In this study we assessed the accuracy of the General Health Questionnaire in detecting psychiatric disorders in general medical out-patients. A total of 290 newly referred patients were interviewed with the Present State Examination. Prior to the interview, 112 patients completed the full GHQ-60, 100 completed the GHQ-30 and 78 completed the GHQ-12. Data from the first group were used to study the full GHQ-60, together with the GHQ-30 and and GHQ-12, when disembedded from the full questionnaire. In a comparison between the disembedded and the separate versions of the GHQ-30 and GHQ-12 we observed considerable variation in the cut-off scores where a certain sensitivity and specificity was attained. In ROC-analysis, the versions were not materially different in their discriminatory capacity (area under the curve). The use of different criteria to define a ‘case’ demonstrated that case severity was another source of increasing cut-off scores. Our data demonstrate that the use of disembedded or separate versions of the questionnaire, together with variation in the case criteria can be a major explanation for variation in cut-off scores that was observed in previous studies.