“We could replace a great deal of radiologists with AI at this moment.” Mitchell H. Katz, MD, president and CEO of NYC Health + Hospitals
We’ve heard versions of this claim for nearly a decade. AI has made real progress. Today, roughly half of U.S. radiologists use AI in some capacity, and over three-quarters of FDA-cleared AI devices are in radiology. Despite this progress, AI is not ready to replace radiologists in any meaningful or scalable way.
Health system leaders suggest that large portions of radiology could already be replaced if regulations allowed it, citing cost savings and improved performance in screening mammography. Although the economics may be compelling, framing this as ‘replacement’ assumes the radiologist’s role is purely image analysis. AI fundamentally changes the economics of knowledge but does not yet replicate judgment. This distinction is where the gap remains.
Rishi SethRad AI
As a radiologist working at the intersection of clinical practice and AI development, I believe deeply in the technology’s potential. I also believe that deploying AI as a replacement for physicians today is premature and carries risks that are not yet fully understood.
Lessons from early predictions
In 2016, Geoffrey Hinton predicted that training new radiologists would soon become unnecessary. Andrew Ng echoed similar concerns, suggesting even highly specialized physicians were vulnerable to displacement. Those predictions have not held up. Not because AI stalled, but because the real-world practice of medicine is more complex than early assumptions suggested. Radiology is not purely image interpretation.
Both have since revised their views. The reality has shifted toward collaboration, not replacement: AI as a tool that enhances efficiency and accuracy rather than eliminates clinical oversight. Jensen Huang has also noted that “the surprising thing is the prediction that radiologists would be the first jobs to go was exactly the opposite.” The trend shows that more radiologists are being hired now due to AI.
That shift reflects a deeper truth: early success in controlled environments often fails to translate into clinical reality, where variability in imaging protocols, patient populations, and clinical context is the rule.
Growing demands
The United States faces a significant shortage of radiologists projected to persist for decades, driven by rising demand, an aging population, burnout, declining reimbursement, and workforce attrition.
While this creates an understandable pressure to automate, clinical readiness requires consistency across settings, judgment in ambiguous cases, and accountability for errors, which are conditions that AI has not yet met.
AI in practice
While vision language models (VLMs) will almost certainly take on a larger share of imaging interpretation in the coming years, they are not a replacement. It is a rational reallocation of a scarce cognitive resource. The full scope of radiology involves contextual reasoning not captured in a single image.
Radiology is not image classification. It is the integration of imaging findings into a clinical narrative. We do not simply identify abnormalities; we determine what they mean in the context of a patient’s history, symptoms, prior studies, and clinical trajectory. We weigh uncertainty, assess risk, and decide what matters.
This is the difference between detection and judgment, and today’s AI falls short on both. Most radiology AI tools are narrow by design, built to detect a single finding on a single modality. In contrast, a complex cross-sectional examination requires simultaneous assessment across dozens of organ systems. Vision language models show promise in this direction, but significant gaps remain. And even where detection is strong, judgment and reasoning are still absent.
Liability is the real cost
The “major savings” argument for replacing radiologists is shortsighted, as radiologist salaries account for less than one percent of hospital operating costs, compared with administrative overhead, which is nearly a third of U.S. healthcare spending.
Rather, the core problem is responsibility: when the radiologist is removed, the clear chain of accountability disappears.
Responsibility becomes diffuse, potentially contested between the hospital, provider, and vendor. Current AI systems are approved only for clinical decision support, and vendors explicitly disclaim diagnostic responsibility, creating an undefined liability framework for hospitals that use them autonomously.
Furthermore, patients largely prefer AI as a second reader, rather than a replacement, and expanding access this way risks lowering the standard of care. Accountability is the hard question that cost models do not address.
Future of AI in radiology
None of this is an argument against AI in radiology. Quite the opposite.
AI that identifies and triages critical findings, reduces turnaround times, tracks incidental lesions, assists in report generation, provides contextual guidance, and identifies potential errors represents a meaningful and necessary advance. These tools augment clinical judgment, reduce cognitive burden, and improve consistency, accuracy, and efficiency.
The role of the radiologist is not disappearing anytime soon. It is, however, shifting. As AI expands access to knowledge, the value of radiology increasingly lies in the ability to apply judgment: to interpret findings in context, weigh uncertainty, and decide what matters for the patient in front of us.
It is highly likely that AI will reach a point where autonomous interpretation becomes viable. But that threshold is not defined solely by technical performance. It requires robust validation across diverse clinical settings, clear regulatory frameworks, and an accountability system that protects patients when errors occur. We are not there yet.
This is not just a technical decision for health systems. It is a clinical and operational one, with real consequences for how risk is managed and where it ultimately sits. The threshold for replacement is not technical performance alone. It is the ability to exercise judgment under uncertainty and to bear responsibility for the consequences.
AI has not yet crossed that threshold.
Until it does, removing the physician from the diagnostic process is not progress. It is a shift of risk onto patients without a system prepared to absorb it.
Rishi Seth, MD, CIIP, is a neuroradiologist and chief medical innovation officer at Rad AI.
The comments and observations expressed are those of the author and do not necessarily reflect the opinions of AuntMinnie.com.