Karan Singhal on AI, Trust and the Future of Health Care

Karan Singhal, who leads health AI research at OpenAI, joins a16z partner Jay Rughani to discuss how AI might support patients and clinicians. Singhal frames the ambition as both making health information more accessible and expanding what medicine can eventually do. These are the company’s goals and research claims, not proof that an AI system can replace a clinician.

Singhal says more than 300 million people use ChatGPT for health and wellness questions each week. He connects this demand to difficulties navigating care: finding a specialist can take weeks or months, a short appointment may leave questions unanswered, and patients may need help understanding what to ask or how to follow up.

He sees an opportunity between appointments, where a patient can organize concerns, prepare questions and review information from a visit. Rughani describes forgetting a question during his own dermatology appointment and facing a long wait for another visit. The discussion concerns support around care, rather than a substitute for examination or treatment.

Singhal recalls the publicly shared story of podcast host Nathan Labenz and his young son: according to Labenz, AI helped the family recognize that symptoms warranted testing, and doctors subsequently diagnosed leukemia. Singhal says the family also used AI to help understand information during care. An individual account cannot establish diagnostic accuracy or the reliability of such tools across patients.

Rughani asks why OpenAI builds medical expertise into a general model instead of using an entirely separate medical model. Singhal says specialized health data and physician feedback are incorporated across training stages, allowing health capabilities to benefit from broader model improvements. That is a design choice, not a guarantee of clinical performance.

From his earlier experience with specialized models, Singhal argues that a smaller separate model may quickly fall behind as general systems improve. His team combines general capabilities with health-specific evaluation and training, hoping to retain the advantages of scale while addressing particular medical tasks.

Rughani imagines a personal assistant that remembers a person’s medical history and preferences, explains information and helps them follow a clinician’s plan. Singhal is enthusiastic about a proactive assistant but explicitly says he does not see it as replacing access to a doctor.

Singhal describes privacy protections for information connected through Health in ChatGPT. OpenAI’s [current help documentation](https://help.openai.com/en/articles/20001036-health-in-chatgpt) distinguishes connected health records and conversations using them from other conversations: the former are not used to train foundation models or target ads, while other chats follow the user’s ordinary data-control settings.

The host recounts another anecdote in which a patient checked a proposed medication and raised a possible interaction with a doctor. Singhal describes a possible safety net for patients and clinicians: flag an issue for professional review rather than silently assume that a suggested action is safe.

He cites a study with Panda Health in which an AI copilot offered clinicians suggestions during their work, such as flagging a possible medication interaction. He also imagines future alerts that prompt a clinician to review medication when a chronic-care patient’s laboratory measurements change. These are study examples and proposals, not individualized clinical advice.

Rughani asks whether personal agents could help with appointments, insurance coverage, discharge instructions and billing disputes. Singhal says fragmented records and workflows create friction for patients and clinicians. He expects agents and computer use to become useful even where legacy software lacks modern integrations.

Trust requires more than an accurate-sounding answer. Rughani asks how the same model can explain a laboratory result to a patient while also helping a specialist understand a complex case. Singhal says the team evaluates and trains responses for different users, including laypeople, experienced patients and clinicians.

He points to HealthBench’s evaluation of communication, including whether technical language fits the audience. [The published HealthBench paper](https://cdn.openai.com/pdf/bd7a39d5-9e9f-47b3-903c-8b847ca650c7/healthbench_paper.pdf) describes physician-authored criteria for health conversations and also discusses the limitations of evaluations. A benchmark score is not the same thing as proven patient benefit in routine care.

A difficult case arises when someone wants reassurance but their symptoms call for a different response. Rughani asks how a product balances user satisfaction against medical quality. Singhal says safety must take priority and that evaluations include situations in which the model should recommend seeking appropriate care despite the user’s preferences.

He says OpenAI works with hundreds of physicians across specialties, languages and countries rather than relying on a handful of clinicians to settle uncertain cases. Experts can disagree, so the team seeks ways to aggregate their judgments and measure how models behave in consequential situations.

What if a doctor disagrees with an AI answer? Singhal says patients can discuss the disagreement and seek another clinician’s opinion, as they can with disagreements arising in other ways. A model response is one input to that discussion, not an independent medical examination.

The conversation then turns to medical uncertainty. New papers inform guidelines, and guidelines must be translated into practice; in some areas evidence is unsettled and professionals disagree. Singhal says an AI system should be able to communicate uncertainty when information is insufficient rather than manufacture certainty.

Singhal cites estimates that uptake of evidence in practice can take years and hopes AI can shorten some delays by making literature accessible at the point of care. The interval depends on the evidence, intervention and setting; his 5–17-year range should not be treated as a universal measure for every guideline or patient.

