Researchers found that radiologists who received step-by-step “chain-of-thought” explanations from GPT-4 achieved higher diagnostic accuracy than those receiving standard large language model outputs, differential-diagnosis lists, or no artificial intelligence support in a randomized vignette study involving 101 US radiologists and 2,020 diagnostic assessments. Chain-of-thought explanations improved physician diagnostic accuracy by 12.2 percentage points compared with the control group and appeared to promote more appropriate reliance on model recommendations by helping physicians override incorrect suggestions while adhering to correct ones.