Much can be learned about the world by examining the discrepancies between what is expected and what actually occurs. Although many formal learning theories make use of prediction error as an important— even necessary—component in explaining behavior, this source of evidence has been largely overlooked in the language-learning literature. In this paper, we show how incorporating prediction error into a model of plural word learning (Ramscar & Yarlett, 2007) can yield a surprising prediction: that at an appropriate point in learning, the tendency of children to over-regularize irregular plurals can be reduced, by exposing them to regular plurals alone. We report on an experiment, which was designed to test the model's predictions empirically. The findings indicate that memory testing on regular plurals led to significant reductions in the rates of plural over-regularization in six-year-olds.