Rughani asks how products serve patients, individual clinicians and organizations. Singhal describes three layers: improve the underlying models, build usable products around them, and work with partners on deployment. Strong model performance alone will not help if people cannot use the system in their actual workflows.

He says security, privacy and compliance are important obstacles for health organizations, and that OpenAI invested in enterprise offerings to address them. Requirements vary by product and use: claims about one healthcare offering should not automatically be applied to an ordinary consumer chat or every connected service.

Singhal describes work with Epic to connect healthcare workflows and allow clinicians to use ChatGPT alongside records. He says integration aims to reduce the need to move repeatedly between disconnected systems. The details of access and availability depend on the actual product and organization.

Mental health raises especially difficult questions about what a user means and what help is appropriate. Singhal says the team works with licensed mental-health experts on safeguards and on a benchmark for evaluating behavior in difficult situations. He acknowledges disagreement among experts and describes consensus-building as part of the work.

Rughani distinguishes health information from actual medical care. A person may need a prescription, examination, laboratory work, imaging or a specialist, and the care relationship also involves empathy and professional responsibility. Singhal agrees that AI cannot provide all of that and says OpenAI expects to work with care providers and other partners.

He imagines AI helping people navigate the care journey, including scheduling and logistics, while other organizations deliver care. He does not expect this to solve every access problem in the short term or for OpenAI to manage every aspect of a patient’s treatment itself.

Asked about partnerships, Singhal highlights UCSF’s work exploring clinician chart review alongside electronic records. He also cites a reported reduction in some administrative workflow costs at AdventHealth. The cost figure is a partner-reported example, not an independent estimate of savings across healthcare systems.

The host asks how a patient should react to seeing a clinician use ChatGPT. Singhal expects attitudes to change as people gain experience, but says that transition is incomplete. The relevant question is how the tool is used, checked and integrated into clinical responsibility, rather than whether a model is simply visible on a screen.

If both patient and doctor prepare with AI, Rughani asks what remains to discuss in the appointment. Singhal emphasizes uncertainty, empathy, history-taking and physical examination. A model may lack crucial information, so the visit is also a way to gather evidence rather than merely confirm a prewritten answer.

For the time between visits, Singhal imagines reminders or flags that help people seek a professional review when something changes. He says OpenAI is focused on health information and does not intend to manage every element of care on its own. Whether any proposed intervention improves outcomes still requires evaluation.

Rughani invokes AI’s surprising discoveries in games and asks whether medicine has had a comparable moment. Singhal refers to AlphaGo’s famous move 37 against Lee Sedol as a metaphor for an unexpected insight that later proves meaningful. He says medical AI is still early and discusses the potential rather than claiming an equivalent breakthrough has already been established.

As a smaller example, he says a clinician-support study flagged potential antibiotic overuse. He hopes future systems might examine longitudinal records and identify risks that are hard for a person to notice across many years. Such predictions would need validation, appropriate data handling and clinical assessment before being acted on.

Rughani asks whether existing data might help answer questions about supplements and other topics where evidence is incomplete. Singhal sees potential in analyzing records and other data, and in helping scientists prioritize and validate ideas. The conversation does not establish that any particular supplement is safe or effective.

The discussion turns to the cost of medical intelligence and the much larger cost of care. Singhal reports that a newer, cheaper model performed better than an earlier model on an internal professional benchmark. He expects the cost of a given level of model capability to fall, but a benchmark and model price do not measure total healthcare costs or clinical outcomes.

He explicitly says AI has not yet had the macroeconomic effect of lowering overall healthcare spending. He sees potential savings in chronic-care management, documentation and administration, but says they depend on integration, deployment and real economic value—not merely on making model responses cheaper.

Rughani raises reports of AI systems used by payers and providers creating an administrative arms race. Singhal hopes that shared information and AI-assisted adjudication could eventually reduce friction. This is his proposed future, and the interview does not demonstrate that such a system would be fair or reduce spending in practice.

For healthcare entrepreneurs, Singhal recommends building around models that may improve quickly and focusing on the last mile of delivering quality care. He expects partners to connect model capabilities with services, workflows and the practical barriers patients encounter.

Procedures, tests and physical interventions remain important parts of healthcare spending. Singhal thinks better information and prevention might sometimes help people reach appropriate care earlier; he sees medical robotics as a longer-term challenge involving technology, reimbursement and integration. These are forecasts rather than established system-wide outcomes.

Looking ahead, he returns to two ambitions: make dependable information more widely accessible, and use AI to uncover insights or capabilities that were previously difficult to obtain. A proactive assistant is one possibility he highlights, though he says good deployment and trust matter alongside model evaluations.

Asked for a five- to ten-year prediction, Singhal points to AI-assisted identification of drug targets or molecules, followed by prioritization and validation. He notes that clinical-trial capacity is limited. Better ideas alone do not demonstrate that a treatment works; studies and clinical trials remain necessary to establish evidence.