Should You Trust ChatGPT With Your Health?｜a16z — BigGo Finance

Karan Singhal, who leads health AI research at OpenAI, says the company's goal is to make "trustworthy, abundant, reliable health intelligence" free or almost free. He argues that the best approach is to use a powerful general AI model with healthcare information included at every stage of training, rather than build a separate model just for medicine. The conversation focuses mainly on how OpenAI is developing its healthcare tools: how many people use them, how the company tests whether they can be trusted, products for doctors and healthcare organizations, and the limits the company sets to avoid acting as a medical provider. The question in the title — whether people should trust ChatGPT with their health — comes up in the discussion of trust and safety, but it is only one of several topics.

Singhal speaks with Jay Rughani, a partner at a16z who invests in healthcare and biotechnology. They discuss OpenAI's healthcare strategy, a story about podcaster Nathan Labenz's son, the HealthBench and mental health tests, the connection with Epic, and Singhal's predictions about falling costs and AI helping discover new drugs. The discussion gives specific numbers from OpenAI, explains the difference between sharing health information and providing medical care, and describes what Singhal says the company will not do.

OpenAI says 300M people a week ask ChatGPT health questions, mostly for free

Singhal describes health as one of three ways OpenAI works toward its mission: building and using artificial general intelligence (AGI), preventing safety risks, and "making upside happen," meaning helping people benefit from AI. He sees healthcare as a clear example of that third goal. His main figure is that more than 300 million people use ChatGPT every week to ask questions about health and wellness, and he says that number is growing quickly. He says most of these users have free accounts, and most live outside the United States. Elsewhere in the conversation, he says ChatGPT has 1.2 billion users each week overall. Based on those figures, health-related use makes up about a quarter of the total.

The need for these tools comes partly from problems with the existing healthcare system. Rughani says more than 100 million people do not have access to a regular primary care doctor, and that people may wait more than 40 days for an appointment. Singhal adds that even in the San Francisco Bay Area, people may wait weeks or months to see a specialist. They may then get only 15 minutes with the doctor and leave with questions still unanswered. Singhal says AI is especially useful between appointments. It can help people find their way through the healthcare system, prepare questions, understand what a doctor said, and decide what to do next.

Claim	Figure (as stated by Singhal)
Weekly ChatGPT users asking health/wellness questions	300M+
Share of those who are free users	"Vast majority"
Share of those outside the US	"Most"
Total weekly ChatGPT users	1.2B
US healthcare spend	$5.3T+ (Rughani), third-largest GDP if a country
US healthcare employment	23 million (Rughani)
People without a regular primary care doctor	100M+ (Rughani)

Specialized training inside a general model beats a standalone medical model

Singhal worked on specialized AI models in an earlier job. He now argues that choosing between a general model and a specialized medical model is a false choice. He says that, over the past year, OpenAI has added specialized healthcare data at every stage of training for every model released in ChatGPT. He calls this "the best of both worlds": the model can use advanced reasoning skills while also learning from medical training materials provided by doctors. In his view, a smaller model built only for healthcare quickly falls behind because general AI models improve so fast.

This means the same model can write code, plan vacations, and answer medical questions, while adjusting how it responds to different users. Singhal says HealthBench, released in May 2025, tested whether the model could change its use of technical language depending on whether it was speaking to an ordinary person or a healthcare professional. He says OpenAI has since improved both how it measures this ability and how it trains its models.

Trust is built on a global physician network, not a few in-house doctors

Singhal makes a clear point about how to test medical AI: a small group of doctors should not be the only people deciding what a correct answer looks like. As he puts it, "If you ask five physicians any question you'll get like eight different answers," meaning doctors can disagree even when answering the same question. OpenAI has therefore built a network of hundreds of experts across many medical specialties, languages, and countries. The company combines their judgments to create data for testing and training its models, paying special attention to questions where experts disagree.

