AI X-RAY READERS OVERCONFIDENT WHEN WRONG

The RadLE 2.0 benchmark specifically measures whether AI models in radiology can identify when they should defer to human experts. Results show that many models provide wrong findings with full certainty, creating potential safety risks in clinical settings. Human radiologists continue to outperform AI systems in both accuracy and judgment. A key challenge is calibration—the ability to express appropriate uncertainty rather than false confidence. Before AI systems can operate autonomously in medical diagnostics, they must develop the capacity to recognize the boundaries of their knowledge. This includes knowing when to abstain from diagnosis entirely and flag cases for human review. The findings underscore that clinical AI deployment requires more than accuracy metrics. Trustworthiness and the ability to communicate uncertainty are equally essential for patient safety and physician adoption.