ML practitioners are increasingly adopting techniques like pruning and quantization to enable the deployment of deep learning models when latency and/or computational efficiency are concerns. Although these techniques may not compromise the overall accuracy of the models, Hooker et al. find that compression may cause disproportionately high errors on certain subsets of data – Compression Identified Exemplars. Consequently, these methods may amplify algorithmic bias by impacting the performance of the model on underrepresented features. The authors suggest that CIE might be used to efficiently audit model fairness.