Large language models improve physician accuracy but lead to false reliance.

AI assistance may improve diagnostic performance while simultaneously introducing automation bias. This experimental preprint study evaluates how physicians use large language models (LLM) for clinical diagnosis. LLM assistance increased diagnostic accuracy, but the 46 physicians testing a proprietary agentic retrieval-augmented LLM were more likely to accept incorrect recommendations when supported by apparently credible citations. Findings highlight a critical diagnostic-safety tradeoff: improved performance can coexist with false reliance, emphasizing calibration, oversight, and human factors.