Trust, Scrutiny, or Collaboration? A Performance-Based Framework for Human-AI Interaction in Medicine

AbstractArtificial intelligence (AI) is increasingly integrated into clinical decision-making, but strategies for human-AI collaboration remain underdeveloped. A recent randomized trial by Qazi and colleagues found that physicians exposed to incorrect large language model recommendations scored 14 percentage points lower on diagnostic reasoning than those receiving error-free suggestions, attributing this to automation bias. We argue that this framing is incomplete: The observed effect more closely resembles the well-documented impact of inaccurate information on reasoning, a phenomenon equally present in human-to-human consultation. More importantly, the appropriate response to AI errors is not uniform skepticism. Whether a physician should defer to, scrutinize, or collaboratively engage with AI depends on two key dimensions: the relative accuracy of humans and AI on a given task, and the degree to which their errors are complementary. We propose a framework mapping these dimensions onto four interaction zones: human-dominant, AI-dominant, hybrid review, and disagreement resolution. Each requires a distinct clinical workflow strategy. This framework is intentionally dynamic: As AI performance evolves and physician expertise develops, optimal interaction strategies must be recalibrated accordingly. Trust calibration, not blanket skepticism, is the appropriate goal. (Funded by the Netherlands Organisation for Health Research and Development.).