Math And Medicine May Not Be The Best Disciplines In Which To Cede The Training Of The Next Generation To AI
By
Leonard Zwelling
https://www.nytimes.com/2026/09/22/science/math-ai-history-understanding.html?searchResultPosition=1
https://www.nytimes.com/2026/09/22/science/math-ai-history-understanding.html?searchResultPosition=1
Artificial intelligence is impacting all aspects of our lives. It is also impacting the lives of those whose job it is to train the next generation of mathematicians and physicians.
In the first story from The New York Times Science Section on September 29, Cornell math professors Steven Strogatz and Alex Townsend write about how AI has solved some of the knottiest problems in mathematics that had gone unsolved when humans tried. Among those humans who have historically tried to solve these problems are mathematicians in training. It is the process of trying to solve the hardest of problems that teach the methods and skills that will be used for a lifetime in math.
In the second Times piece from September 27, geriatrician Rachel Bedard thinks the benefits of AI to physicians are immeasurable, but perhaps its frequent use is not the optimal way for medical students and residents to learn their profession. Like mathematicians in training, doctors in training need to learn to think on their own about difficult diagnostic and therapeutic challenges so that they will be prepared to use these skills later in their careers. There are many times and places that call for medical intervention in the absence of access to sophisticated technology. These include rural health settings, massive casualty events, and the battlefield.
The common thread here is how can humans be taught to think in a highly complex scientific discipline if they are trained using AI to solve problems for them rather than learning to use their own brains?
I cannot meaningfully comment on any aspect of mathematics including the wisdom of using AI to solve problems successfully that have evaded human solutions for decades. Calculus baffled me in high school and in college. However, if these problems were worth solving, and I have to assume they were, then surely the use of AI to solve previously insoluble problems was beneficial.
Likewise, in an area I know a bit about, medical diagnosis, I recently used Open Evidence, a large language model for doctors in which the information within the machine is based solely on results from peer-reviewed publications. I queried the model to generate a differential diagnosis for a patient I was asked to interview with some unusual and confounding signs and symptoms and who was without a definitive diagnosis.
Open Evidence, the large language model, performed magnificently. It gave a perfectly comprehensive differential, a list of tests to rule-in or rule-out alternative ailments, and then listed the corresponding appropriate treatments for each condition on the list. There is no way on my best day as an internist, which was many years ago, I could have done as well as quickly. So, what’s the problem with having medical students and residents use AI during training?
Here’s the flaw. I had to sit with the patient and ask the right questions to get the list of signs and symptoms to type into Open Evidence. I also had to sift through these data to create a comprehensible story even for the machine. It is unlikely that a patient such as this one could have fed the machine a cogent story all alone. The patient was all over the place in listing the symptoms and the timing of their initial manifestation. I had to repeat questions frequently to get an accurate timeline of the patient’s symptoms and dig deeply into current problems sorting out what the patient called something to what I might call it in medical lingo. Then I had to construct the story to give to Open Evidence. My skill in creating an accurate picture is what Open Evidence will rely on to give me a good answer.
It is possible that a patient, a keyboard, and the large language model alone could have arrived at the same differential diagnosis, but I doubt it and that’s the point.
It is still necessary to train physicians to take a good history, know the right questions to ask, and do a detailed physical exam to accurately record significant findings before enlisting the help of Open Evidence for diagnostic and therapeutic purposes. In other words, AI is just another tool in the doctor’s black bag.
In reading about the dilemma that mathematicians have in training the next generation and seeing for myself, as a patient, the inability of young physicians to synthesize a cogent differential diagnosis without the help of CT scans, MRIs, many consultants, and thousands of dollars in lab tests—none of which might they have at their fingertips some time later in their careers—I agree with the conclusions of these editorial writers.
AI will advance mathematics and medicine, but those learning these disciplines ought to get back to the basics first or they themselves will be no better than expensively-educated machines. I, for one, do not want a robot for my doctor unless it’s a DaVinci with a real, human surgeon on the other end.
