Human-AI dyads in radiology: insights from human factors research

Artificial intelligence (AI) systems are increasingly embedded within radiological practice, most commonly as decision-support tools intended to augment diagnostic accuracy, efficiency, and consistency. Although substantial advances in algorithmic performance have been demonstrated across multiple imaging tasks, translation into consistent clinical benefit has been variable.

This observation has shifted attention from model-centric evaluation toward the interaction between humans and AI systems within complex clinical workflows. In this review, we examine the evidence on human-AI dyads with a particular focus on diagnostic radiology and propose Kremer’s O-Ring theory as a unifying conceptual framework. The O-ring model conceptualizes system performance as the product of multiple interdependent components, such that overall reliability is constrained by the weakest element rather than average performance. Applied to radiology, this framework provides a coherent explanation for why improvements in algorithmic accuracy alone may fail to yield proportional gains in patient outcomes.

We perform a conceptual review of empirical evidence on human-AI collaboration, automation bias, verification behavior, and workflow integration and outline a future research agenda emphasizing system-level design. An O-Ring-informed perspective underscores the need to address human factors and technical systems in concert to support safe and effective integration of AI into radiological practice.