Some noise has been made lately about the imminent redundancy of the job of physician, thanks to the ascendancy of artificial intelligence (AI) medical technology.
The latest volley in the discussion comes via a JAMA essay, terrifyingly titled, “Will Autonomous AI Exceed AI-Aided Physicians as the Best Medical Care?” The authors quickly answer the headline question in the affirmative. But they go further, asserting that not only will AI render physicians obsolete but, in fact, it already has.
The piece lays out support for their claim and acknowledges some areas that remain to be researched on the topic. But there are two fundamental gaps in the essay, domains where not only does AI fall short in medical care but ones where it’s likely to always do so.
The first: user error.
There are, admittedly, studies that show better performance by AI than physicians in some scenarios. What everyone seems to forget, though, is that the specially trained or developed LLMs used by researchers in these studies are pretty much unavailable to anyone other than the scientists conducting the research themselves.
And your average Joe or Jane Patient also doesn’t have, and can’t get, access to the more common AI tools used by healthcare professionals (like OpenEvidence). Those require an NPI number, which isn’t granted based on how many followers you have on TikTok or how worried you are about that rash.
Argue, if you would like, that everyone should have access to these more sophisticated models, but the fact remains that not only do we not, but we probably never will.
Perhaps that’s a good thing. For as those trained in AI know, output from these models is only as good as the input.
Scientists researching how AI performs against human physicians have an intimate understanding of how to write a prompt. They give it the exact quantity and type of information required, provided in the format that is most likely to yield high-quality results.
Most of us, however, are typing our symptoms into Gemini (via Google) or free versions of ChatGPT or Claude late at night, bleary-eyed, while the newest episode of “The Traitors” streams in the background. We’re unlikely to provide the kind of relevant medical history a physician is trained to elicit, and it’s quite possible we’ll include outdated, misunderstood or just wrong information in our prompts. We might have faulty or incomplete memories of the lab results or diagnostic tests we’ve had done.
In most cases, we might be no more likely to get an accurate AI diagnosis than we were before AI, when Dr. Google was the greatest threat to medical professions.
Now, lest you think that while physicians won’t lose their jobs, they’ll become glorified prompt engineers, we come to the second, more important domain where AI can’t rival human physicians: medical mysteries.
AI remains thoroughly incapable of identifying novel pathogens, unknown diseases or syndromes. If it hasn’t been named, categorized and entered into its training data sets, it might as well not exist. LLMs are like robots that expertly shelve books at a library, faster than any human ever could. But what do they do when it finds a book that it’s never heard of, one that doesn’t fit into any existing category?
An example: Dr. Louis Weinstein sat with a pregnant Navajo woman at her bedside in the 1980s as she died from a condition that physicians at the time couldn’t explain. If that woman had plugged her symptoms into ChatGPT (or even OpenEvidence), the AI model would not have been able to conceive that she was suffering from some as-yet unnamed condition.
We don’t know exactly what an LLM would have diagnosed her with, but we do know that it would not have spent years finding and treating similar patients, isolating their commonalities, eventually naming their condition with an acronym listing the main symptoms, making it easier for physicians to identify and treat patients with HELLP syndrome.
Humans are not, and never will be, computers. While, sure, that occasionally has its drawbacks, the limitation comes with a creativity of thought that can’t be algorithmically duplicated. We are willing to question what AI always accepts.
AI’s strength comes in pattern recognition. That may be all well and good for diseases that fit an established pattern and are already well-understood, but there are many more parts of the human body that remain a mystery.
AI will be a key part of medicine, but without an intrepid human mind, it will sometimes fall short, trying to make the proverbial peg fit into one of the previously bored holes.
It will never drill a new hole.
So, before we decide that physicians are no more important in 2026 than VCR repair, we should pause. Medicine will have to adapt, and we’ll need to have conversations about ethics and access.
But I don’t plan to end my relationship with my PCP in favor of ChatGPT, and I don’t think you should, either.
Inside the human body, there is too much left to know and too many worlds to explore, worlds that only humans are capable of discovering.
Photo credit: Online Marketing at Unsplash