A Nature Communications study found that large reasoning models improve primary diagnosis accuracy over standard language models but also increased false positives by more than 34%. Physician collaboration improved outcomes, suggesting that human-in-the-loop review can offset some model errors. The key operational issue is not accuracy alone, but calibration trade-offs—where higher sensitivity can manifest as more over-calling. The results highlight that clinical integration of AI systems must consider error types and how they propagate in real decision workflows. The study underscores why deployment requires more than benchmark performance, including monitoring for false-positive rates and structured review processes.