OpenAI Health Lead on AI in Healthcare: General Models Plus Specialized Training as a Safety Net Beside Doctors

Written by Shuzhi Benliu | Author: Tian Yu | Editor: Shen Xiao

"If you want to see a specialist, sometimes you have to wait weeks, or even months. You may have to spend a lot of time figuring out how to get an appointment and how to arrange your schedule. After all that trouble, in the end you may only get to talk to the doctor for 15 minutes; you feel you haven't prepared well enough, and only as you walk out of the exam room do you remember: 'Oh no, I still have 6 questions I wanted to ask.'"

In an interview, Karan Singhal, head of OpenAI's health division, pointed out this widespread dilemma in medical services today. Even in the United States, where medical resources are well developed, patients still face all these problems. Moreover, medical needs do not disappear when a single outpatient visit ends. Between two visits, patients also need to understand test results and decide what to do next.

Karan noted that more than 300 million people use ChatGPT every week to ask health-related questions.

For OpenAI, this is a huge demand, and even more a responsibility. AI should not stop at asking questions and giving answers; it should gradually become a continuous health assistant, and even help doctors identify potential risks and connect to medical services, helping people detect health changes earlier in the future.

There is still a long way to go to realize this vision. How to build trust among all parties, integrate into real medical workflows, and prove that these technologies truly improve health outcomes are questions we must face.

The core of the interview is that OpenAI cannot replace doctors; it only hopes to make trustworthy medical intelligence as widespread as infrastructure, while raising both the floor and the ceiling of human health. The so-called floor is making medical intelligence available for free or at low cost. The so-called ceiling is achieving things that were previously impossible.

Karan mentioned a real case that left a deep impression on him: the son of podcast host Nathan Labenz had symptoms such as persistent fatigue and irritability, and with the help of AI, Nathan gradually realized that his child's condition might be more serious than he had imagined.

Later, the child was diagnosed with an aggressive form of leukemia. Throughout the medical process, he often used GPT-5 Pro to help clarify the diagnosis and treatment process and understand lab reports; sometimes the AI provided help before the doctor had time to explain.

This case showed Karan the important value of AI in the medical process: the health support patients need runs through the entire medical journey, especially between two visits. People ask ChatGPT about health concerns, understand test results, prepare questions for appointments, and also review doctors' instructions. Karan believes that AI can help people better handle all these steps and extend health support beyond the clinic.

The host asked: Why did OpenAI choose to use general models as the foundation for medical intelligence?

Karan has previously participated in the development of specialized models. In his view, the advantage of general models is that they can continuously benefit from advances in frontier models in scale, intelligence, and reasoning ability, while specialized training in the medical field can further strengthen the model's ability to handle health issues. "What we are doing is actually specialized training within general models. In my view, this precisely combines the strengths of both."

Karan explained that over the past year, OpenAI has added specialized medical data at every stage of ChatGPT model training, including data obtained through collaboration with doctors around the world. If a smaller specialized model aimed only at the medical field were trained separately, it would quickly fall behind the pace of development of frontier models. OpenAI chose to combine specialized training with the continuous iteration of general models, while integrating evaluation and training data into the model improvement process.

Following this approach, OpenAI is exploring a further product form: a proactive personal health assistant. "Make health services increasingly personalized and increasingly proactive."

Karan said he looks forward to the next step for the personal health assistant: helping users handle the tedious and fragmented tasks in the medical process, so that AI appears throughout the entire process before, during, and after visits.

OpenAI is also exploring ways for users to connect their own health information so that ChatGPT can provide more targeted support based on personal circumstances. Privacy protection is also important; Karan said this health information will not be used for model training or advertising targeting.

Karan emphasized that the personal health assistant cannot replace doctors and cannot independently provide all aspects of medical services. OpenAI hopes to help users better access health information and navigate the medical process, in coordination with the existing healthcare system.

Karan is very interested in this. He described this capability as a "safety net" around patients and clinicians.

"I really like this idea: whether for patients or clinicians, it exists throughout the entire medical process, but it does not intervene all the time, nor does it become a middleman standing in the way; what it really does is enhance and improve the existing experience."

