Alvira Tyagi is a first-year medical student at the Icahn School of Medicine at Mount Sinai. She was part of a research team examining the limitations of ChatGPT Health in a study, which had findings published in Nature Medicine.

Most first-year medical students spend their time mastering anatomy, memorizing biochemical pathways, and adjusting to the pace of clinical training. For Alvira Tyagi, that first year coincided with an opportunity to understand the rapid transformation in how patients seek health information with AI tools.

“In January, OpenAI launched ChatGPT Health, and I was immediately curious as to how people were using it,” she says. ChatGPT Health is a service dedicated to answering health and wellness questions, with options to connect to medical records and wellness apps.

The team found that textbook emergencies were correctly triaged. However, more than half of true emergencies were under-triaged, and the service’s suicide crisis safety alerts were inconsistent and lacking. The full findings, in an article with Ms. Tyagi as the second author, were published as “ChatGPT Health performance in a structured test of triage recommendations” in Nature Medicine in February.

“I did not expect to be involved in AI-driven health care research so early as a student,” says Ms. Tyagi. “Being part of work that could directly impact patient outcomes has been incredibly meaningful.” Read on to learn how she began working at the intersection of AI and health care, and the importance for students to be familiar with this rapidly evolving field.

How did this research project get started?

It started with me shadowing Dr. Ramaswamy in the Urology Department. In-between surgeries, we talked about our interests in AI in health care, and I learned we had a robust department at Mount Sinai that focused on AI and research. We continued having conversations about AI in health care, and when OpenAI released ChatGPT Health, the discussions intensified. Immediately, we were texting about the implications of this tool, which coalesced into the idea of a study to examine it. The project started out with the two of us, but with the help of leadership from AIHH, and other physicians, we managed to find collaborators and were able to begin the study quickly.

What was it like being on the research project as a student?

At first, I was intimidated. I was a first-year student working alongside physicians with far more experience in AI and clinical medicine than I had. It took some time to realize that I didn’t need to match their background to contribute meaningfully. I brought a different perspective. I could think through how someone my age would realistically use a tool like ChatGPT Health—how we’d phrase questions, what we might take at face value, and where misunderstandings could happen.

That lens helped us step outside a purely clinical viewpoint. We knew we needed to move quickly. From its release, ChatGPT Health was already being widely used, and we felt a responsibility to evaluate it while people were actively using it. We completed the data collection within two weeks because we wanted to better understand its safety profile and identify any potential limitations as early as possible. Our goal was not to diminish the value of AI in health care, but to approach it thoughtfully by examining where it performs well and where caution may be warranted.

I have always enjoyed writing, whether for leisure or through my work with my undergraduate newspapers—so it naturally became a larger part of my role in the project. In addition to contributing to data collection, I took on significant responsibility for drafting and editing the manuscript. This involved many late nights and multiple rounds of revisions, but I valued that process. Given that the tool was already being widely used, it was essential that we communicate our findings clearly, accurately, and with appropriate nuance.

I was genuinely excited to be part of the project, especially as a student stepping into the world of AI research for the first time. At the same time, I felt a deep sense of responsibility. I wanted to contribute meaningfully alongside experienced researchers, and I was acutely aware that our findings could influence how people understand and use this technology. That awareness pushed us to be especially rigorous. We carefully crafted our prompts and clinical scenarios to be as comprehensive and realistic as possible.

Was it hard balancing school work and being on this project?

My school work always came first, and I was careful to keep that as my priority. Because of that, much of the research work happened in the evenings. It could be demanding at times, but I truly enjoyed it. Being part of a project that was unfolding in real time, and working alongside people who made the process engaging and collaborative, felt energizing rather than exhausting.

What also made this project so meaningful was that it never felt disconnected from my education. It was a different kind of learning: hands-on, fast-paced, and collaborative. There was constant progress and discussion, and that experience offered something you simply cannot replicate in a classroom. The structure of the medical education program at the Icahn School of Medicine also helped tremendously. The flexibility and autonomy built into our curriculum made it possible to take on a project like this while staying on track academically. In the end, it was demanding, but it resulted in work I am genuinely proud of.

