The apocalyptic narrative writes itself: artificial intelligence will soon read X-rays and CT scans faster and more accurately than any human, rendering radiologists obsolete within a decade. This story has been circulating since at least 2016, when Geoffrey Hinton, the godfather of deep learning, declared that medical schools should stop training radiologists because the technology would soon outperform them. Nearly a decade later, radiology departments are busier than ever, radiologist salaries remain among the highest in medicine, and the profession has not collapsed. What happened instead is far more instructive about how AI actually changes professional work.
The grunt work revolution
The reality of AI in radiology is aggressively mundane. Algorithms now excel at specific, repetitive tasks: measuring nodules across sequential scans, flagging potential pneumothorax cases for urgent review, quantifying liver fat deposits. These are not the glamorous diagnostic breakthroughs of popular imagination but rather the tedious measurements that once consumed hours of a radiologist's day. A chest X-ray triage system does not replace the radiologist's judgment; it reorganizes the queue so that the most urgent cases surface first. The doctor still reads every image. They simply read them in a smarter order.
This pattern—automation of the mechanical, preservation of the judgmental—appears across every serious deployment. AI handles the volume problem. Radiology workloads have grown relentlessly for decades, driven by aging populations, defensive medicine, and ever-improving imaging technology. The number of images a radiologist must review has increased by orders of magnitude since the 1990s. AI tools function as cognitive exoskeletons, not replacements, allowing the same number of doctors to process vastly more studies without proportional increases in burnout.
The liability question nobody discusses
There is a reason hospitals deploy AI as a "second reader" rather than a primary diagnostician, and it has little to do with technological capability. When a radiologist misses a tumor, the liability chain is clear: the physician, the hospital, the malpractice insurer. When an algorithm misses a tumor, the legal and ethical terrain becomes treacherous. Who bears responsibility—the software vendor, the hospital that purchased it, the radiologist who relied on it, the administrator who mandated its use? Until this question has clear answers, AI will remain a tool that assists rather than decides. The technology may be ready for autonomy. The legal and institutional frameworks are not.
This institutional friction is underappreciated in discussions of AI disruption. Professions are not merely collections of tasks but dense webs of licensure, liability, reimbursement codes, and institutional relationships. Transforming radiology requires not just better algorithms but new billing structures, updated malpractice frameworks, revised training curricula, and renegotiated relationships between radiologists and referring physicians. Technology moves in months; institutions move in decades.
The skills that matter now
Young radiologists entering the field today face different demands than their predecessors. The ability to rapidly interpret straightforward cases—a skill that once distinguished competent practitioners—matters less when AI handles initial triage. What matters more is the capacity to adjudicate difficult cases where algorithms express uncertainty, to integrate imaging findings with complex clinical contexts, and to communicate nuanced probabilistic assessments to anxious patients and harried clinicians. The radiologist of the future is less a pattern-recognition machine and more an expert consultant, a role that requires deeper clinical integration rather than isolation in a dark reading room.
Our take
Radiology offers a preview of AI's actual impact on knowledge work: not the dramatic displacement of headlines but the gradual redistribution of cognitive labor. The profession is not dying; it is molting, shedding mechanical tasks while its core—judgment under uncertainty, accountability for decisions, integration of information across contexts—becomes more rather than less valuable. Those who predicted radiologists would vanish confused the ability to perform a task with the willingness of institutions to let machines bear responsibility for it. The former is a technical problem, largely solved. The latter is a civilizational question we have barely begun to answer.
