In a simple data entry task, feedback and error correction routines will allow users to detect misrecognitions and to re-enter incorrect data. Even when these resources are available, a small proportion of data strings are still entered incorrectly. Examination of these cases indicates that the source of these errors depends on the modality of feedback used. With visual feedback, users sometimes fail to monitor incorrect feedback. With spoken feedback, the problem is more likely to involve confusions within the task dialogue. The implications of these findings for the design of feedback and error correction procedures are discussed.