Singhal says safety must come before telling users what they want to hear. OpenAI tests whether its models tell people to seek medical care even when those people would rather hear that everything is fine or want to try another treatment. He says the models' results on these tests have improved "pretty dramatically." For political or genuinely disputed topics, such as unapproved peptides and vaccines, the stated policy is to follow standards set by experts rather than take independent positions. When science itself is uncertain, the models are trained to recognize and explain that uncertainty after searching for information. HealthBench also publicly measures how well the models express uncertainty.

"Health is a place where it moves at the speed of trust. And so you need to be able to not only make the models better and measure them, but also kind of communicate and bring the world along."

Singhal also points to a major problem with the healthcare system: medical guidelines can take about five to 17 years to become part of everyday medical practice, depending on the study. He says ChatGPT for Healthcare and ChatGPT for Clinicians aim to shorten that delay to weeks, days, or even hours by giving doctors access to current research during a 15-minute appointment.

OpenAI wants to be the front door and the quarterback, not the doctor

Singhal draws a clear line between giving people medical information and providing medical care. He does not expect OpenAI to provide every kind of care and says directly that "there's no future where any one company is delivering all the medical care." He sees OpenAI as the front door and the "quarterback": helping people find the right care, manage practical details, schedule appointments, and recognize when they need medical attention. Partners would provide the actual care. He explicitly says OpenAI will not be "in the business of managing all these things by ourselves."

The idea of AI agents shows what this could look like in practice. Rughani describes a growing group of personal agents, including Instinct, Muse, Grokbot, and OpenAI's DOTS. These tools can take a goal and carry out several steps using a person's apps and payment details. In healthcare, they could book appointments, check insurance coverage, guide patients through care after leaving the hospital, or challenge billing mistakes. Singhal is excited about this possibility. He points out that much of healthcare still uses older electronic health record (EHR) software running on virtual machines, which he says is "absolutely right for agents and computer use." He predicts that many barriers between patients, doctors, and health information "will dissolve in the coming months."

However, Singhal expects human care to remain important. Taking a patient's medical history, examining them physically, showing empathy, and providing a second opinion from a person will still matter. He also notes that AI cannot know everything about a patient before an appointment, because the visit itself provides new information. When asked what patients and doctors will discuss after both have used AI to prepare, he answers: "the human aspect of healthcare."

Epic integration and named health-system partners carry the enterprise motion

Singhal explains OpenAI's approach to working with healthcare organizations in three steps: make the models good, build tools that make them easy to start using, and then work with partners to put them into practice. He says work on compliance, including HIPAA requirements and privacy protections for healthcare data, took about two years and ended roughly a year before the episode. This work was intended to remove security problems that had stopped healthcare organizations from adopting the technology.

The most specific partnership example involves Epic. OpenAI announced a way for healthcare providers to use Epic through ChatGPT, as well as a ChatGPT sidebar built into the Epic system that doctors already use. Singhal names two healthcare organizations:

• UCSF — one of the early users of the connection between Epic, electronic health records, and ChatGPT. It uses the system to review patient records and reduce the need to switch between tabs while providing care.
• AdventHealth — a major user that is putting AI agents to work across its digital workspace. Singhal says the organization reported a reduction of about 80% in the cost of administrative workflows.

He also mentions a study with Panda Health that tested an AI assistant for doctors as an extra layer of safety. The system could spot potential problems, such as interactions between medications. One pattern they observed was that doctors sometimes used antibiotics more than necessary. The AI alerts encouraged them to consider the risk of antibiotic resistance, rather than focusing only on the individual patient's situation.

Mental health gets its own benchmark, built on licensed experts

Singhal says the same basic safety standards apply across OpenAI products, including DOTS. The company's work on mental health also aims to measure the positive benefits of AI, not just whether it can avoid making a crisis worse. Its approach is similar to HealthBench: work with licensed mental health experts around the world, agree on the right way for a model to respond in difficult situations, and publish a shared test to measure its performance. Singhal notes that mental health experts disagree even more than doctors in physical healthcare. That makes reaching agreement especially important. The mental health benchmark was released shortly before this episode.

