The present study used eye-tracking to examine the relationship between attention and category learning in probabilistic environments. While training, participants received either perfect feedback (100% accurate), or one of three different levels of probabilistic feedback (87.5%, 75% or 62.5% accurate). It was found that participants in the 87.5% condition were more accurate than participants in the other two probabilistic feedback conditions. However, despite their greater accuracy, participants in the 87.5% condition continued to attend to irrelevant information as frequently as those in the other two probabilistic conditions. This shows that: (1) cues that are not utilized in making a categorization decision may still be frequently attended to, and (2) attentional learning is not as tightly coupled to improving accuracy as current formal models suggest.