AI in Radiology: Collaboration, Not Replacement

Geoffrey Hinton's 2016 prediction that AI would replace radiologists within five years has not come true; instead, radiology is growing, with a projected 26% increase in practitioners over the next three decades. As of early 2026, about three-quarters of the 1,400 AI-enabled medical devices cleared by the FDA are for radiology. AI tools can draft reports, flag urgent images, and identify abnormalities invisible to the human eye, but human error rates of 3–5% translate to 40 million errors worldwide annually. The challenge is combining AI precision with human experience. Radiologists must evaluate AI decisions, a process requiring "mental rewiring," says Paul Yi of St. Jude Children's Research Hospital. Neural networks are "black box" systems, making it hard for radiologists to know when to override them. Curtis Langlotz of Stanford notes that AI and human intelligence are different: AI examines every pixel without fatigue, while radiologists understand disease context. Nina Kottler of Mosaic Clinical Technologies highlights automation bias and automation complacency as pitfalls. Training is essential; a 2026 AMA survey found over a quarter of physicians received no AI training. Kottler advocates for AI tools to report confidence estimates. Langlotz's key insight: "Radiologists who use AI will replace radiologists who don't."