In a column and podcast episode for KevinMD, retired surgeon and patient advocate Alan P. Feren argues this week that artificial intelligence is eroding physicians’ clinical judgment, turning doctors into editors of machine output rather than originators of diagnosis.
He calls the remedy “actual intelligence first, artificial intelligence second.”
Feren, who practiced as a surgeon and now works as a patient advocate, published the essay “Actual Intelligence: the skill AI cannot replace” on KevinMD and discussed it on The Podcast by KevinMD, hosted by internist Kevin Pho.
The piece lands amid rapid adoption: an American Medical Association survey in 2026, cited by Feren, found that 81 percent of physicians now use AI tools, while 88 percent worry the technology is eroding clinical skills.
The core of his argument is not that AI makes obvious errors. It is subtler. “AI forces you to edit instead of originate,” Feren told Pho on the podcast.
When a tool surfaces a plausible answer in a time-pressed clinic, he argues, it can shut down the differential diagnosis — the doctor’s own list of possible explanations — before it has been built. “Skill loss is what happens when you begin to just edit the output as opposed to originating it,” he said.
That matters to patients because the failure mode is invisible. A chart can look complete while missing the one detail that changes everything, and the clinician, presented with an authoritative-sounding answer, may never look for it.
AI answers can be plausible yet incomplete
- An AI answer can be plausible and still incomplete — it only knows what it was given.
- Tell your doctor what an AI tool suggested to you, and what it left out.
- The physician, not the software, remains responsible for the diagnosis.
The evidence he leans on
Feren’s concern is anchored in a 2025 Polish study of gastroenterologists using AI to help identify adenomas — precancerous growths in the colon. The study suggested a statistically significant drop of about 6 percent in adenoma detection when doctors who had first worked with AI assistance then worked without it three months later.
He acknowledged on the podcast that several later studies did not replicate the finding, and he presented it as a signal worth tracking, not a settled conclusion.
He also flagged a problem with the tools’ scholarship itself. In his own writing, he said, “probably up to 5 percent of the references that are surfaced are either nonexistent or incorrect” — the phenomenon known as hallucination, where a model fabricates plausible-looking citations.
For the next generation of physicians, Feren pointed to what some commentators now call “never-skilling”: trainees who rely on AI from day one may never build the diagnostic instincts that experienced clinicians developed by seeing patient after patient.
He drew a parallel with his own surgical training after the introduction of laparoscopic cholecystectomy — gallbladder removal through small incisions — where knowing when not to operate, and when to convert a minimally invasive procedure to open surgery, was a judgment that kept patients alive.
A case that shows the gap
To illustrate, Feren described a hypothetical vignette, not a real patient: a 52-year-old fit woman with pain between the shoulder blades — the interscapular region — that does not resolve. An AI chatbot suggests a muscle strain. Her doctor, hearing the same story, agrees.
But the chart holds two facts the AI was never given: dyslipidemia (abnormal blood lipids) and a father who died of a myocardial infarction — a heart attack — at 52. In a woman in her fifties, interscapular pain with that family history can be angina, reduced blood flow to the heart, not a pulled muscle.
“There’s just no replacement, I don’t think, for clinical judgment and our individual experience,” Feren said on the podcast. His pushback at arguments that AI reasoning already exceeds physicians’: the software does not know when a patient hesitates, or how a chatbot the patient used beforehand has already shaped the story.
The other side of the argument
Feren directed his disagreement at health policy Ezekiel Emanuel, who has argued recently that AI’s clinical reasoning is more accurate than physicians’ and that inserting a human into the loop can, in some studies, reduce AI’s accuracy. Feren’s counter, as he put it on the podcast, is that AI “really doesn’t know the patient’s circumstances” — the hesitations, the chart details, the prior chatbot that narrowed the story before the doctor ever heard it.
The two positions are not settled between them; the debate is live in the literature and in opinion pages, including a New York Times op-ed by Rachel Bedard, “Being a Doctor Will Never Be the Same After AI,” which Feren cited.
What he asks of clinicians and patients
For practicing physicians under productivity pressure, Feren’s advice is to establish clinical reasoning before looking at AI framing — “Take what fits, verify, and then discard,” as he put it. For patients, he said AI can help organize a story to bring to the office, but they should also tell their clinician what the tool failed to capture.
| Figure | Value | Source cited by Feren |
|---|---|---|
| Physicians using AI (2026) | 81% | American Medical Association survey |
| Physicians worried AI erodes clinical skills | 88% | American Medical Association survey |
| Adenoma detection drop after AI use, three months later | ~6% | Polish gastroenterology study, 2025 |
| AI-surfaced references nonexistent or incorrect | up to 5% | Feren’s own writing experience |
Chest or back pain in over-50s needs prompt medical evaluation
This is a professional-practice debate, not a treatment, so there is nothing to avoid or take. The relevant safety point comes from Feren’s own vignette: new, persistent chest, upper-back, or interscapular pain — especially in someone over 50, or with risk factors such as dyslipidemia or a family history of early heart attack — warrants prompt medical evaluation, not reassurance from a chatbot.
Patients using AI to research symptoms should bring both the tool’s conclusions and its omissions to their clinician, and treat any AI-surfaced answer as a question for a professional, not a diagnosis.
Where the debate goes next is unresolved. Replication of the Polish deskilling finding remains open, Feren himself noted, and no regulator named in the source material — the FDA, EMA, or NHS — has issued a position on clinician deskilling. What is on record is Feren’s closing formulation: actual intelligence “has to be practiced. It’s not merely present in the room.”
References
- Alan P. Feren, “Actual Intelligence: the skill AI cannot replace,” KevinMD.
- The Podcast by KevinMD, interview with Alan P. Feren, hosted by Kevin Pho, KevinMD.
- American Medical Association physician AI survey, 2026, as cited by Feren.
- Polish study of gastroenterologists and AI-assisted adenoma detection, 2025, as cited by Feren.
- Rachel Bedard, “Being a Doctor Will Never Be the Same After AI,” The New York Times, as cited by Feren.
Disclaimer: This article reports on a published opinion piece and podcast discussion. It is not medical advice. For personal symptoms or health decisions, consult a qualified clinician.