What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal

Healthcare’s scale has not eliminated scarcity: the discussion cites more than $5.3 trillion in spending, over 100 million people without a regular primary care doctor, and waits exceeding 40 days for many who have one. OpenAI’s Karan Singhal says more than 300 million people ask ChatGPT health and wellness questions each week, most on free accounts and most outside the United States. OpenAI wants to make trustworthy health intelligence free or nearly free, and eventually offer a frontier-level health model to ChatGPT’s 1.2 billion weekly users. Singhal sees AI helping people make sense of concerns between appointments, prepare questions, navigate care, and understand clinicians’ advice—not replacing medical care.

Singhal says OpenAI trains healthcare expertise into its general-purpose ChatGPT models, combining frontier capabilities with healthcare-specific data, physician input, and evaluations. He argues a separate, smaller medical model could quickly fall behind as general models advance. A global physician network informs training and testing, but medical answers are not always settled: Singhal says five physicians might give eight different answers. Health Bench, released in May 2025, tests issues such as whether a model adapts its language for clinicians or laypeople and acknowledges uncertainty. He says safety must take priority over reassurance, including telling users to seek care when a situation appears unsafe. If a patient or doctor disagrees with ChatGPT, he recommends treating it like other medical disagreements, including seeking a second opinion.

The intended destination is a proactive personal health assistant that uses an individual’s context while augmenting clinicians. Singhal describes studying a clinician co-pilot with PEN to Health that could flag medication interactions for review, or alert a doctor to a diabetic patient’s rising A1C. AI might also flag possible medication-adherence problems or help clinicians notice patterns across a person’s history. Singhal says agents could help navigate fragmented care and disconnected data, but many health systems still run on legacy software, and clinicians may reach electronic records through virtual machines. OpenAI is working with Epic and other partners on workflows; ChatGPT is available within Epic, including through a sidebar. UCSF is an early adopter of an Epic connection for chart review and efforts to reduce fragmented work.

AI could also bring newer evidence into appointments: Singhal says clinical guidelines can take five to 17 years to reach everyday practice, and OpenAI’s healthcare offerings aim to make recent biomedical literature available to patients and clinicians. He notes that visits may last about 15 minutes, so tools that help answer questions during the encounter could matter. Mental health requires particular care: OpenAI works with licensed experts to define model behavior and measure supportive responses as well as safety. Singhal says experts disagree more on mental-health scenarios than physical-health ones; Mental Health Bench provides a shared measure, while safeguards address users in distress or crisis. He expects acceptance to grow, but says appointments must still leave room for empathy, uncertainty, examination, history-taking, and investigation.

Singhal’s ambition is a health “Move 37”: AI findings that surprise experts and prove useful, though he says such examples are still early. He imagines models drawing on a decade of electronic health records, chats, and other personal history to identify risks people might not readily infer. Agents could extract more value from underused records, claims, and longitudinal data, while AI might speed the validation process that Singhal sees as a greater constraint than generating ideas or treatments. He also describes a PEN to Health study in which AI flagged clinicians’ tendency to overuse antibiotics, underscoring that decisions should account for antibiotic resistance as well as the individual case.

Singhal expects automation in chronic-care management, documentation, administration, and back-office work to put pressure on healthcare spending, but says AI has not yet produced a macro-level spending effect; savings depend on integration, workflows, deployment, and scale. He says GPT6 Luna, then OpenAI’s cheapest frontier API model, outperformed 5.6 sold on Health Bench Professional at 34 times lower cost, while acknowledging that current models still need improvement on difficult clinician workflows. AI billing and paperwork bots may instead fuel a costly arms race: Singhal cites a Blue Cross Blue Shield Association analysis attributing about $1 billion in increased spending to such bots. He favors a shared account of what happened during a visit, followed by AI adjudication and limited appeals. Over the next year or two, he expects better prevention and targeting to help some patients avoid expensive procedures; personalized medicines and faster clinical trials are a five-to-ten-year prospect, while healthcare robotics faces longer-term technology, reimbursement, and integration hurdles.