Underfitting occurs when a model is too simple to capture the underlying trend in data, resulting in high bias. It performs poorly on both training and test sets. Solutions include adding more features, increasing model complexity, or reducing regularization.