Medicine Is Not a Data Problem
Why Silicon Valley's quest for a 'Dr. Algorithm' misses the entire point of healing.
In Greg Egan’s haunting short story, "Learning to be Me," every person is implanted at birth with a "jewel," a tiny computer that spends a lifetime perfectly learning and mimicking the brain's every thought and action. At a chosen moment, the biological brain is removed, and the jewel takes over, granting a kind of immortality by replacing the fallible, mortal self with its perfect, durable, digital twin. Egan, a notoriously private computer programmer who has never been photographed, writes with the terrifying clarity of an engineer reverse-engineering the soul. His story is a chilling thought experiment that pushes the logic of replacement to its ultimate conclusion, forcing us to ask: if a machine can perfectly replicate every function, what is left of the original? What, precisely, is the self we are trying to preserve?
This is no longer just the stuff of Australian hard science fiction. In the summer of 2025, a team at Microsoft AI unveiled a jewel of their own. They called it the Microsoft AI Diagnostic Orchestrator, or MAI-DxO, and it represented the most potent version of this replacement anxiety made real. They took on one of the most sacred rituals in medicine: the complex diagnostic challenges published in the New England Journal of Medicine (NEJM), cases so thorny they often stump panels of human experts. The AI, working sequentially like a real clinician, could ask questions, order tests, and reason its way to a conclusion. The result was a thunderclap. On 304 of these complex cases, MAI-DxO achieved a diagnostic accuracy of up to 85.5%. For comparison, a group of 21 experienced physicians, tackling the same problems, hit the mark just 20% of the time. The machine wasn’t just good; it was more than four times better than the experts, and it did so while spending less money on tests.
The headlines wrote themselves: AI had arrived, not as a clumsy intern, but as a medical superintelligence. Egan’s jewel, it seemed, was polishing itself for duty.
The Fine Print on the Miracle
Before we hand over the keys to the clinic, however, a closer look at the fine print is in order. The Microsoft paper, while impressive, was a preprint, not yet subject to the full crucible of peer review. And the experimental design, while clever, was also perfectly artificial. The human doctors were hamstrung, forced to work in a vacuum without access to the very tools that define modern medical practice: colleagues, textbooks, and online resources. The "patients" weren't messy, emotional, forgetful humans; they were neat, orderly text files—the literary equivalent of a patient who never interrupts and has perfect recall. The NEJM cases themselves are curated success stories, puzzles with known solutions, a far cry from the ambiguous, inconclusive chaos of a real emergency room where false positives and red herrings are the norm.
This isn't to dismiss the achievement. It’s a powerful proof of concept. And it’s not an outlier. Researchers at Google have been developing their own formidable AI, the Articulate Medical Intelligence Explorer, or AMIE. In text-based diagnostic conversations, AMIE has demonstrated performance on par with, and sometimes superior to, human physicians in both diagnostic accuracy and, astonishingly, conversational quality and empathy. Now, with multimodal capabilities, it can even "see," interpreting medical images like X-rays and CT scans during a diagnostic dialogue.
The trend is undeniable. We are building machines that can, in certain well-defined, data-rich environments, reason about medicine with breathtaking speed and accuracy. This reality has reignited a fundamental debate about the very soul of medicine, a conflict personified by two competing visions for the future of the physician.
The Silicon Prophets and Their Visions
The contemporary debate about AI in medicine can be understood as a clash between two powerful worldviews, a proxy war for what medicine should even be. Is it an engineering problem to be optimized, or a humanistic art to be restored?
On one side, we have the provocateur, the techno-optimist who sees a flawed system ripe for disruption. There is no better avatar for this view than Vinod Khosla. An engineer by training and a co-founder of Sun Microsystems, Khosla’s entire worldview was forged in the crucible of Silicon Valley, where legacy systems exist to be dismantled. His critique of medicine is not just academic; it’s personal. A skiing accident left him deeply frustrated with the inconsistencies and opacity of the healthcare he received, transforming his theoretical belief in disruption into a visceral, personal mission.
In 2012, long before the current AI boom, he dropped a bomb on the medical establishment in the form of an essay titled, "Do We Need Doctors Or Algorithms?". His thesis was simple and terrifying: technology, specifically AI, will eventually replace 80% of what doctors do. Khosla argues that medicine as currently practiced is a "science of traditions," rife with cognitive biases, inconsistencies, and preventable errors. He envisions a future where "Dr. Algorithm" handles the heavy lifting of diagnosis and treatment planning, processing "helluvabytes" of data that no human could possibly comprehend. The human doctor’s role would be relegated to providing an empathetic interface, a comforting hand to hold while the machine does the real thinking. For Khosla, the human element is a bug to be patched, and the algorithm is the elegant, efficient solution. He sees a future of abundance where expertise is essentially free, and AI-driven primary care could be a government service for all.
