MIT researchers have identified and fixed a critical flaw in AI model training that causes overconfidence in wrong answers. The breakthrough enables AI systems to appropriately express uncertainty without sacrificing performance.