Magnetic resonance (MR) imaging of the shoulder necessitates high spatial and contrast resolution resulting in long acquisition times, predisposing these images to degradation due to motion. Autocorrection is a new motion correction algorithm that attempts to deduce motion during imaging by calculating a metric that reflects image quality and searching for motion values that optimize this metric. The purpose of this work is to report on the evaluation of 24 metrics for use in autocorrection of MR images of the rotator cuff. Raw data from 164 clinical coronal rotator cuff exams acquired with interleaved navigator echoes were used. Four observers then scored the original and corrected images based on the presence of any motion-induced artifacts. Changes in metric values before and after navigator-based adaptive motion correction were correlated with changes in observer score using a least-squares linear regression model. Based on this analysis, the metric that exhibited the strongest relationship with observer ratings of MR shoulder images was the entropy of the one-dimensional gradient along the phase-encoding direction. We speculate (and show preliminary evidence) that this metric will be useful not only for autocorrection of shoulder MR images but also for autocorrection of other MR exams.