The prediction was everywhere a decade ago: radiologists would be the first physicians made obsolete by artificial intelligence. Machines could read scans faster, spot tumors earlier, and never suffer from fatigue or distraction. The profession, it seemed, was living on borrowed time.
That prediction was wrong — but not in the way radiologists hoped. The specialty has not been eliminated. It has been transformed into something its practitioners barely recognize, a discipline where the core skill is no longer pattern recognition but something far more slippery: knowing when to trust the machine and when to override it.
The new workflow
In most major hospital systems today, AI pre-reads every imaging study before a human radiologist sees it. The software flags abnormalities, measures lesions, tracks changes from prior scans, and generates preliminary reports. The radiologist's job has shifted from primary reader to quality controller, from detective to judge.
This sounds like a demotion, and in some ways it is. The satisfaction of spotting a subtle finding that others missed — the professional pride that drew many physicians to radiology — has diminished. The machine usually sees it first. But the work that remains is arguably more consequential: deciding which of the AI's flags represent genuine pathology and which are false alarms, integrating imaging findings with clinical context the algorithm cannot access, and communicating uncertainty to referring physicians who increasingly expect binary answers.
The cognitive burden
Radiologists report that their days feel different now. The volume of studies has increased — AI efficiency has raised throughput expectations — but the nature of attention has changed. Instead of sustained focus on a single image, the work involves rapid toggling between AI suggestions and original data, a kind of supervisory vigilance that is mentally exhausting in ways that are difficult to quantify.
Research on automation in aviation and nuclear power has long documented this phenomenon: when humans monitor automated systems rather than performing tasks directly, their engagement drops, their skills atrophy, and their ability to catch errors degrades. Radiologists are now living this literature. The question is whether the profession can adapt its training and workflow to account for it.
What training looks like now
Radiology residency programs have begun teaching AI literacy as a core competency. Trainees learn not just anatomy and pathology but also how algorithms are trained, where they fail, and how to interpret confidence scores. Some programs simulate scenarios where the AI is deliberately wrong, forcing residents to practice independent judgment under pressure.
This represents a genuine intellectual shift. The radiologist of the future is not simply a physician who reads images but a hybrid professional fluent in both clinical medicine and machine learning epistemology — someone who understands why a model trained on one population may fail on another, and who can articulate that uncertainty to colleagues and patients.
Our take
The story of AI and radiology is not a story of replacement but of redefinition. The profession has survived, but its practitioners are doing different work than they trained for, experiencing different satisfactions and frustrations, and developing different skills. This is probably the template for how AI will reshape most knowledge work: not sudden obsolescence but gradual transformation, a slow rewriting of job descriptions until the original role is unrecognizable. Whether that transformation is progress depends on what we valued in the original work — and whether we are honest about what we are losing.
