What a 1992 Essay by a French Philosopher Can Teach Us About Artificial Intelligence in Healthcare
A few weeks ago, I attended a demonstration of an artificial intelligence tool for radiology. The presentation was impressive. A mammogram appeared on the screen and, within seconds, the system automatically highlighted the areas it considered suspicious. The radiologist still had the final say. No one suggested otherwise. Yet, as I listened to the discussion that followed, I noticed the conversation drifting in a familiar direction. The questions quickly became the same ones we hear almost every time AI is discussed: How long will it take before artificial intelligence replaces part of the radiologist’s work?
As a physician, I spent a long time believing that AI’s greatest impact would be speed. Finding literature in seconds. Automating repetitive tasks. Detecting patterns that would take a person weeks to uncover. I thought about efficiency. I thought about productivity. It never occurred to me that the most important change might lie somewhere else.
I often recommend reading the 1992 essay by the French sociologist and philosopher Bruno Latour (1947–2022), Where Are the Missing Masses? The Sociology of a Few Mundane Artifacts. His central idea is remarkably simple, yet its implications are profound. We tend to describe decisions as if they were the product of a single individual. We say that a physician decided on a treatment, that a researcher chose a particular strategy, or that a driver decided to slow down. For Latour, this way of telling the story oversimplifies reality. Decisions do not emerge from isolated individuals. They emerge from a network in which people, technologies, institutions, rules, infrastructure, and accumulated knowledge continuously interact.
Latour illustrates this idea with the well-known example of a speed bump. The driver is still the one pressing the brake pedal, but it would be difficult to argue that the speed bump played no role in that action. The road changed the driver’s behavior. The decision remains human, but it can no longer be explained simply by looking at the driver alone.
We often talk about clinical judgment as though it belongs exclusively to the physician. Yet anyone who has participated in a multidisciplinary case discussion knows that this picture is far from complete. A clinical decision is shaped by the patient’s medical history, imaging studies, laboratory results, pathology reports, clinical guidelines, the experience of the healthcare team, treatment availability, and, quite often, the patient’s own preferences. The physician signs the decision, but the reasoning behind it has never occurred in isolation.
Until recently, medical technologies provided us with information, and from that point onward our work began. Artificial intelligence is starting to occupy a different place. An algorithm reviews a mammogram before the radiologist and highlights the regions it considers relevant. Another identifies microscopic patterns in a biopsy that deserve a second look.
For months, we have described these advances by saying that artificial intelligence interprets images or analyzes data. But as I reread Latour, I began to wonder whether that description misses something important. The truly interesting development is not that an algorithm can interpret a mammogram. It is that its interpretation becomes part of the process through which a physician forms clinical judgment. The radiologist still examines the image, remains responsible for the report, and makes the final decision. But she no longer begins from exactly the same place.
That distinction may seem subtle, but I believe it fundamentally changes how we understand what is happening. For generations, physicians learned by engaging directly with the evidence. We read entire papers, reviewed images one by one, debated difficult cases with colleagues, and, perhaps most importantly, made mistakes. Clinical judgment did not emerge overnight. It developed slowly, after years of living with uncertainty. I often wonder how that process will evolve when an increasing share of the evidence arrives already organized—and accompanied by an initial interpretation generated by an algorithm.
What I do know is that I find the debate over whether AI will replace physicians increasingly uninteresting. That question assumes that clinical judgment once belonged exclusively to the professional, and that a machine is now trying to take its place.
Perhaps clinical judgment has always been the product of a network in which people, tools, institutions, and knowledge interact. If that is true, then artificial intelligence is not replacing clinical judgment. It is joining a network that existed long before it did. And perhaps that is the most important change of all.
A few years from now, we will probably still say that a clinical decision was made by a physician. That will almost certainly remain true. What may no longer be true is the assumption that the reasoning behind that decision began and ended with that physician alone.
