Expected Improvement is a criterion used in Bayesian optimization to balance exploration and exploitation by quantifying the expected gain from sampling a new point. It helps in efficiently finding the global optimum of a function by prioritizing areas with high potential for improvement over the current best observation.