Learning curves are an important tool in cognitive diagnostics modeling to help assess how well students acquire new skills, and to refine and improve knowledge component models. Learning curves are typically obtained from a model estimated on real data obtained from a finite, and usually limited, sample of students. As a consequence, there is some uncertainty associated with estimating the model from that sample, and a risk that the inferences made using learning curves derived from the estimated model are over-confident one way or another. Based on previous work modeling the uncertainty on Additive Factors Model parameters, we derive a principled way to quantify the confidence in learning curves associated with each knowledge component. We show that our approach leads to relatively tight bounds on the learning curves, much tighter than a naive approach relying only on parameter uncertainty. This also reveals a disparity across knowledge components regarding how confident one can be in how well these skills are mastered. [For the full proceedings, see ED607784.]