The Diagnostician’s Brain: Why Judgment Beats Logic in the Age of AI in Healthcare

Artificial intelligence is often mistaken for general intelligence. In reality, AI has mastered one specific form of intelligence: logic. Data in the world exists in relationships—A equals B, B equals C, therefore A equals C—and computers reproduce and remix those relationships with remarkable, reliable speed. That is a genuine achievement.

But logic, it turns out, is the most basic form of intelligence a physician uses, not the most sophisticated. Understanding what actually separates expert clinical judgment from computation matters enormously as AI tools enter daily practice.

The Narrative Logic of Clinical Judgment

Human cognition does not organize information the way a database does. It organizes information as narrative. Show a person a still image of ancient Greek soldiers hiding inside a massive wooden horse, and their brain doesn’t just catalog “horse, soldiers, sea”; rather, it instantly activates a 10-year story of deception, war, and a long journey home, complete with themes and a trajectory.

A computer, by contrast, can label every object in that image perfectly while grasping none of the plot.

This narrative capacity does two remarkable things once the brain settles on an interpretation of a scene. First, it projects forward: A surgeon looking at an intraoperative field can imagine branching future complications, including ones they have never personally encountered, because the brain understands how bodily “stories” tend to unfold. Second, it works backward: A clinician seeing a college student home from his first month away with a persistent sore throat can reconstruct the likely chain of exposure that led to a diagnosis of mononucleosis, then project forward again to anticipate a spectrum of outcomes, from full recovery to the rare complication of a fragile, enlarged spleen.

That dual capacity—reconstructing the past and forecasting the future from a single data point—is not a parlor trick. It is the foundation of diagnostic reasoning, and it depends on something computers lack: common sense built from lived experience.

The ‘Sources of Power’ Behind Clinical Expertise

Researcher Gary Klein studied naturalistic decision-making among firefighters, military commanders, and NICU nurses and found that the nurses who were best at detecting a critically ill infant were not outperforming the algorithms or the vital sign monitors through superior technical knowledge alone. They had built what Klein called “sources of power”: pattern recognition, mental simulation, and intuitive judgment developed through sustained immersion in real, high-stakes environments.

A recent study of emergency department diagnosis—drawing on more than 40 interviews, 60 critical diagnostic incidents, and 180 hours of direct observation—found the same pattern. Broad technical knowledge mattered, but it sat alongside a constellation of other skills entirely absent from any textbook or dataset. Notably, pure logical brilliance without those other skills does not necessarily produce a good clinician. Fictional diagnostician Dr. House is the cautionary extreme: encyclopedic and inferentially sharp, but someone few of us would actually want running a real care team.

Detecting What the Chart Doesn’t Say

One of the sharpest illustrations of this kind of reasoning predates modern neuroscience by more than a century. In Arthur Conan Doyle’s short story “The Adventures of Silver Blaze,” Sherlock Holmes solves the disappearance of a prize racehorse not by analyzing evidence in front of him but by noticing evidence that is absent: that the watchdog never barked on the night of the theft. That absence tells Holmes the intruder must have been familiar to the dog, pointing to the horse’s own trainer as the culprit.

In medicine, the equivalent is often not what appears in the chart, but what should be there and is not, such as a missing symptom, an absent risk factor, a pattern that doesn’t fit, or a clinical silence that changes the meaning of everything else. This is anomaly detection, and it is central to expert diagnosis.

A patient may present with seven out of eight classic findings of heart failure, but if they are also losing weight—a finding incompatible with a straightforward heart failure exacerbation—an expert clinician’s brain flags the inconsistency and demands an explanation before accepting the obvious answer, considering alternatives like hyperthyroidism, endocarditis, or malignancy.

Computers excel at analyzing the data placed in front of them. Drawing meaningful inference from the data that is conspicuously missing remains a distinctly human strength.

Why Clinical Judgment Needs More Than Data

If you could combine the entire legal code, all case law, and every piece of personal data a person has ever generated into a single unbiased decision-making system, would that be an improvement over human judgment?

A recent film, Mercy, imagines exactly this scenario, with an AI court system empowered to decide guilt and innocence based on total information about human actions and communications. Most people’s instinctive reaction to that premise is discomfort, even when the logic seems airtight on paper.

A more grounded version of this question comes from pediatric burn care, where the number of decisions required for a single patient—when to operate, which antibiotic to use, whether irritation signals infection, when to transfer care—is enormous. Posed with a choice between a burn surgeon with seven years of hands-on experience and an algorithm trained on hundreds of thousands of prior cases from electronic health record data, audiences overwhelmingly choose the surgeon to make those decisions for the patient.

They do so not because the algorithm’s statistical recommendations are wrong but because what is actually being asked for in that moment is not the statistically optimal answer.

It’s judgment.

It’s someone who knows not just the biology but the biography of this particular child, this particular family, and this particular health system, informed by having lived through similar moments before.

Training Clinicians for Judgment, Not Memorization

This reframes what medical education is for. We do not spend years training physicians primarily to memorize facts. Those are increasingly discoverable in seconds. We train them to develop judgment. We train them for the capacity to customize medical knowledge for the specific patient sitting in front of them, the same way a good lawyer customizes the legal code for a specific case rather than simply reciting it.

As AI tools increasingly handle the logic-based, pattern-matching layer of medicine, the differentiating skill for physicians becomes even more clearly the narrative, anomaly-detecting, judgment-forming capacity that no dataset fully captures.

That is not a reason to fear these tools. It is a reason to be deliberate about what we ask them to do, and what we insist on preserving as irreducibly human.

New physicians should not worry about machine learning; they should focus on how to become a learning machine.

Gurpreet Dhaliwal, MD, spoke at the 2026 TDC Group Executive Advisory Board Meeting in Napa, California, which gathered top healthcare executives, academic researchers, and clinical leaders to discuss “Healthcare AI in 2026: Emerging Issues as the Field Matures.”

Dr. Dhaliwal is a clinician-educator and Professor of Medicine at the University of California, San Francisco. He is the site director of the internal medicine clerkship at the San Francisco VA Medical Center, where he teaches medical students and residents in the emergency department, urgent care clinic, inpatient wards, outpatient clinic, and morning report. He studies, writes, and speaks about how doctors think—how they make diagnoses, how they develop diagnostic expertise, and how they interface with technology to augment their

The opinions expressed here do not necessarily reflect the views of The Doctors Company. We provide a platform for diverse perspectives and healthcare information, and the opinions expressed are solely those of the author.