Ten years ago, a question from the audience unsettled me enough to change the direction of my career.
It was 2016. I was still in residency, speaking at a national conference about the nuances of workflow optimization in imaging IT. After my talk, a data scientist from IBM walked to the microphone and asked, in effect: why bother with human radiologists when algorithms are clearly better at the work?
She also mentioned a renowned scientist who had recently gone on record to say the same. The smartest people in the world were asking this question.
I did not have a satisfying answer. I also did not know much about machine learning or the work of data scientists. But I had spent years of my life preparing for a profession that someone was suggesting might be unnecessary. The question bothered me.
So I decided to learn everything I could about it.
A familiar question, ten years later
I thought about that moment when I read Robert F. Kennedy Jr.’s recent remarks that AI could give patients a second opinion “much better informed than any doctor in the country.”
My reaction as a radiologist was less shock than recognition. We have been hearing versions of this argument for a decade.
As the argument reaches more of medicine, colleagues in other specialties are encountering a question that has shaped radiology’s professional conversation.
Six physician organizations, including the American Medical Association, responded with a joint statement emphasizing clinical context, professional judgment, and responsibility for care. They warned that portraying AI as inherently better informed could undermine patients’ trust. They also explicitly supported physicians helping lead responsible AI development and use.
Those concerns are legitimate. “Better informed” is not the same as more accurate, and being more accurate on a defined task is not the same as delivering better care.
Which system, doing what, for which patients, with what evidence? Those are necessary questions.
But for individual physicians, the response cannot end with explaining why the claim is too broad.
If the possibility bothers you, let that discomfort become a reason to learn.
What happened after that conference
Over the next several years, I learned everything I could about the technology behind that question.
Between 2016 and 2018, I worked through An Introduction to Statistical Learning (it still holds up). I learned that cleaning the data and defining the predictive task were more important than training the model itself. I became very familiar with DICOM.
I placed on the leaderboard of a Microsoft machine learning competition in 2016, using a boosted random forest model and a Gates Foundation dataset to predict women’s health risk.
Along the way, a co-founder and I created an NLP solution to automatically audit clinical notes and radiology reports, hoping to catch medical errors before they happened. We launched a startup and became finalists in UPenn’s business competition (it didn’t do well—another story for another time).
In 2018, I published a peer-reviewed study with Drs. Hanna Zafar, Maya Galperin-Aizenberg, and Tessa Cook, one of the earlier studies to systematically evaluate machine learning and NLP for large-scale mining of unstructured radiology report data.
Later that year, I joined the Cleveland Clinic.
Today, I get to spend much of my workday thinking about AI: how it changes healthcare workflows and whether it can reduce the burdens that contribute to clinician burnout. I get to talk about what it means to be responsible for AI in the same discussion as being responsible for patients. And best of all, I get to do this work with folks who really, really care.
The question that frightened me did not become irrelevant.
It became part of the fuel that shaped what I now call a career.
Earn the disagreement
I am not suggesting that every physician needs to become a machine-learning researcher. My path was one path. But there is a difference between disagreeing with a claim because we understand its limitations and disagreeing because we dislike its implications.
And when rigorous evidence shows that an AI system performs better than we do on a task, we should be the first to recognize it. A commitment to evidence-based medicine cuts both ways.
Learning also requires more than trying a chatbot once. It means examining evidence, working with people who understand the methods, and observing how a tool behaves in an appropriately governed setting.
And it means building. It is the widest gap you cannot bridge by reading papers. “Vibe code” an app for your personal life. Make an AI agent to answer questions about something you know by heart. Do it with the utmost care and follow hospital policies—but you must try. After all, that’s what our patients will be doing for their own health.
Familiarity is a beginning; expertise requires knowing when a tool’s apparent competence does not apply.
Ten years ago, I could have treated that question as something to rebut and forget. Instead, it pushed me to learn a technology that now shapes much of my work.
If you are a physician worried about what AI means for your patients or your profession, take the concern seriously. Learn enough to disagree for good reasons, and enough to recognize when the technology has something to teach you. Our relevance will not come from insisting that AI cannot do our work. It will come from understanding it well enough to transform what we do.
