The triage note read, “The patient wants her thyroid checked.”
On the stretcher, fully dressed, sat a 65-year-old woman with her family members at the bedside. She said she had been feeling weak and tired lately, so bad that even her voice had turned hoarse. She’d never had any problems with her thyroid in her life, but she had looked it up; maybe it was her thyroid or anemia. That was what the AI said.
I thought about an old coffee mug at the nurses’ station brought in by one of the providers. It has the slogan “Please don’t confuse your Google search with my medical degree.” I can understand where the irritation comes from. As clinicians, we feel like our decade of training and twenty years in medicine are being weighed against your 30 seconds of typing into a smartphone. We’re not just dealing with Google searches now; we have patients who are having long discussions about their health care with generative AI.
Like everyone else, I use AI to answer everyday questions. If I want to know what Bob Dylan’s last album was or look for a quick and easy cookie recipe, I know exactly who to ask. It’s like having a brilliant best friend who will only occasionally lie to you. It makes sense that if something is going wrong in our bodies, AI would be the first place we look. It’s a reasonable and rational decision. I acknowledged the patient’s concerns about anemia or thyroid issues. I told her that these are a good start, but that’s just the beginning of the journey.
Real physician voices, twice a week
Free, and one click to unsubscribe.
There are a couple of things that laypeople don’t understand about medical education. The first is that the volume of information is incomprehensible. One individual can’t come close to understanding everything we know about the human body. One of my medical school professors told us about a textbook titled The Pupil. It was in two volumes with well over 2,000 pages, just about diseases of the eye.
The second is that when medical students learn about disease processes, like congestive heart failure, they study an organ system and disease, then learn how the disease presents. When the patient comes to us, they don’t have a sticker on their forehead that says congestive heart failure. They present with symptoms first, and we work backwards to identify the illnesses that could cause them. Medical students basically learn one way and have to practice in reverse.
A large volume of information with a complex web of interconnections is precisely the kind of thing AI should be good at, but there are a couple of caveats. AI produces a list of possibilities when you enter symptoms and sometimes understands the probability of each, but it can’t examine the patient or grasp what really scares them. When a patient tells me they’re having palpitations, they want to make sure they don’t have an adrenal gland tumor, despite AI telling them the condition is possible but unlikely. I have to delicately explain to them that the three shots of espresso they just had are about a million times more likely to be causing the symptoms than an adrenal gland tumor. And when AI doesn’t know the answer, it rarely admits it. It can be confidently incorrect, something people call AI hallucinations. An error in a cookie recipe can be forgiven, but an incorrect dose of a chemotherapy agent cannot.
As I gathered information on my patient with weakness, I went down the standard list of follow-up questions. How long has this been going on? Have you done anything out of the ordinary? My next question hit the nail on the head. I asked her if anything had made it better or worse. She just said, “Sleep.”
I stopped and looked up from the patient’s chart. “What do you mean, ‘sleep’?”
“Oh, when I wake up, I’m completely fine. I get worse as the day goes on, and by the end of the day I can barely talk. Then I go to sleep, and everything goes away by morning.”
Weakness that gets worse as the day progresses and resolves with rest, a hallmark symptom of a particular disease. I asked her to hold her head still and watch my hand, which I held high above my head, forcing her to maintain a sustained upward gaze. My arm started to tire after a minute, but by this time both of her eyelids were starting to droop. After two minutes, she couldn’t hold her eyes open.
It was the only time in twenty years of emergency medicine that I suspected new-onset myasthenia gravis, a rare autoimmune disorder that used to have a poor prognosis but has more effective treatments now. When I go back and ask AI to list all the causes of weakness, there are over 100 items, and myasthenia gravis is pretty far down that list. A positive response to a single follow-up question moves it to the very top of the list.
So we shouldn’t dismiss a patient who turns to AI for help. Many times AI comes up with a similar list that any diagnostician would create. Someday, it may be able to examine the patient, but right now it can’t ask a patient to move their gaze and watch what happens. Clinicians who use AI as intended, to handle large volumes of information and catch what they may have missed, will practice better medicine than those who don’t. AI can help you know what to look for, but it can’t do the looking for you.
Duane Penshorn is an investigative nonfiction author and thriller writer. A practicing emergency physician, former ER medical director, and hospital chief of staff, he draws on decades of frontline clinical and forensic training to diagnose the hidden corporate frameworks shaping modern life. He writes about what happens when institutional design clashes with real-world patient care, exposing the metrics and bureaucratic bottlenecks that drive clinician burnout and threaten safety.
Penshorn brings the same clinical realism to his fiction as an award-finalist science fiction and medical thriller author. As the founder of Astus Press, he writes across genres to explore the unintended consequences of complex modern systems, advocating for a return to intuition over dashboard compliance. He practices at Baptist Neighborhood Hospital and lives in San Antonio, Texas.
More can be found on his website.
