Using both behavioral and eye-tracking methodology, we tested whether and how asking students to generate predictions is an efficient technique to improve learning. In particular, we designed two tasks to test whether the surprise induced by outcomes that violate expectations enhances learning. Data from the first task revealed that asking participants to generate predictions, as compared to making post hoc evaluations, facilitated acquisition of geography knowledge. Pupillometry measurements revealed that expectancy-violating outcomes led to a surprise response only when a prediction was made beforehand, and that the strength of this response was positively related to the amount of learning. Data from the second task demonstrated that making predictions about the outcomes of soccer matches specifically improved memory for expectancy-violating events. These results suggest that a specific benefit of making predictions in learning contexts is that it creates the opportunity for the learner to be surprised. Implications for theory and educational practice are discussed.