OpenAI's Karan Singhal on Trusting ChatGPT With Health Questions

Singhal says more than 300 million people weekly use ChatGPT for health and wellness types of questions. He says OpenAI takes that as an incredibly important responsibility to get right, and says that usage has been growing rapidly over time.

He describes the motivation in terms of the difficulty of getting care. Even in the Bay Area, he says, a specialist appointment can take weeks or months, and a visit may last only 15 minutes, leaving patients with unanswered questions. He hopes AI can be part of the care journey between visits, which he says is when AI helps people most on the consumer side.

Singhal recounts a story he finds moving: Nathan Levent, who hosts the Cognitive Revolution podcast, described a long process with his very young son, who suffered a lot of fatigue and irritability. AI helped the family realize, while they were on vacation, that the situation was serious, and encouraged them to seek testing and to take it seriously even when doctors were less sure and seemed to downplay it.

According to Singhal, the son was diagnosed with an aggressive form of leukemia. Levent reportedly used GPT-5 Pro throughout his care journey for navigating the process and understanding lab reports, including before a specialist had time to help interpret them. Singhal says this brought the family a lot of peace of mind, and notes that Levent tells the story in his own podcast.

Asked about a future in which information reaches the user at the right time to nudge better health behaviors or adherence to a doctor's protocol, Singhal says OpenAI is very excited about a proactive personal health assistant. He is clear that he does not see it as any replacement for having a doctor or access to one.

He describes the aim as making health information more personal to a user's context. OpenAI's consumer offering lets users connect their health information, which he says is used securely and not for training data or ad targeting. Most of the roughly 300 million weekly health users are on free plans and most are outside the US, which he says shapes how the company thinks about distribution.

Singhal describes a series of experiments OpenAI ran with Penda Health to study a clinician co-pilot. The tool was deployed as a safety net for clinicians during their work: it proactively suggested that something might be off, such as a medication interaction the clinician might not be tracking, and could help adjust the plan.

He says he expects the same pattern on the patient side. He gives the example of a diabetic patient whose A1C rises, and says a proactive agent could make sure the medications get reviewed by a doctor and that the change is flagged for appropriate review. He adds that he wants AI to augment the existing care experience rather than always interfering or acting as the middleman.

Singhal says OpenAI evaluates and trains its models by working with experts around the world and collecting feedback in various ways, so that behavior is tailored to whether the person is a lay user, an experienced lay user navigating a complex care journey, a clinician, or another user persona.

He says the original HealthBench evaluation, released in May 2025, included a part focused on how the model adjusts its level of technical jargon for a lay user versus a clinician. Since then, he says, the company has improved at measuring whether models can cater to different users and at training them toward that.

Singhal identifies disagreement as a second component of trust. A person may share all their symptoms while hoping to hear that everything is fine. From a consumer engagement view they may want reassurance, but medically they may need to hear something very different.

He says relying on a few physicians is not the best approach, because asking five physicians a question can produce eight different answers. OpenAI therefore built a global physician network of hundreds of experts across dozens of specialties, languages and countries of practice, to aggregate expert wisdom rather than depend on one or two people in areas of disagreement. When the physicians' view points one way and the user wants comfort, he says the company prioritizes safety, and evaluations track situations where the model should say plainly that something is unsafe and the user should seek care, even if the user does not want to hear it.

Asked how a product should respond to gray-area topics such as unapproved peptides or vaccines, which are often politicized in the media, Singhal says that wherever there is debate or no consensus, whether or not it is politicized, OpenAI relies on experts to define the standards for how the model should respond. He says that is the approach the company takes across the board.

Singhal says a model can be very good, but if the activation energy to use it for improving health is too high, people will not do it. He says OpenAI's product efforts focus on reducing that barrier for patients, clinicians and enterprises, pointing to ChatGPT Health, ChatGPT for Healthcare on the enterprise side, and ChatGPT for Clinicians.

For enterprise offerings, he says the company spent about two years on security and compliance, including offering HIPAA compliance and privacy protections so healthcare organizations can handle the data they need and trust the service. He adds that partnerships and deployment form a third part of the strategy.

Asked how a product should interpret a vague request such as wanting to feel better over the next few weeks, Singhal says OpenAI is committed to helping users who are having a difficult time or are in crisis, regardless of the surface. He says the company applies the same safeguards it uses across its products, including the mental health product he refers to, as a baseline safety floor.

