18 May 2026 · LinkedIn
At a Crain's New York Business panel in March, the CEO of Westchester Medical Center reported that the AI his system uses for low-risk breast cancer screening produces a false negative roughly three times in every ten thousand reads.
Sitting alongside him, Mitchell Katz, who runs the eleven-hospital NYC Health + Hospitals network, told the room that he could replace a great deal of his radiologists with AI right now, if the state regulator would let him. He then asked the other CEOs in the room whether any of them could think of a reason not to push for the rule change.
What is significant here is not the prediction. Predictions about AI displacing radiologists have circulated for the better part of a decade, mostly wrongly. What is significant is that the question on the table has changed.
For most of that decade, the operative question was whether the technology could do the work. That question is now being closed. Three-quarters of the more than one thousand AI applications the FDA has cleared for medical use are in radiology. The performance numbers on specific tasks are no longer in doubt. The remaining argument is about liability, accreditation, and who carries the risk when a read goes wrong.
That is a different category of argument. It is the argument a profession has when its institutional defences are the only defences left.
Software engineering has been having the same argument, in a different register, for about two years. The question there is no longer whether AI can write production code; it is what a senior engineer is for, what a junior engineer is being trained towards, and what the unit of professional value becomes when generation is cheap. The radiologists are now arriving at that same threshold.
The narrower question is what happens to the people. There are roughly fifty thousand radiologists in the United States. They are among the most highly trained and highly compensated professionals in the labour market, with a decade or more of post-graduate education behind each licence.
The model Katz proposes, of AI for first reads and radiologists for abnormal findings, is the exact pattern by which expert professions thin out: the apex retains its function while the base of the pyramid disappears.
The thing nobody at the panel asked is where the next generation of senior radiologists is meant to come from, once the work that produces them has been automated away.
Join the conversation on LinkedIn →
← All writing