Cost per unit of health intelligence is falling fast, but macro savings have not arrived

Singhal supports his argument about falling costs with one comparison: GPT-6 Luna, which he describes as the cheapest advanced model available through the API, scored better on HealthBench Professional than GPT-5.6 Sol, which he calls the best model "a month ago." He says GPT-6 Luna costs 34 times less. He uses this example to argue that AI is becoming cheaper at a given level of ability faster than earlier technologies did. He expects the cost of producing the same level of health intelligence to "plummet," while the standard for earning people's trust rises over time. The goal, he says, is to make this technology widely available rather than keep it to themselves: "We don't wanna hoard that health intelligence, we want everybody to use it."

He is careful when discussing the possibility of saving money across the $5.3 trillion US healthcare system. In his view, health intelligence has not yet made a measurable difference to overall US healthcare spending. Areas where AI could save money include managing long-term illnesses, writing medical notes, handling administration, and carrying out back-office tasks. But he says these savings will require connecting AI to existing systems, changing how work gets done, and using the tools widely. Cheaper models alone will not be enough.

Rughani raises the reported financial battle between healthcare providers using AI and insurance companies using AI to challenge their bills. He refers to a New York Times article and an analysis by the Blue Cross Blue Shield Association that put the extra spending at about $1 billion. Singhal calls this a short-term problem that should not exist. He suggests using AI to record an accurate account of each medical visit, sharing that record between providers and insurers, and having an AI system decide billing disputes. A limited appeals process would allow people to challenge those decisions. He says this could avoid the need for both sides to keep competing against each other with AI.

The next five to ten years hinge on proactive prevention and AI-accelerated drug discovery

Singhal uses the idea of a "move 37" to explain what he hopes AI will contribute to medicine. The phrase refers to a famous AlphaGo game against Lee Sedol, when a move that experts first thought was a mistake turned out to be brilliant. In healthcare, Singhal expects AI to find important patterns that people might miss. One example is models combining ten years of information about a patient, including electronic health records, insurance claims, and chats, to predict health risks that no person could keep track of mentally. He also says AI agents could search through information already stored in electronic health records and insurance systems but not yet being used, and find new medical insights.

Looking five to ten years ahead, he identifies two main areas: using AI to suggest new drug targets and molecules, and using AI to decide which candidates should be tested first and help confirm whether they work. The second area matters because there are more possible drug molecules than there are available places in clinical trials. AI could help researchers choose better candidates and run trials faster. For medical procedures, he expects AI to help identify problems earlier so patients can avoid costly procedures later. He sees the role of robots in these procedures as a longer-term question because both the technology and the way healthcare pays for it still need to develop. He also highlights a third group that could benefit from AI: scientists. He expects AI to help solve problems with testing and confirming discoveries, rather than simply coming up with ideas.

His advice to healthcare founders is straightforward: build products using AI models that will become more capable, and prepare for a future in which AI agents and computer use change how healthcare technology works. "Bet on the models getting better," he says. He presents this as the single most important decision founders can make.

The main unresolved issue in the conversation is the gap between OpenAI's reach and the limits it says it will respect. The company has a large audience, with 300 million people using ChatGPT for health questions each week, as well as falling costs and systems for testing its models. Yet Singhal repeatedly says OpenAI will not provide medical care itself, decide disputed medical questions independently, or manage long-term illnesses on its own. That means the potential savings in a $5.3 trillion healthcare system depend on partners and changes to everyday work that OpenAI cannot control. They also depend on reaching a level of trust at which both patients and doctors feel safer with AI involved, a level Singhal suggests is getting closer but has not yet been reached. HealthBench, HealthBench Professional, and the mental health benchmark are the tools OpenAI uses to make that level of trust measurable by outside groups. Whether the healthcare industry adopts these tests as shared standards is an important question to watch.