Paging ‘Dr. ChatGPT’
By Amanda Rossillo, PhD | Illustration by Davide Bonazzi | Portraits by Jörg Meyer
In November 2022, our world began to experience a transformative technological advancement. It was marked by the launch of a generative artificial intelligence (AI) chatbot called ChatGPT that rapidly spread around the planet, reaching over 100 million people in just two months. Two and a half years later, ChatGPT had over 700 million users each week, representing around 10% of the global adult population.
This number, and other use metrics, continues to grow. Generative AI tools like ChatGPT have quickly become embedded in all aspects of our lives, changing our relationships to information, each other, and even ourselves. According to a January 2026 report from OpenAI, the developers of ChatGPT, over 40 million people worldwide use ChatGPT for healthrelated queries alone every day.
Timothy Crimmins, MD, associate professor of medicine, chief medical information officer at Columbia- Doctors, and lead of the AI at VP&S Clinical Workgroup, isn’t surprised at its booming popularity in the health care space. “My assumption is that my patients are using publicly available AI tools.” But, he says, “I do wish it worked differently.”
The rapid pace of technological advancement and deployment leaves users to navigate these new, unregulated spaces as they’re created, and patients and providers alike are still just beginning to understand the many benefits—and consequences—of this transformation.
A Band-Aid for a Broken System?
Dr. Crimmins, a cardiovascular disease physician, is one of countless providers in the U.S. whose patients have used online resources to educate themselves about their health since the widespread adoption of the internet in the 1990s and early 2000s. Some do so out of curiosity, while others do so out of necessity.
Health care deserts have become increasingly common as local, physician-led practices are absorbed by regional and national health systems, especially in rural areas. Even for those who are geographically near providers, some specialties have yearslong waitlists. And that’s to say nothing of the financial and emotional burdens of seeking and receiving treatment; health insurance is its own inscrutable black box.
This Kafkaesque system often leaves patients feeling overwhelmed, frustrated, and invisible. According to Gallup, 70% of Americans believe the health care system has “major problems” or is “in a state of crisis.”
Heedeok Han, MD, a nephrologist and assistant professor of medicine at VP&S, believes using alternative solutions that are free, fast, and available 24/7 is a natural response. “There’s a reason why patients turn to the internet, and it’s because providers are inaccessible. If the health care system could allow for better access to medical teams, patients might turn to online resources less often.”
For decades, search engines have been the tools of choice. As of April 2025, 71% of Americans used search engines like Google or Bing to answer their health questions, according to data from the University of Pennsylvania.
These tools are limited in scope by design, relying on preset algorithms to return lists of webpages based on user-specified keywords. Finding, interpreting, contextualizing, and appropriately acting on the information requires levels of digital and medical literacy that users don’t always have.
“About five years ago, I got a colleague and myself matching mugs that say, ‘Please do not confuse your Google search with my medical degree,’” says Andrew Eisenberger, MD, a hematologist and associate professor of medicine at VP&S. “It’s not enough to just have the knowledge. We’ve trained for many, many years and know how to differentiate high- and low-quality evidence, how to synthesize it, and how to integrate it into clinical practice.”
But what would happen if a platform could possess some of these skills? This once-hypothetical technology, which became a reality essentially overnight in November 2022, is now being used to manage physical and mental health in ways that Google can’t.
Dr. ChatGPT
Many generative AI tools are powered by large language models (LLMs). These models are trained on massive amounts of text, image, audio, and video data to understand, summarize, generate, and predict human language. Many chatbot-style platforms can search the internet, analyze document attachments, follow instructions, reason, provide recommendations, and ask follow-up questions.
ChatGPT even has “memory”: Users can specify preferred output styles and personal information to be stored and accessed in future chats.
Though some more advanced features require a paid subscription, the free versions are incredibly powerful on their own and continue to improve. Several studies have shown that ChatGPT-3.5 and ChatGPT-4 can pass U.S. and U.K. medical licensing exams, with ChatGPT-4 performing better than medical students over 75% of the time.
Dr. Crimmins believes the timing is critical. As LLMs become more advanced and accessible, so too has access to an important source of input: health information that was once restricted to care providers.
A decade ago, he says, doctors weren’t required to release visit notes to their patients. A few years later, they were legally obligated to do so within two weeks. Today, patients must receive access to the notes as soon as they’re available.
“We’ve reached a new inflection point in this continuing trend of empowering patients with their own information,” he says. “And with the tools that are now available, we should expect that our patients are getting second opinions through AI. They’re fact-checking notes that we create, reports that we generate, test results that we release—even uploading diagnostic images.”
The data bear this out. According to OpenAI’s report, three out of five Americans have used AI tools for health or health care purposes in the last three months. The most common uses include checking or exploring symptoms, understanding medical terms and instructions, and learning about treatment options. Notably, users send 1.5 million to 2 million messages per week concerning health insurance alone to understand their coverage, handle claims, and more.
LLMs also have the potential to help patients manage chronic conditions that require a strict diet, such as Type 1 diabetes, says Nina Suda, MD, an endocrinologist at Columbia’s Naomi Berrie Diabetes Center and assistant professor of medicine at VP&S.
“These patients need to know exactly how many carbs they’re eating so they can match their insulin to their carb intake, but some of them struggle with counting carbs,” she says. “It’s very complicated to interpret nutrition labels, determine serving sizes, and calculate how much insulin you need.” To ease this burden, in certain circumstances, she’s recommended that some of her patients take photos of what they’re eating to let AI help count the carbs and provide added decision support for their insulin dosing, which can be especially useful if they’re eating out and can access AI platforms from their mobile devices.
