AI chatbots reading X-rays can be dangerously confident even when they're wrong

A new benchmark called RadLE 2.0 reveals that many medical AI models provide incorrect radiology findings with high levels of confidence. The study shows that these systems often fail to identify when they should defer to human expertise, highlighting a significant safety gap in current automated diagnostic technology. Human radiologists continue to outperform these models, suggesting that AI is not yet ready to function autonomously in clinical settings.