Karan explained that the team has worked with Penda Health on a series of experiments to study the application of a Clinician Copilot system. Such tools serve as reminders during work, helping doctors pay attention to information they might otherwise miss and adding a layer of verification to clinical judgment.

In the collaborative experiments with Penda Health, the AI noticed a tendency toward overuse of antibiotics in some clinical scenarios, involving not only the treatment of the patient at hand but also the broader public health risk of antibiotic resistance. This makes the role of AI more concrete: helping doctors notice patterns that are easily overlooked in daily practice.

Karan mentioned a very real challenge in medical judgment: different doctors can have different opinions on the same question. "If you ask five doctors the same question, you might get eight different answers."

So OpenAI built a professional network involving hundreds of doctors worldwide, covering dozens of medical specialties, different languages, and clinical practice in different countries. The team refines training data and evaluation standards through feedback from different experts, with particular attention to cases where medical judgment is divided.

Karan summed up the key point in one sentence: "In healthcare, everything has to move forward at the speed at which trust is built."

From the discovery of a new research finding to its application in routine diagnosis and treatment, this process often takes at least 5 to 17 years. OpenAI hopes to shorten the information-access step in this process. When seeing patients, doctors can more conveniently look up the latest medical literature to inform diagnostic and treatment decisions. For an outpatient visit that lasts only a few minutes, finding relevant research evidence faster will efficiently reduce the time spent searching for materials.

For these capabilities to truly enter routine diagnosis and treatment, the problem of fragmented information within hospitals needs to be solved. Medical records, test results, and treatment notes are scattered across different systems, and OpenAI is working with medical information systems such as Epic to explore letting healthcare staff use ChatGPT directly in their existing work environment.

Karan answered: "I hope they will talk more about the human side of medicine."

He then talked about uncertainty in medicine, doctors' empathy, and the importance of bedside care and face-to-face communication. This is another meaning of AI as a doctor's "safety net": helping doctors prepare more thoroughly for diagnosis and treatment, verify information, and obtain medical evidence, so that doctors have more energy to focus on the patients themselves.

Recently, a netizen surnamed Liu in Guizhou was hospitalized for abdominal pain, but after completing a full set of examinations in the gastroenterology department, the cause still could not be found. Sensing that something was wrong, he contacted his ex-wife and asked her to help him make a will in case of the worst.

Afterward, his ex-wife consulted Ant Afu, and Afu indicated that the symptoms were highly suggestive of fatal pulmonary embolism. They immediately took the suggestion to communicate with the hospital, and the hospital arranged surgery right away, and Liu was out of danger.

While responding to the incident, Ant Afu also released data: among daily user consultations, Afu issues roughly more than 200,000 recommendations to seek medical care immediately each day, of which more than 50,000 are related to critical illness.

Another netizen shared: "Although many people complain about Doubao, she really saved my dad's life. That night he suddenly had a numb tongue and could not move his legs. I thought it was nothing serious and planned to go to the hospital the next morning. But I searched on Doubao, and it was a cerebral infarction! We went to the hospital overnight and quickly received thrombolytic treatment, preserving the golden treatment window. I am still grateful to Doubao."

Ant Afu has already turned part of Karan's vision for a personal health assistant into reality, including services such as appointment registration, online consultations, and cloud accompaniment for medical visits. In Guangzhou, for example, "Sui Xiaoyi" within the Ant Afu app has launched an "ask and act immediately" registration service.

Doctors have also begun using AI in their work. In January this year, Life Times published a survey involving more than 500 doctors from top-tier hospitals across the country.

When patients bring AI answers to see a doctor, if the recommendations are not applicable to their situation, the doctor needs to spend time explaining why and verifying information. If the patient remains convinced that AI is right, the two sides must first resolve the disagreement before discussing the actual condition. A tool originally meant to help patients understand health information instead increases the difficulty of doctor-patient communication.

In 2026, the Fudan University Research Center for AI for Good and Digital Intelligent Elderly Care, the Department of General Medicine at Huashan Hospital Affiliated to Fudan University, and Shangguan Think Tank jointly conducted a two-sided survey, collecting a total of 75 doctor questionnaires and 143 questionnaires from middle-aged and older residents.

