ChatGPT Wasn’t Wrong; It Was Solving the Wrong Problem

It will probably be my undoing that I have basically checked out with regard to reading the AI vs. doctor studies. They seem to have so little to do with the medicine I practice every day. I cannot imagine a time when a person, a patient, will be satisfied making a complicated decision about their health with only a computer consultant.

Dr. McBride has obviously not checked out. Here she articulates some of my irritations with the new genre of AI vs. MD.

Adam Cifu

A patient in her mid-60s, weary from caring for her mother with dementia, asked me to prescribe a GLP-1. “ChatGPT said it could help my memory,” she reasoned. “I can’t remember anything.”

The problem? The treatment for caregiver burnout doesn’t come in a shot. What my patient needed was more sleep, less alcohol, and time to catch her breath. ChatGPT wasn’t wrong that GLP-1s may protect brain health, but its advice didn’t touch the realities of her life. The chatbot had answered what it was asked. It couldn’t surface the questions she meant to ask.

Last month in JAMA, bioethicist Ezekiel Emanuel and colleagues argued that AI is already superior to physicians at five cognitive tasks, including eliciting the patient’s story, and that keeping doctors in the loop with AI actually degrades care. The piece flooded my LinkedIn feed: doctors horrified, AI evangelists applauding. The evidence they cite is impressive. But nearly all of the data rests on methods that sidestep real life: in most studies, the LLM was handed a curated case vignette and asked to spit out a diagnosis. The studies behind the history-taking claim barely had AI talk to a patient—two of three analyzed electronic medical records; the third had doctors typing like chatbots to actors. It’s no surprise the models do well; LLMs excel at generating answers from organized information.

My patient isn’t an outlier. According to a June KFF poll, 29% of adults now use AI chatbots like ChatGPT for health information monthly—nearly double two years ago. Patients are not wrong to reach for these tools given the barriers to care—cost, access, being dismissed or unheard.

But the most pernicious problem in US healthcare isn’t a lack of solutions; it is the absence of space for question-formation—the slow, unglamorous work of discovering what a patient is actually asking—before anyone writes a prescription. At the very moment primary care is on life support, receiving less than 5 percent of US health care spending, chatbots are poised to industrialize the transactional medical encounter that patients (and doctors) resent most.

In January 2025, researchers at Harvard and Stanford asked a different question: What happens when the model is asked to interview patients the way a clinician does, instead of receiving a pre-packaged vignette? Diagnostic accuracy fell sharply: in dermatology cases, from 82 percent to 63 percent. As a primary care practitioner for 25 years, I’m not surprised. Clinical information derives from more than words—the moment my patient welled up with tears discussing her caregiving responsibilities, the stiffening of her body when I asked about alcohol use. These are data too.

The newest systems are built to close this gap. Earlier this month, Google announced that AMIE, its medical AI, now conducts real-time video consultations—and clinical evaluators rated its history-taking as thorough as primary care physicians’. But look at who it was interviewing: patient actors, performing scripted cases. An actor has been handed her story in advance. My patient hadn’t. Her real problems couldn’t have been packaged up because she had not yet formed them for herself.

Even Emanuel et al. concede that communication between patient and model remains “a particular point of failure.” As a primary care doctor whose recommendations rest entirely on human-human communication, I might swap “particular” for “catastrophic.” In Google’s study, the actors rated the AI favorably on empathy and rapport, but rapport with an actor is rehearsed. Even Google admits that actors cannot replicate the complexity and unpredictability of real patients. The machines win at answer generation; what hasn’t been shown is their ability to build what undergirds every answer: trust from a person who is sick, scared, and not yet sure what her problem is.

There are questions that emerge only between two people who have time and a reason to be honest with each other—the question beneath the question. Patients will disclose remarkable things to a machine. But discovery is not the same thing as disclosure. The art of medicine includes helping a patient articulate her true problems—medical, social, and emotional—in such a way that this information shapes our treatment plan. It requires time and patience and sometimes awkward silences. That is not something any chatbot has been shown to do.

This is not an argument against AI in medicine. In my own practice, an ambient AI scribe drafts my visit notes so I can look at my patient rather than at a screen. AI drafts letters to insurance companies that refuse to cover my patients’ medications. These are the administrative tasks that don’t require a human—and that have been crowding out the human experience of medical care for decades.

So here is the assignment. Give AI the gruntwork that has hollowed out primary care: the documentation, the prior authorizations, the inbox triage that forces doctors to look at screens instead of the patient. Then pay for human conversations, not just the procedures and prescriptions that follow them. Otherwise, AI will simply shrink the fifteen-minute appointment to ten. Use LLMs for diagnoses—great—if they are so much better than us, but only after the conversation, after trust is built, after the patient has had time to surface her real problem. And teach clinicians—and patients—the skill no model yet replicates: knowing which question to ask, and which question to ask yourself first. Even JAMA’s authors call for urgent work on reimbursement and medical education. At least we agree on the path forward, even if we disagree on the destination.

My patient didn’t need a prescription drug; she needed a human to understand her. The allure of the quick fix is human nature; the problem is that it’s inhuman to suggest there are easy solutions to being alive. Questions like: What keeps you up at night? What’s going on in your life? are the ones no one—human or machine—thinks to ask. Let’s build a system that gives doctors time to ask them.

Lucy McBride, MD, is a primary care physician in Washington, DC, and the author of Beyond the Prescription: A Doctor’s Guide to Taking Charge of Your Health.