He says the mental health benchmark goes beyond measuring that floor and asks how models can proactively help raise the ceiling, providing positive mental health experiences. The benchmark was built with licensed mental health experts around the world on tricky scenarios. Singhal says people disagree even more here than on physical health, and the aim was to reach enough consensus to measure performance. The benchmark was released very recently, and like HealthBench and HealthBench Professional is meant as a shared yardstick for the industry.

Singhal says he does not think AI can or will provide all medical care, and draws a distinction between health information and medical care. He says no single company will deliver all of it, so OpenAI's role is to enable the whole ecosystem and to make the last mile of care easier for partners, so that the last mile of impact on patients looks like the last 100 meters for them.

He says OpenAI has a relatively unique front door to health. It can help people navigate questions and journeys, and as personal agents improve, may help with scheduling and logistics, acting as a quarterback for a person's health. He does not think this will resolve all the challenges in the short term, and says the company is mainly looking to enable partners that deliver care, most of which use its models in various ways, to make operations smoother so patients are not left in a fragmented experience.

Asked how people would react if they saw ChatGPT open on their doctor's screen, Singhal says he hears differing things and that attitudes are changing over time, with the window of acceptable views shifting in real time. He suggests there may once have been concern that clinicians would push back on patients' AI use and patients on clinicians' AI use, but he thinks both parties find AI incredibly useful for their own personal purposes.

He says that in practice, both parties may feel safer with AI in their care journey, whether it is the clinician using it or the patient using it as a sanity check. He says the attitude is definitely changing but he would not say it is fully there yet, and that this is his prediction of how things will go.

Singhal describes how models taking in a user's longitudinal history over ten years, across their records, wearable data and chat data, could make predictions about health and risks that seemed impossible before, because a human cannot integrate all of that data in their head.

He notes that consumers often ask questions whose answers have changed over time, such as whether smoking, coffee or sugar is bad for you, and newer questions such as creatine or melatonin, where evidence may be limited or not widely agreed upon. He says one of the most interesting things about agents is that they can solve bottlenecks in analyzing data that already exists across the healthcare ecosystem, including EHR data, longitudinal histories and claims.

Singhal points to healthcare spending of more than $5.3 trillion, which he says is the biggest part of the economy and the biggest employer, with 23 million Americans working in healthcare. He says making models better at the frontier and making them widely accessible is one of the most important ways OpenAI can pursue its mission, and that this is how it has deployed health intelligence to many people for free.

He cites GPT-5.4 Luna, which he says is OpenAI's cheapest frontier model on the API. In a test on HealthBench Professional, a benchmark of clinician workflows, he says it performed better than GPT-5.6 Soul, which he calls the company's best model about a month earlier, despite being 34 times cheaper.

Singhal says he does not think health intelligence has yet had a macro effect on the cost problem he was asked about. He says there are areas such as chronic care management, documentation, administration and back-office workflows where a lot of money is spent in ways that probably do not need to be spent that way. He says he is very optimistic about an overall compression of spending over coming years, but does not think it will come just from making health intelligence abundant.

He says it will come from integration, workflows, deployment and scaling, and from making sure the models are valuable for economically valuable tasks. He calls abundant intelligence a prerequisite for that future rather than the whole answer.

Asked what specialist applications should build on top of the stack assuming models become far more capable, Singhal says OpenAI thinks of much of its health impact as coming through partners and the last mile of delivering care, rather than only through its own products.

He says people in healthcare sit on a spectrum from betting against the models to betting neutrally or with them, and that the most important advice is to bet on models getting better. He suggests building toward a future in which agents, computer use and other frontier technology accelerate healthcare. He also expects medical care to become more accessible to patients with AI, including monitoring the care journey.

Singhal says OpenAI is focused on two high-level goals. The first is raising the floor of human health by making trustworthy, reliable health intelligence abundant for all clinicians, users and patients, to the point where people feel safer with AI in the care journey. He says that depends not only on good models and evaluations but also on deployment, activation energy and products.

The second is raising the ceiling, which he describes as a proactive personal health assistant that helps navigate care, schedules appointments and flags places where a person can intervene today to prevent costly deterioration later. He also points to unlocking the value of existing data, often stored in various places, to produce new insights that benefit patients.