On the other side stands the humanist, the insider seeking to reform the system from within. Dr. Eric Topol is not a Luddite; he is a pioneering cardiologist and geneticist who has been at the forefront of digital medicine for decades. He’s the last person you’d expect to be sounding the alarm about technology’s dehumanizing potential, which is precisely what makes his argument so powerful. In his book Deep Medicine, Topol argues that AI’s greatest promise is not replacement, but redemption. He sees a system already broken, not by a lack of data, but by a lack of time and presence. He diagnoses the modern condition as "shallow medicine," characterized by rushed appointments, dehumanizing EHRs, and staggering rates of physician burnout.
The numbers back him up. In 2024 and 2025, physician burnout rates remain alarmingly high, hovering between 43% and 49%, with a majority citing bureaucratic burdens from EHRs and administrative tasks as a primary cause. Topol’s argument is that AI can be the cure for this administrative disease. By automating the scut work—the endless clicking, the note-taking, the data entry—AI can give doctors the "gift of time". Time to talk, to listen, to think, and to restore the deep, empathetic human connection that is the true secret of patient care.
Where Khosla sees a flawed "practice" to be optimized, Topol sees a "shallow" and dehumanized system to be restored. Khosla's AI is a replacement, set to take over 80% of physician tasks; Topol's is an assistant, designed to automate administrative work and give back the "gift of time" for deep empathy. The conflict between these two visions is fundamental. It is a debate about the very definition of medicine.
The Ritualist's Lament and the Birth of Medical Abstraction
Into this binary debate steps a third, crucial voice, one that complicates Topol’s neat vision of restored humanity. Abraham Verghese, the celebrated physician-novelist at Stanford, offers a more profound and unsettling critique. Born in Ethiopia to Indian parents who were teachers, Verghese’s perspective is uniquely attuned to the power of story and ritual in shaping human experience. He is both a man of science and a man of letters, and it is this dual vision that allows him to see what many others miss. For Verghese, the issue isn't just about time; it's about the sanctity of the physical encounter.
His central argument is that the physical exam is not merely a data-gathering exercise; it is a "sacred ritual". It is a "primitive dance" where the act of a doctor laying hands on a patient builds a bond of trust and validates the patient’s story on their own body in a way no technology can replicate. This ritual, he argues, is being lost as our attention shifts from the patient in the bed to the "iPatient" in the computer—a digital construct that gets meticulous care while the real human often feels ignored.
Verghese’s perspective forces us to ask a difficult question. Topol suggests AI will free up time for empathy. But what if some of the "inefficient" work we are so eager to automate is the empathy? The time-consuming, hands-on ritual of the physical exam is, in Verghese's view, a primary vehicle for demonstrating care and building trust. If AI-powered remote monitoring and diagnostic imaging reduce the need for this physical encounter, we risk throwing the baby out with the bathwater. The very tool meant to enable more humanity could, if we’re not careful, lead to even greater distance.
This tension, this trade-off between human connection and technological data, did not begin with silicon. It began with a rolled-up piece of paper in 1816. In the early 19th century, a French physician named René Laënnec was faced with a dilemma. Examining a young, overweight female patient with heart trouble, he found the standard method of diagnosis, placing his ear directly on her chest, to be both ineffective and socially awkward. Inspired by children playing with a wooden tube to transmit sound, he rolled a sheet of paper into a cylinder, placed one end on the patient’s chest, and the other to his ear. The effect was immediate and profound. The muffled sounds of the heart and lungs became clearer, distinct, and amplified. He had invented the stethoscope, a technology born from a need for distance. In a moment of tragic irony, Laënnec would later die from tuberculosis, a disease whose progression he had so meticulously documented in others using his own invention.
This simple wooden tube was the first piece of "augmented intelligence" in medicine. It introduced a layer of technology between two humans, prioritizing objective, interpretable data over direct physical connection. Like AI today, Laënnec’s invention was met with a mix of enthusiasm and ridicule before its undeniable utility led to its universal adoption. It was the original sin of medical abstraction, the first step on a path that leads directly to the EHR, the remote monitor, and the uncanny intelligence of MAI-DxO and AMIE.
The Uncanny Intern and the Atrophy of Skill
As we move from the philosophical to the practical, the picture gets even murkier. AI is no longer a future hypothetical; it's in the clinic now, and its performance is both astonishing and troubling.
The diagnostic accuracy of modern AI is, in some narrow domains, breathtaking. Studies have shown AI models achieving over 90% accuracy in detecting breast cancer from mammograms, often outperforming human radiologists. AI is being used to analyze medical images, predict disease progression, and personalize treatment plans with incredible precision.
