A study published last month showed that a computer vision model correctly identified pneumonia on chest X-rays with greater accuracy than a panel of board-certified radiologists. Another model detected early signs of breast cancer on mammograms that human readers had missed. A third flagged subtle fractures on CT scans that had been labeled normal in the initial report.
The headlines were predictable: “AI beats doctors.” “Radiologists obsolete.” “The end of human diagnosis.”
The reality, as usual, is more complicated. AI can perform well on a specific task, using clean images, in a controlled setting. Deploying that same model in a busy hospital, with real patients, messy data, and overworked staff, is a different problem entirely.
The Benchmark Gap
Most of the impressive AI results you read about come from research datasets"carefully curated, anonymized, and labeled images. The models are tested on held-out data from the same distribution. They are compared against radiologists who are reading the images without clinical context, without access to patient history, and often under artificial time constraints.
In a real hospital, none of those conditions hold.
Images come from different machines, with different settings, from different manufacturers. A patient moves during the scan. The image includes artifacts. The radiologist knows the patient is 82 years old, has a history of smoking, and came in with shortness of breath. That context changes how an image is read.
An AI model does not have that context unless it is explicitly given. And even then, it does not “understand” it the way a clinician does.
The Workflow Problem
Even if an AI model is accurate, it has to fit into a radiologist’s day. A radiologist may read 100 studies in a shift. Each study takes a few minutes. The AI needs to integrate with the picture archiving and communication system (PACS), the electronic health record (EHR), and the reporting software.
If the AI produces results in a separate window, radiologists will not use it. If it adds clicks to their workflow, they will not use it. If it slows them down, they will not use it.
The best AI tools are invisible. They flag a finding in the corner of the image the radiologist is already looking at. They suggest a measurement that matches what the radiologist would have done. They do not interrupt.
Building that kind of integration is not a machine learning problem. It is a software engineering and user experience problem. And it is harder than publishing a paper.
The Liability Problem
When a radiologist misses a finding, there are established processes: peer review, quality assurance, malpractice insurance. The radiologist is accountable.
When an AI model misses a finding, who is responsible? The hospital that deployed it? The vendor that built it? The radiologist who signed off on the report? The answer is not clear, and hospitals are understandably cautious.
Most hospitals that use AI today use it as a “second reader”"a safety net that flags potential misses, rather than as a primary diagnostic tool. That is sensible, but it also means the AI is not replacing human labor. It is adding to it, at least for now.
The Training Data Problem
AI models are only as good as the data they are trained on. If a model is trained mostly on images from one hospital, one manufacturer, one patient population, it may not generalize well to others.
There are documented cases of AI models performing worse on images from underrepresented demographic groups. There are models that learned to recognize the hospital that produced the image rather than the disease in the image"because all the cancer cases happened to come from one scanner.
Building models that work across different hospitals, different populations, and different equipment requires diverse data. Collecting and labeling that data is expensive, slow, and subject to privacy regulations.
What Is Actually Happening
AI is already in use in some hospitals, mostly for specific tasks: detecting lung nodules on CT scans, measuring breast density on mammograms, flagging intracranial hemorrhage on head CTs. These are narrow, well-defined problems where the model’s output can be easily verified.
The broader vision"an AI that reads any study, understands the clinical context, and produces a diagnostic report"is still years away. The technology is advancing, but the deployment challenges are substantial.
For now, the most realistic role for AI in radiology is as an assistant, not a replacement. It can help prioritize urgent cases. It can reduce fatigue-related errors. It can give a second opinion on difficult images. It cannot, and should not, be trusted to practice medicine on its own.
The radiologist is not obsolete. The radiologist who refuses to use AI may be.
Sources and Further Reading
-
Nature Medicine " AI in medical imaging: current status and future directions (2026)
https://www.nature.com/nm/ -
U.S. Food and Drug Administration " AI/ML-Enabled Medical Devices (2026)
https://www.fda.gov/ -
European Society of Radiology " AI in Radiology: A Guide (2026)
https://www.myesr.org/ -
The Lancet Digital Health " Clinical deployment of AI in radiology (2026)
https://www.thelancet.com/journals/landig/ -
Stanford Medicine " AI and the Future of Radiology (2026)
https://med.stanford.edu/
Use original reporting and public scientific literature. Do not copy paragraphs from the sources.