Should students be thinking about AI more?

As medical students, we’re trained to understand clinical systems and patient care. It can be easy to view AI as something reserved for computer science experts and engineers, and that it’s separate from us and the work we do as clinicians. But that is becoming less and less true by the day. Patients now have direct access to AI technology, and many will go to doctor appointments having already used them to research symptoms or interpret medical information. At the same time, in our current health care system, patients may wait months to see a physician. In that gap, AI tools can function as a kind of interim resource—offering information, reassurance, or sometimes misinformation—before a patient ever steps into a clinic.

Because of this, it falls on us as future doctors to understand these AI health care technologies before patients come to see us. Understanding and discussing the AI-generated information a patient has already seen may soon become a routine part of taking a patient history. We cannot effectively counsel patients about tools they are using if we do not understand how those tools work, what their limitations are, and where they may fall short. As part of a generation of physicians training alongside these technologies, we have a responsibility not only to react to AI’s presence in medicine, but to engage with it thoughtfully and proactively.

What advice do you have for students who are interested in AI research?

For students who are not sure whether they can even get started, you absolutely can. You don’t need to be an engineer or have years of technical experience to contribute meaningfully. AI research, especially in health care, needs people who can think critically, ask good questions, and communicate clearly. Then, for those who aren’t sure how to get started, start having conversations—with classmates, professors, and doctors. A simple conversation in between patient cases is what transformed my shadowing experience at the Urology Department into this research project. There are so many talented scientists and faculty at Mount Sinai, and simply engaging with them by asking questions, sharing your interests, and expressing curiosity, can open doors.

Sometimes all it takes is one thoughtful conversation to set something much larger in motion. Being open to opportunities and willing to learn really makes a difference. I had never done AI research before this project, so stepping into it required me to get comfortable with not knowing everything. But I came to understand that AI is developing so quickly that no one has it completely figured out. Even people with years of experience are still asking questions and adjusting as the field evolves. That realization made it feel less about being an expert and more about being engaged. You don’t have to start with deep technical knowledge; you just have to be willing to listen, learn, and contribute where you can. In a space that’s changing this fast, humility and curiosity go a long way.

Isaac Kohane, MD, PhD, Chair of the Department of Biomedical Informatics, Harvard Medical School, was keynote speaker of the Icahn School of Medicine at Mount Sinai’s Windreich Department of AI and Human Health (AIHH) December 2025 session of AIHH Grand Rounds.

Health care systems across the country have been increasingly using artificial intelligence (AI) systems to assist and augment what clinicians and researchers can achieve. As adoption of machine learning accelerates, thought leaders have been scrutinizing how AI is being embraced.

“Many doctors are already using these tools, such as OpenEvidence, but without visibility or oversight by health care systems,” says Isaac Kohane, MD, PhD, Chair of the Department of Biomedical Informatics, Harvard Medical School. OpenEvidence is an AI-powered clinical decision support and medical search engine.

Dr. Kohane is a prominent researcher in biomedical informatics and AI whose nearly 400 papers have been cited more than 95,000 times, according to Google Scholar. He was the keynote speaker of the Icahn School of Medicine at Mount Sinai’s Windreich Department of AI and Human Health (AIHH) December 2025 session of AIHH Grand Rounds. Dr. Kohane wants to see not just more use of AI, but more responsible use—a theme of his lecture, which was titled “A Tipping Point for Clinicians’ Influence Upon AI-Driven Clinical Decisions.”

“I chose these topics for Grand Rounds because I view the Icahn School of Medicine and its leadership as among the most forward-looking in the country,” says Dr. Kohane, “and therefore they should be truly focused on setting an example in terms of accelerating adoption options that are both safe, and also enabling patients and clinicians to benefit from the complementarity of AI to human expertise, as well as changing the promotion process to reflect greater engagement with reproducibility and robust research.”

