Wednesday, December 2 | 9:30 a.m.-10:30 a.m. | W3-SSIN06-3 | Room E450B
Rather than eliminating diagnostic risk in radiology, AI shifts it, according to an analysis set to be presented in Chicago.
AI is increasingly part of radiology workflows, yet its diagnostic failures in routine use remain poorly characterized, noted presenter Guillaume Herpe, MD, PhD, of the University of Poitiers in France, and colleagues. The group set out to classify these errors and see whether they match the patterns long documented for human radiologists.
The researchers analyzed 942 incident reports submitted to a multicenter post-market surveillance registry in late 2025. The reports involved 10 commercial AI tools used in chest, abdominal, and neuroimaging at six institutions, and two senior radiologists categorized each one.
Missed findings were the most common error, followed by workflow and integration failures. About half of the errors mapped to established human error categories, mostly perceptual, according to the authors. Nearly one in five, however, stemmed from deployment instability and performance drift, problems with no human equivalent. Human oversight limited downstream impact in 78% of cases.
“AI diagnostic errors substantially overlap with human radiologist error patterns while revealing system-level and temporal vulnerabilities unique to AI,” the researchers wrote. “Rather than eliminating diagnostic risk, AI redistributes it — reducing perceptual misses while introducing deployment and drift-related failure. Human oversight remains essential in AI-augmented workflows.”
Head to this Wednesday morning session for more information.
