with the caveat, “I wouldn’t say we’re fully there.”

7. Information, not all of care — and the hunt for a “move thirty-seven for health”

• Rughani frames medical care as a bundled product — diagnosis, empathy, and legal liability. Singhal says, “I don’t think AI can or will provide all of that.” OpenAI’s role is to make “the last mile of actually having an impact on patients just look like the last hundred meters for all of our partners” — a “quarterback for your health” for navigation and logistics, while relying on partners to deliver care.
• What remains for the visit once both sides have pre-processed information through AI? “The human aspect of healthcare” — empathy, bedside manner, physical examination, and history-taking. Singhal says, “I don’t think all the information exists for the AI to just figure out what should be done.”
• On AlphaGo’s move 37 analogy — a prediction a human might not have made, but that seems correct and insightful in hindsight — Singhal says, “we’re very early.” In the clinician-copilot study, a common pattern the models flagged was avoiding antibiotic overuse and considering the global issue of resistance, not just the local case. The future version: models integrating a user’s longitudinal history over ten years and across EHR and chat data to make predictions humans couldn’t easily make by integrating all that information themselves.
• A third constituency beyond patient and clinician is the scientist. Agents may mine healthcare data “stored in boxes in various places” to generate new insights, and accelerate validation — which Singhal thinks may be a bottleneck, rather than proposing new ideas or treatments.

8. The cost curve, the billing-bot arms race, and what to build

• The economics centerpiece: GPT-6 Luna, OpenAI’s cheapest frontier API model, outperformed 5.6 Sol — “literally our best model a month ago” — on HealthBench Professional “despite being 34 times cheaper.” Singhal’s bet is that, as models reach a level of intelligence clinicians and patients can trust in the care journey, the cost at that level will fall — toward “commoditization of health intelligence. We don’t want to hoard that health intelligence.”
• Against Rughani’s macro framing — $5.3T in healthcare spending and a third-largest GDP if healthcare were a country — Singhal’s hedge is that health intelligence “has not yet had a macro effect.” Spend compression requires integration, workflows, and deployment at scale; abundant intelligence is a prerequisite, not the mechanism by itself.
• Rughani cites a New York Times article and a Blue Cross Blue Shield Association analysis, saying billing and paperwork bots increased spend by “a billion dollars or something like that.” Singhal describes possible short-term pain, but says there is no reason the arms race must continue. His ideal is a shared record of what happened in a visit that an AI can help adjudicate; he calls the arms race “a bit of a short-term distraction.”
• For builders: healthcare founders sit on a spectrum from betting against the models to betting for them — “bet on the models getting better. That’s probably the most important thing.” Procedural care’s near-term macro factor may be prevention, avoiding procedures through earlier action; robotics is “not a next-few-years kind of thing.” Singhal’s five-to-ten-year prediction is new treatments from AI-guided target and molecule generation, plus AI-guided prioritization and validation of clinical trials, since candidate molecules already exceed available trial slots.