Fifty-seven subjects with shoulder pain, the cause of which was not clear on clinical examination, underwent magnetic resonance imaging (MRI) and arthroscopy. With arthroscopy as the standard of reference, the accuracy of MRI for diagnosing shoulder disease was assessed. Pathologic conditions were organized into 3 groups: (1) labral tears, (2) rotator cuff disease, and (3) other pathologic conditions. At arthroscopic evaluation there were 15 labral tears, 24 rotator cuff tears, and 27 other abnormalities. MRI identified 11 of the labral injuries, with accuracy of 62%, sensitivity 73%, and specificity 58%. Rotator cuff tears were identified by MRI with accuracy of 68%, sensitivity 96%, and specificity 49%. When partial rotator cuff tears were differentiated from complete tears, only 16 of 24 injuries were identified by MRI. Nonrotator cuff, nonlabral disease was identified with accuracy of 75%, sensitivity 63%, and specificity 87%. In our center MRI does not appear to be an accurate effective tool for assessing shoulder pathologic conditions in patients in whom the clinical picture is not clear and therefore may not be of assistance in surgical planning for patients with these difficult conditions.