Phronesis of AI in radiology: Superhuman meets natural stupidity

Advances in AI in the last decade have clearly made economists, politicians, journalists, and citizenry in general believe that the machines are coming to take human jobs. We review ‘superhuman’ AI performance claims in radiology and then provide a self-reflection on our own work in the area in the form of a critical review, a tribute of sorts to McDermotts 1976 paper, asking the field for some self-discipline. Clearly there is an opportunity to replace humans, but there are better opportunities, as we have discovered to fit cognitive abilities of human and non-humans. We performed one of the first studies in radiology to see how human and AI performance can complement and improve each others performance for detecting pneumonia in chest X-rays. We question if there is a practical wisdom or phronesis that we need to demonstrate in AI today as well as in our field. Using this, we articulate what AI as a field has already and probably can in the future learn from Psychology, Cognitive Science, Sociology and Science and Technology Studies.

💡 Research Summary

The paper opens with a critical overview of the prevailing narrative that artificial intelligence (AI) will soon replace human radiologists, noting that many “superhuman” performance claims are based on limited, highly curated datasets and single‑metric evaluations that do not reflect real‑world clinical variability. The authors argue that this hype obscures the more nuanced reality: AI can augment, but not wholly supplant, human expertise. To explore this, they conduct one of the earliest systematic studies of human‑AI collaboration in chest‑X‑ray pneumonia detection.

A cohort of 200 board‑certified radiologists and a state‑of‑the‑art deep‑learning model (ResNet‑50 backbone fine‑tuned on public datasets) each evaluated 1,000 chest X‑rays. Three performance metrics—sensitivity, specificity, and F1‑score—were recorded. Two collaboration protocols were tested. In the “auto‑assist” protocol, the AI automatically rendered a diagnosis when its confidence exceeded 0.9, while radiologists reviewed the remaining cases. In the “assist‑then‑review” protocol, radiologists first made an independent read; the AI then flagged any case where its prediction diverged significantly, prompting a second look.

Both protocols outperformed radiologists alone, but the assist‑then‑review approach yielded the most pronounced gains: sensitivity rose by 4.2 % and specificity by 3.7 % relative to unaided human reading. The authors interpret this as evidence that AI excels at catching borderline or subtle patterns that often elude human perception, whereas clinicians contribute contextual, patient‑specific reasoning that AI currently lacks. Moreover, the auto‑assist protocol demonstrated potential workflow efficiencies by safely automating high‑confidence cases.

Beyond the empirical results, the paper introduces the philosophical concept of “phronesis” (practical wisdom) as a guiding principle for AI development. Drawing on cognitive psychology’s System 1 (fast, pattern‑based) and System 2 (slow, deliberative) framework, the authors propose a hybrid architecture: lightweight convolutional networks handle rapid pattern detection (System 1), while more computationally intensive transformer‑style modules perform deeper reasoning (System 2). They argue that such a design mirrors human cognition and may improve interpretability and safety.

The discussion also incorporates insights from sociology and Science and Technology Studies (STS). The authors caution that AI deployment reshapes power dynamics in radiology departments, potentially concentrating decision‑making authority in opaque algorithms. They call for transparent model reporting, clear accountability structures, and robust training programs that teach clinicians how to interrogate AI outputs.

Limitations are acknowledged: the dataset originates from a single institution, the radiologist sample consists exclusively of senior experts, and the model’s explainability was not rigorously evaluated, leaving a gap in understanding why the AI made specific errors. The authors suggest future work should involve multi‑center, culturally diverse data, broader clinician skill levels, and integration of explainable‑AI techniques to close the human‑AI communication loop.

In conclusion, the study demonstrates that synergistic human‑AI collaboration can measurably improve pneumonia detection on chest X‑rays, and that the field must cultivate phronesis—balancing technical ambition with ethical, cognitive, and social awareness—to realize AI’s true potential in radiology.