Radiology Holds Three-Quarters of Cleared Medical AI—and a New Veto Problem

Radiologists are not being displaced by AI image readers. They are being asked to become the final check on systems that may outperform them on some tasks, yet still make unfamiliar mistakes that can be hard to explain or detect. That is a demanding safety role, not a ceremonial one.

Radiology is the clearest current concentration of medical AI: about three-quarters of roughly 1,400 AI-enabled medical devices cleared by the Food and Drug Administration by early 2026 were designed for the specialty. These tools can draft reports, flag images needing urgent attention, detect abnormalities, identify tumor subtypes, and trace lesion boundaries.

That reality is far from the replacement forecast Geoffrey Hinton made in 2016, when he predicted computers would replace radiologists within five years. Radiology employment has continued to grow instead. But the failed forecast does not mean the technology is marginal: AI-assisted systems can improve detection in defined settings, including colonoscopy, where an analysis of 43 clinical trials found more polyps detected than with conventional procedures.

The work shifts from reading to judging

The safety problem comes from a mismatch in strengths. Neural-network image systems can examine every pixel and compare an image with patterns learned from prior examples. They generally do not show clinicians the reasoning behind an output, however. A radiologist must therefore decide when to accept or reject a conclusion from a system whose path to that conclusion is opaque.

Average accuracy alone cannot settle that decision. In an illustrative example, an AI tool detects 95% of lung nodules and radiologists detect 90%; the radiologists may still catch some nodules the tool misses because human and machine errors are not identical. The value of the combined system depends on preserving those complementary catches, rather than treating the higher headline score as a mandate to defer.

Two ways oversight can fail

• Automation bias: a clinician gives excessive weight to an AI result, including a false positive that identifies disease where none exists.
• Automation complacency: a clinician fails to challenge a false negative, such as an AI system missing blood in a brain scan.

Training is part of the clinical system

The risk is not theoretical. A cited study found that experienced radiologists’ mammography accuracy dropped substantially when incorrect AI predictions influenced their decisions. The failure can also run in the other direction: a conspicuously silly AI error may lead a clinician to dismiss a system too broadly, including when it has found a genuine abnormality.

That makes generic instructions to “use AI responsibly” inadequate. Clinicians need to know an individual tool’s reliability and the conditions in which it fails, such as scans degraded by patient movement. Nina Kottler, chief medical AI officer at Mosaic Clinical Technologies, also argues that systems should provide a confidence estimate with each evaluation rather than only a yes-or-no answer, because no radiologist can retain the limits of multiple tools from memory.

Radiology’s lesson is less that humans must remain in the loop than that the loop must be designed. Monitoring when physicians agree or disagree with AI, teaching known failure modes, and giving them information calibrated to a specific result are proposed ways to make that collaboration safer. How to optimize the human-machine team remains unsettled, even as the number and complexity of these systems continue to increase.