Although many learning theories make use of negative evidence, it is often overlooked in the language-learning literature, leading both to claims that learning simple aspects of grammar is logically impossible and appeals to a universal grammar. Here we investigate the ability of young children to correct their tendency to over-regularize plural nouns. We present an error-driven model of plural learning that makes a surprising prediction: at an appropriate stage in learning, childrens tendency to over-regularize irregular plurals can be reduced through exposure to regular plurals alone. We describe a simulation and a behavioral experiment showing that, consistent with the models predictions of U-shaped learning, memory testing on regular plurals led to significant reductions in plural over-regularization in six-year-olds, while increasing over-regularization in four-year-olds. Prediction error appears to be a strong corrective source of evidence in learning, suggesting that learning language may be far more possible than is sometimes supposed.