Learning how to allocate attention properly is essential for success at many tasks. Extant theories of categorization assume that learning to allocate attention is an error-driven process, where shifts in attention are made to reduce error. The present work introduces a new measure, error bias, which compares the amount of attentional change in response to incorrect responses versus correct responses during category learning. We first confirm that prominent categorization models predict high amounts of error bias. We then test this prediction against human eye-tracking data from 384 participants. Across 7 of 8 data sets we find that participants show minimal or no error bias. This finding suggests that attentional learning mechanisms, as implemented in influential computational models, cannot be generalized to account for measures of overt attention.