But this power comes with a paradox. A surprising number of studies have found that simply giving a doctor an AI assistant does not automatically lead to better results. In some cases, the performance of AI alone has been shown to be superior to that of a physician working with the AI. It seems we haven't yet figured out how to dance with our new partners. Physicians, it turns out, often undervalue or ignore the AI’s suggestions, especially when they contradict their own initial impressions.
Worse, these uncanny interns exhibit very human-like flaws. A 2024 study published in NEJM AI found that models like GPT-4 are susceptible to the same cognitive biases that plague human doctors, such as framing effects and hindsight bias—in some cases, even more so. The machine, trained on our data, has learned our bad habits.
This leads to the most insidious risk of all: skill atrophy. I’ve seen this firsthand in my own field of software development. On forums like Reddit, you’ll find senior coders admitting that over-reliance on AI assistants is causing their fundamental skills to deteriorate. They’ve started scheduling "AI Detox" days just to keep their own minds sharp.
Now, translate that to medicine. If a physician increasingly delegates diagnostic reasoning to an AI, will their own clinical muscles weaken? It’s the classic "pilot on autopilot" problem. A generation of doctors could become brilliant managers of AI-generated suggestions but lose the deep, intuitive expertise to know when the AI is subtly wrong, or to handle a crisis when the technology fails. The primary risk of medical AI may not be its incompetence, but its seductive over-competence, creating a dangerous dependency where the human is still accountable but no longer fully capable.
The Ship of Theseus in a White Coat
This brings us back to where we started, with Egan’s jewel and the philosophical heart of the matter. It’s the Ship of Theseus problem, retooled for the medical age. If you replace every plank of a ship, is it still the same ship? If an AI replaces every cognitive function of a doctor, is it still a doctor? Egan’s story pushes Khosla’s vision to its logical conclusion: a world where the biological self is discarded in favor of a perfect, immortal, digital copy. It forces us to ask: if an AI can perfectly replicate a doctor's mind, what is left of the doctor? What, precisely, is the "self" we are trying to preserve?
But there is a powerful counterargument, one offered by another giant of contemporary science fiction, Ted Chiang. Where Egan is a programmer exploring the physics of consciousness, Chiang is a technical writer by trade, and his stories are less about the code and more about the user manual for being human in a world full of new technologies. In his brilliant novella, "The Lifecycle of Software Objects," Chiang presents us with "digients," AI-powered virtual pets. The crucial insight of the story is that these digients cannot be "hothoused"—that is, they cannot be trained on a massive, static dataset in a simulated environment and emerge with true intelligence. They must be raised, slowly, over years, in a rich, interactive, embodied world. They need to play, to form relationships, to experience setbacks, to learn through a social process that mirrors the development of a human child.
Chiang’s story is a devastatingly elegant critique of the entire large language model paradigm. It suggests that our current AIs, for all their power, are fundamentally limited. They are masters of mimicry, brilliant at remixing the vast archive of past human knowledge. But they lack lived, embodied experience.
And this reveals the doctor’s true, unassailable moat. It is not their textbook knowledge; an AI can access that better. It is not even their diagnostic pattern-matching; an AI can learn to do that incredibly well. The doctor's most profound and durable advantage is the wisdom forged through years of lived, embodied experience with other human beings. It is the knowledge gained not from a dataset, but from sitting in a room with a patient, from the ritual of touch, from navigating thousands of messy, unquantifiable human stories. This is a form of intelligence that, by its very nature, cannot be downloaded.
The Centaur, Not the Cyborg
So, where does this leave the future of the physician? The answer is not a simple binary of replacement or augmentation. The future physician is not a cyborg, with human parts methodically replaced by superior technology. The better metaphor is the centaur: a seamless, hybrid entity where the human brain and the AI work as one, each playing to its strengths. The human provides the wisdom, the strategy, the ethical framework, and the empathy. The AI provides the boundless memory, the computational power, and the tireless analysis of data.
The doctor's job description is being rewritten. Their most critical roles will no longer be as a repository of knowledge, but as a master interpreter, translating the AI’s cold, probabilistic outputs into a coherent, compassionate, and actionable narrative for the patient. They will be an ethical navigator, making the hard calls when the data points one way but human values—dignity, hope, quality of life—point another. The physician becomes the final arbiter of what is not just possible, but right. In a world drowning in data, their highest calling will be to act as a guardian of context, to remember the story, to see the person behind the genome, the life behind the lab values. And to fight the atrophy of their own skills, the doctor must cultivate a healthy, informed skepticism of the machine, knowing its limitations, questioning its outputs, and having the courage to turn it off.
The stethoscope began a 200-year journey of translating the patient into data. AI is the final, breathtaking step on that path. The physician of the future will not be judged by their ability to read the data; the machine will do that better. They will be judged by their wisdom in reading the human. Their most profound, most valuable, and most uniquely human act will be to reach through the silicon and pull the patient back out.
I prefer Kremer s stetoscope!