MIT researchers demonstrate that best-performing machine learning models can become worst-performing when applied to new data environments, revealing hidden risks from spurious correlations in medical AI and other critical applications.