Get the latest Syracuse news delivered right to your inbox.
Subscribe to our newsletter here.

Artificial intelligence has percolated into nearly every facet of daily life. People are using it to analyze the news, summarize books, meal prep and write their breakup texts. AI is also serving as a personal, 24/7 pocket doctor.

In January, OpenAI announced the launch of ChatGPT Health, indicating this experience will likely become more common.

As a medical student, I’ve seen this play out across medical specialties throughout my clinical experiences. It’s not always easy to get an appointment, and as humans, we naturally worry about our health. In the past, patients often went to the clinic after Googling their symptoms.

Now, much of that information is increasingly fed into AI models and search engines. While neither scenario replaces a doctor, searching symptoms on Google often leads to articles from experts and reputable institutions. AI models, though, don’t analyze or evaluate the accuracy of the information before presenting it.

I’ve seen patients come in after using AI to summarize lab reports or MRI reports, confused about their results as AI gave them an incorrect diagnosis or misinterpreted data. Some have agonized over the possibility of a brain tumor simply because an online source suggested their migraine could be as serious as cancer.

Going to the doctor can be expensive, inconvenient and time-consuming. While AI may seem like an easy replacement, it fails in key areas, making it far from an ideal substitute for an in-person appointment. A recent study in Nature found that none of the tested models were ready for direct patient care, despite strong performance on some measures.

To get useful results from an AI model, the user must input their own information. In contrast, when patients visit a clinic, an emergency department or urgent care, doctors are trained to guide the conversation to gather the information needed for diagnosis. Patients don’t always know which details are important and which are extraneous. The obvious or most debilitating symptoms may not be the ones that confirm the diagnosis.

Even more so, patients often lack the language to accurately describe their symptoms. As a future neurologist, a common example of this is when patients claim to feel “dizzy.” To them, the sensation is clear, but “dizzy” could mean lightheaded, unsteady, woozy or vertigo.

These nuances matter deeply and decrease the sensitivity of the AI model. Enter “dizzy” into ChatGPT, and you may be met with information about fainting when the problem is actually related to the inner ear.

Also, AI doesn’t usually provide a single answer. Continuing with the “dizzy” example, it may list several possible diagnoses, including benign paroxysmal positional vertigo, labyrinthitis, an ear infection, concussion, a brain tumor, anemia, heart problems or simply dehydration. It’s nearly impossible to decide which one is correct.

As a student doctor, that’s what my training is for: piecing together the tapestry of symptoms, labs and imaging. I’m trained to provide my patient with the most likely diagnosis and create a plan that works for both the disease and their lifestyle.

AI models don’t whittle down the options, and the recent Nature study found that people typically perform poorly when making these decisions.

This isn’t to say that AI can’t be useful for patients or physicians alike. A few years ago, a mother utilized ChatGPT to help diagnose her son after years of symptoms and pain. But, the information input into the model came from an MRI report read by a physician, and the ultimate diagnosis still required review of the images by a neurosurgeon.

While ChatGPT has proved useful in some cases, there are other instances where AI delays care. One patient experienced multiple episodes of vision changes following a procedure and consulted ChatGPT in an attempt to save a trip to the emergency department. The patient delayed seeking care for 24 hours, which postponed the diagnosis of a transient ischemic attack, also known as a mini-stroke.

The AI model provided unclear and, at times, inappropriate recommendations based on potentially outdated information.

Even when announcing its new product, ChatGPT Health, OpenAI admits it’s designed to support, not replace, medical care and isn’t intended for diagnosis or treatment. While the product is not yet available to the general public, people are already using ChatGPT to diagnose their ailments.

These AI models are simply not ready to be used to provide accurate diagnoses or treatment plans. Relying on them for that purpose does a disservice to one’s health.

These models and the information they provide aren’t infallible. The information can be inappropriate, inconsistent or incorrect. Taking ChatGPT’s health information at face value can be somewhat clarifying at best and deadly at worst. While AI can provide useful information at times, it’s vital to bring that information to a trusted medical provider for review and clarification.

These models shouldn’t replace a physician. It’s hard to distinguish fact from fiction when a model simply spits out information. A doctor or provider can help sort it out; that’s what they’re trained to do.

Samantha Ballas is a medical student. She can be reached at ballass@upstate.edu.

Published on February 25, 2026 at 8:54 pm