The pitch has been consistent for nearly a decade: AI will read medical scans faster, cheaper, and more accurately than radiologists ever could. Geoffrey Hinton, the computer scientist often called the godfather of deep learning, suggested years ago that training radiologists was essentially pointless given AI's trajectory. The profession, we were told, was walking dead.
Radiology departments worldwide now deploy AI tools for everything from detecting lung nodules to flagging potential strokes. The technology works. It also hasn't replaced a single radiologist. Understanding why illuminates something important about how artificial intelligence actually integrates into skilled professions—and why the replacement narrative consistently misses the point.
The worklist problem
A radiologist's day revolves around the worklist: a queue of studies awaiting interpretation. Before AI, that queue operated on a first-in, first-out basis modified by clinical urgency flags set by ordering physicians. The system worked adequately but contained an obvious inefficiency—a scan showing a massive brain bleed might sit behind routine chest X-rays simply because it arrived later.
AI's most valuable contribution has been triage. Algorithms scan incoming studies and push critical findings to the top of the queue. A radiologist in Manchester might see a flagged pulmonary embolism within minutes of acquisition rather than hours. This genuinely saves lives. It also increases radiologist productivity, because the cognitive load of worrying about what's buried in the queue diminishes.
But triage is assistance, not replacement. The algorithm flags; the radiologist diagnoses. The distinction matters legally, clinically, and practically. When an AI system misses a finding—and they do, regularly—the radiologist remains responsible. When it flags a false positive—and they do, frequently—the radiologist must investigate, adding work rather than subtracting it.
The liability vacuum
No AI system currently approved for clinical use in major markets carries diagnostic responsibility. The Food and Drug Administration clears these tools as decision support, not decision makers. Insurance frameworks, malpractice law, and hospital credentialing all assume a licensed physician stands behind every interpretation.
This isn't a temporary regulatory lag. It reflects something fundamental about medical diagnosis: someone must be accountable when things go wrong. A patient harmed by a missed cancer needs recourse. An algorithm cannot be sued, cannot lose its license, cannot explain its reasoning in court. The radiologist can.
Some AI advocates argue this will change as systems become more reliable. Perhaps. But reliability isn't the core issue. Medicine requires judgment calls in ambiguous situations, communication with patients and referring physicians, integration of clinical context that doesn't appear on the scan. A shadow on a lung image means something different in a twenty-five-year-old nonsmoker than in a seventy-year-old with a history of malignancy. The radiologist knows this because they read the chart, call the oncologist, remember the patient from last year.
What actually changed
Radiology has transformed, just not through replacement. Workflow software incorporating AI has made departments measurably more efficient. Subspecialization has intensified—a neuroradiologist today might read nothing but brain MRIs, developing expertise no generalist could match. Communication expectations have shifted; radiologists increasingly consult directly with clinical teams rather than issuing reports into the void.
The profession has also become more interesting. Tedious tasks—measuring tumor dimensions, counting lung nodules, calculating ejection fractions—increasingly fall to algorithms. The radiologist's role shifts toward synthesis, communication, and judgment. This resembles what happened to accountants after spreadsheet software: the profession didn't disappear, it elevated.
Training programs have adapted accordingly. Residents learn to work with AI tools, understanding their capabilities and limitations. They also learn, perhaps more explicitly than before, what machines cannot do: build trust with a frightened patient, navigate institutional politics, make the call when the scan is genuinely ambiguous.
Our take
The AI-will-replace-radiologists narrative was always more about Silicon Valley's self-image than medical reality. It assumed diagnosis was pattern matching, that expertise was mere data processing, that accountability was a detail to be sorted later. Radiology's actual transformation—toward augmented capability, shifted responsibilities, and elevated judgment—offers a more honest template for how AI changes skilled work. The machines do what machines do well. The humans remain human, with all the irreducible complexity that implies. Geoffrey Hinton's prediction looks less like prophecy and more like a category error.