The AIHH Grand Rounds is a monthly seminar series that showcases developments in how AI, science, and medicine intersect, and features an open discussion to foster collaboration. The inaugural session launched in September 2025.

How should health systems think about engaging with AI as it pertains to patients, clinicians, and researchers in a way that is beneficial to all parties? Dr. Kohane discussed the following themes during the seminar.

Transforming the institution with AI

By their nature, large health care systems in the United States are high-revenue, low-margin businesses, and because of that, they face challenges in moving rapidly with change to avoid disruptions.

Institutionally, AI adoption has found more comfort and scalability on the administrative side of operations, including reimbursement and corporate functions. AI is a critical lever, but not a priority for health care system spending presently, according to Dr. Kohane.

However, the application of AI on the clinical side, including continuity of care, clinical operations, and quality and safety, remains nascent or in pilot stages.

“It’s actually the doctors who are leading [with AI adoption], even when their own institutions are not supporting them directly,” says Dr. Kohane.

That landscape is slowly changing as health care leaders begin to engage their clinicians with AI support where it is needed now, but it should not be at the cost of extended, effortful multi-year governance conversations, Dr. Kohane pointed out. The incentives for using AI in the practice of medicine must be focused on improving care rather than maximizing revenue.

“And so, I anticipate that the future first adoptions will happen in specialized high-end services like concierge services, primary care, or cancer care,” he says. “But eventually, it would become a requisite for the safe practice of medicine, and for meeting the expectation of our patients, that ultimately our health care systems will be propelled into more significant engagement [with AI].”

Transforming publishing and literature review with AI

“Every part of the scientific publication process—that is, the generation of manuscripts and review of manuscripts—is going to be augmented by AI,” says Dr. Kohane. “That is going to present, or is already presenting, challenges that the whole peer-review publishing industry is not well equipped to handle.”

Dr. Kohane discussed a case study in which he created a hypothesis that was incorrect, and with AI tools was ultimately able to generate data that were not only fictional, but designed by AI to avoid detection by the majority of fake-data-generation detectors.

“We’re going to really have to address, first and foremost, the incentives that drive perverse behaviors,” he says. An industry that prizes publication volume, or publishing in high-profile publications over producing work with actual scientific impact—such as important but unglamorous replication studies—is only going to drive bad actors.

In the right hands, AI will increase the efficiency and quality of scholarly scientific review. AI can serve as a prism that allows clinical and laboratory experience to be distilled into new knowledge, forming a substrate for truly lifelong medical education. “However, we have to reset the culture and incentives,” Dr. Kohane says.

Transforming AI with human values

In an industry where urgency and time matters, AI presents a strong value proposition with its capability to process large datasets and execute large volumes of actions in a blink of an eye. Time-consuming tasks can be automated by AI, but when decisions that pertain to the care of individuals with unique needs are left to a normative model that adheres to overarching policies, the individual’s needs might not be met.

The solution is not to turn away from AI, but to develop personal models that account for the needs of not just the patient at hand, but also their caretakers, doctors, or any other relevant stakeholders, says Dr. Kohane. It is about building human values within an AI model, which can flag when an individual case does not align with the normative model.

That work to develop such projects falls on the health care system, says Dr. Kohane. He introduced the Human Values Project, an international initiative led by Harvard Medical School’s Department of Biomedical Informatics, which aims to characterize how AI models respond to ethical dilemmas in medicine, measuring both their default behaviors and their capacity for alignment. And he proposed that researchers at the Icahn School of Medicine have that potential to develop their own human values-based AI models.

“My takeaway from presenting and participating in the AIHH Grand Rounds really stemmed not from the presentation itself, but from discussions I had afterwards with various leaders of the AI efforts,” says Dr. Kohane. “My sense was that more than most institutions, [Mount Sinai’s] leadership was willing to invest and take a chance on pilots of deployments of these technologies to learn fast and adapt fast. And at the same time, everybody recognized that this is very challenging, given our current regulatory environments and incentives.”

Dr. Kohane ended his presentation with a line of wisdom for participants to consider: “There is no one to lead this in the direction we want, other than us.”