Automation Bias and Deskilling in Radiology AI: A Confidence-stratified Triage Model for Workflow Integration

Radiology is adopting artificial intelligence (AI) faster than it is studying what AI does to radiologists' perceptual expertise and interpretive skill. The question has partly shifted from whether these tools perform well on benchmark tasks to what happens to clinical judgment when use of these tools becomes routine. Two risks accompany this transition: automation bias, the tendency to follow AI suggestions without adequate scrutiny, and deskilling, the erosion of perceptual and interpretive ability through disuse, typically revealed when AI support is withdrawn. Incorrect AI output degrades accuracy regardless of experience, the benefit of assistance varies unpredictably across readers and tasks, and performance falls after AI withdrawal; neither risk has been studied longitudinally in radiology. A confidence-stratified triage model is proposed in which calibrated AI confidence, not findings alone, routes cases to low-, intermediate-, or high-confidence streams: output withheld until independent reading, shown with uncertainty signals and a checklist, or shown with random blinded audit. Demonstrated local confidence calibration is a precondition for deployment, and stream thresholds and audit rates should be weighted by task and potential clinical harm, not error probability alone. Rather than applying one workflow to every case, the model redistributes radiologist attention to where AI error is most likely and most consequential, operationalizing four principles: keeping the radiologist in the decision loop, making AI uncertainty visible, inserting targeted cognitive friction, and governing the human-AI interaction as a system. A research agenda to determine whether these countermeasures reduce automation bias and skill erosion in practice is also proposed.