Signal Dual- versus Single-Suggestion AI Support for Radiographic Interpretation in Residents: Randomized Multireader Study

Summary

A randomized multireader study conducted at three hospitals in China between July and September 2026 compared single- versus dual-AI-suggestion support for radiographic interpretation among 123 analyzed residents with less than three years of clinical experience. Participants were randomized to receive suggestions from GPT-5.4 alone, GPT-5.4 plus Kimi-K2.6, or GPT-5.4 plus Gemini-3.6 Flash, and interpreted 60 radiographs before and after receiving AI support. Radiology residents showed significantly greater accuracy improvement with dual-suggestion support than with single-suggestion support, while non-radiology residents showed no significant difference. When the shared GPT-5.4 suggestion was incorrect, dual-suggestion support markedly improved accuracy over single-suggestion support in both radiology and non-radiology residents. The effect of dual-suggestion support differed significantly by medical specialty. The findings suggest that presenting multiple AI suggestions, rather than a single one, can reduce clinicians' over-reliance on an individual model's errors, with direct implications for how clinical AI tools should be designed and deployed.