At a time when forecasts indicate that artificial intelligence will take over the role of human radiologists — the actual impact of AI in radiology emerges as a striking counter‑example. 

The recent article, “AI isn’t replacing radiologists” (Works in Progress, September 25, 2025, by Deena Mousa) portrays a more nuanced reality. Even though AI tools have leapt forward in number of models approved by the Food and Drug Administration (FDA) and the increased utilization of those models, radiologists have not, by any stretch, been rendered obsolete. On the contrary, despite the hundreds of AI imaging models approved by the FDA, workforce shortages in radiology are widespread. Radiology residency programs are not keeping pace with demand for radiologists, whose services are highly sought after — making them highly compensated, with salaries climbing — especially as digital imaging studies proliferate while the population ages.

AI models have become astonishingly capable on certain types of diagnostic tasks. Some tools now detect pneumonia, identify nodules, spot stroke or clots, breast cancer, lung cancer, and more — often across multiple imaging modalities. In benchmark tests, many AI models outperform (or match) radiologists in narrowly defined tasks under tightly controlled conditions. But as the article underscores, translating benchmark success into real‑world impact has proved challenging. 

While over 700 radiology AI tools have cleared FDA approval, they tend to cluster around certain high‑volume, high‑visibility conditions — breast cancer, lung cancer, stroke — rather than the often more complex, less common subspecialties (e.g., vascular, head & neck, thyroid, spine). The author cites two key reasons: 

1. For some imaging modalities or patient populations, there have not been enough well‑annotated examples to build robust, generalizable AI models. Non‑standard imaging modalities like ultrasound pose special challenges.
2. Fully autonomous diagnostic models face steep legal and ethical barriers. Insurers and regulators have been reluctant to allow software to autonomously produce diagnostic reports without human oversight. Lawsuits and malpractice risk loom large. Coverage or reimbursement is often limited to “assistive” tools rather than autonomous ones.

The picture that emerges is that AI tools excel in narrow, well‑defined tasks, freeing radiologists to focus on the broader, more human parts of their work. Radiologists are in demand now more than ever to: interpret the exponentially growing numbers of imaging studies; bring context to findings; perform the necessary supervision of diagnostic services, and; interact with patients. 

So while AI has transformed some aspects of radiology, it hasn’t replaced radiologists — and the reasons are complex. The work of radiologist involves human judgment, nuance, and interaction. Their work remains beyond what current AI tools can replicate. Thus even in places where AI aids in image interpretation, radiologists remain indispensable.