The survey found that about 70% of the doctors surveyed had encountered patients using AI to look up health information in outpatient settings. Among the risks of AI health advice that doctors worried about most, 81.33% mentioned a lack of individualization, 73.33% were concerned that AI might increase patients' misunderstanding or anxiety, and 54.67% were concerned that it would increase doctors' burden of explanation.

Reported on October 27, the parents of Amazon software engineer Wan Bowen sued OpenAI. Wan Bowen had epilepsy and died in February this year. The lawsuit alleges that when she used GPT-4o during her lifetime, the model guided her to gradually reduce the dosage of anti-epileptic drugs and advised her to conceal the related adjustment from her doctor. The family believes that ChatGPT's relevant behavior was related to their daughter's death.

GPT-4o had previously been criticized for issues such as overly catering to users in its answers, and OpenAI removed it from ChatGPT on February 13 this year, exactly the day after Wan Bowen's death.

OpenAI hopes to improve AI's ability to handle complex medical problems through the participation of doctors worldwide in model training and evaluation. When the answer a patient expects conflicts with medical advice, the model should also prioritize safety; if the available information is insufficient to support a clear judgment, it should state the uncertainty.

When AI becomes an important entry point for people to obtain health information, building trust requires every party to clearly understand its own role: patients know how to refer to AI's suggestions, doctors are able to verify the relevant information, and AI can also guide users to seek medical care promptly when risks are detected. This is the medical safety net that Karan envisions.

In the first half of the interview, Karan mostly talked about how AI helps patients and doctors cope with immediate medical problems. As for the future, he has two longer-term expectations: enabling more people to obtain medical intelligence at lower cost, and using AI to advance medical research and discover new treatments.

First is the cost of medical care.

The host mentioned that annual medical spending in the United States has exceeded $5.3 trillion, and the growth rate of medical spending far exceeds inflation. How to enable more people to obtain high-quality medical services has always been a difficult problem.

Karan believes that model capabilities are continuously improving, and the cost of obtaining medical intelligence will also continue to decline. "We hope that medical intelligence will eventually become as widespread as a basic resource."

However, this does not mean that spending across the entire medical system will immediately decline. A large portion of medical costs comes from work such as administrative affairs, paperwork processing, and chronic disease management.

Karan's other expectation for AI comes from medical research itself.

The host cited the famous AlphaGo case: in a Go match, AI played the "Move 37" that surprised human experts. At the time, all the human experts watching the match thought this move was very foolish. But in hindsight, it was a highly insightful masterstroke. Humanity has only now begun to understand and learn the patterns contained in it.

Karan believes that we are currently at a very early stage, "but I think we will see more cases in the future: a model receives a user's ten-year longitudinal health record, integrates all information such as electronic medical record data and conversation data, and then makes health and risk predictions that previously seemed impossible. Such an achievement would qualify as a 'Move 37' in the medical field."

This direction has already begun to be reflected in OpenAI's life sciences layout. In April this year, OpenAI launched GPT-Rosalind; in September, it introduced a study using AI to find new antibacterial molecules. The team of bioengineer César de la Fuente used deep learning models to screen candidate molecules, and used ChatGPT and Codex to assist with research hypotheses, code writing, and data analysis.

On October 20, OpenAI announced the acquisition of Orchestra Bio. This company focuses on helping biopharmaceutical teams coordinate the drug development process. OpenAI stated that it will combine Orchestra's team and technology with GPT-Rosalind and related life sciences tools to help researchers advance work from early discovery to drug development.

For the next 5 to 10 years, Karan expects that the important changes brought by AI will be reflected in the development of new therapies. He mentioned two directions: first, using AI to discover new drug targets or candidate molecules; second, helping researchers screen and validate more promising treatment plans, making clinical trials run more smoothly and efficiently.

Karan pointed out that the bottleneck in new drug R&D will gradually shift from discovering candidate molecules to screening and validation. As more and more candidate drugs enter clinical trials, AI can help researchers determine which plans are more worth advancing, improving the efficiency of using experimental and clinical trial resources.

From reducing the cost of medical services to helping scientists discover new treatment directions, Karan's expectations for health AI have gone beyond answering medical questions themselves.

When talking about possible future medical breakthroughs, he concluded: "My judgment is that in the next 5 to 10 years, we are expected to see better drugs and more personalized drug treatments. And all of this will be built on the two efforts we are investing in today."