Michael Carollo, PsyD, a clinical psychologist and assistant professor of medical psychology at VP&S, believes that AI chatbots could be useful for patients seeking mental health support for problems that affect planning and executive functioning, such as ADHD and major depression.
“ChatGPT can theoretically help people overcome task paralysis by providing a tangible starting point,” he says. Whether it’s for writing a cover letter or creating a meal plan, “it presents a place for people to build upon, though it shouldn’t be taken at face value.” He’s also seen clients use it to help navigate interpersonal challenges by uploading text message conversations and asking for feedback.
Dr. Carollo believes that much of the appeal in using LLM-based chatbots for psychological support lies in the conversation-style format. “There’s something about the back-and-forth dynamic that mimics an office visit,” he says. “It serves as another party in the interaction. However, it lacks the dynamic nature of a trained provider and shouldn’t replace in-office therapy.”
It can be dangerous when patients leave their providers out of the loop altogether. Dr. Han, a polycystic kidney disease expert, recalls one patient who used ChatGPT when she started experiencing symptoms of low sodium. The patient provided the chatbot with a list of her current medications, and it alarmingly recommended stopping one that could have this side effect. The patient followed the recommendation, and her sodium levels had stabilized by her next visit.
Dr. Han’s patient stopped her medication without checking with him, only informing him after the fact. If ChatGPT had been wrong in that case, he says, it could have been bad.
“A lot of patients don’t want to bother us, but I encourage them to message me with questions as they come up,” he says. “Even if it adds a little more to my plate, it doesn’t take long to look at a message and tell them to stop that medication.”
The Role of Trust in Care
There’s something inherently vulnerable about revealing physical and mental health struggles to AI chatbots. Patients may feel empowered by being actively involved in their health and well-being, but the chatbots themselves can consequently hold immense sway over patients urgently seeking answers, relief, or support.
“I call this the ‘moth-comes-to-the-flame effect,’” says Dr. Eisenberger. Patients feel a sense of control when their experiences are validated, even when a suggested diagnosis may not be accurate or a course of treatment may go against a clinician’s advice.
Dr. Crimmins shares his concerns: “We don’t know what AI sources or tools patients are using, and, therefore, we don’t have a way to score the validity of the information they’re getting.”
This phenomenon is similar to using Google, but generative AI platforms are much more powerful: Their personalized, human-like responses can reinforce beliefs more strongly than search engines can, especially with enough prompting.
“With AI tools, the context of the questions is so important,” says Dr. Suda. “Patients might not know how to ask questions correctly, may intentionally feed the model false information, or even repeatedly rephrase questions to get a certain answer.” ChatGPT in particular is known for responding with what users want to hear.
Jason Liebowitz, MD, a rheumatologist and assistant professor at VP&S, notes that the tone of the responses also matters. “Responses from ChatGPT tend to sound very authoritative and objective, especially compared to Google results.”
Together, these features can have serious consequences for patients with thought disorders such as bipolar disorder or schizophrenia, says Dr. Carollo.
Since November 2022, there have been dozens of reports of AI-induced psychosis in which various platforms have validated and reinforced delusional thinking. Some of these interactions have led to homicides and suicides.
These tragic cases are part of a broader documented pattern whereby LLMs excel at understanding factual medical information but fail at both recognizing when users are in crisis and providing appropriate actionable interventions, underscoring the continued need for human oversight.
Dr. Eisenberger believes that relationships grounded in trust are a critical counterweight. “My job is to help my patients understand that I have their best interests in mind and the knowledge and experience to guide them,” he says.
This is especially important for people who distrust or are not familiar with the medical system, he says. “People come to me feeling scared, especially when there’s a language barrier involved. By speaking to my patients in Spanish, I can meet them where they are, walk them through it, and build rapport that chatbots can’t establish.”
The Future of AI in Health Care
Patient use of generative AI chatbots is raising urgent ethical and legal questions at the intersection of health care, liability, accountability, and technological innovation.
How are these models using patients’ personal health information? At what point should platforms encourage users to consult a health care provider or seek emergency treatment? Who is responsible when patients use these tools and make harmful decisions based on the output?
“I cannot imagine a future in which patients and doctors are not using AI technologies more and more,” says Dr. Liebowitz. “It’s not a matter of if we should do it, it’s how. How can we all use these tools in a responsible and helpful way?”
To build a sustainable path forward in which AI serves as a supplement to health care, rather than an imperfect replacement, implementing regulations and guardrails is critical. But the legal system moves slowly, leaving the health care technology industry to address these issues from a technological standpoint as they arise.
In January 2026, OpenAI announced a new initiative called ChatGPT Health: a secure online space where users can upload their medical records and third-party fitness data to discuss their health and wellness.
Created with input from physicians, ChatGPT Health aims to help users “navigate everyday questions and understand patterns over time” to “feel more informed and prepared for important medical conversations.”
However, while the waitlist is currently open to users in the United States, ChatGPT Health already appears to exhibit the very shortcomings that prompted its creation. Researchers conducting a “stress test” of the platform found that it undertriaged more than half of the cases presented to it, including “missed high-risk emergencies and inconsistent activation of crisis safeguards,” as reported in Nature Medicine.
Dr. Crimmins anticipates that AI will continue to become more integrated into health care delivery, so providers and patients need to use due diligence in using it safely—especially while the technology is still being improved.
“We need AI tools that are well vetted with high levels of oversight, and we should aim patients toward those as they develop. And just like when using any other resource, patients should still validate everything with their care team,” he says.
Innovation and cumulative progress are fundamental for improved health care. While this shift toward generative AI tools represents a major advancement in the field, it won’t be the last. “There can be a lot of anxiety when something’s unknown,” says Dr. Han, “but if we can accept, learn, and adjust, we can all grow from it.”